Describes the (expected) average change in the outcome variable for each one-unit change in the independent variable for continuous variables, or the average change in the outcome variable for one category of the independent variable compared with a reference category for categorical variables
In our example, the baseline disease-COPD, ILD, or cancer (the reference category)-is the independent variable, and length of ICU stay and receipt of palliative care elements are the outcomes of interest. In addition, the regression models also included other independent variables considered as potential confounders, such as age, sex, and minority status. In the linear regression model, the length of ICU stay for patients with ILD was longer than for those with cancer (β = 2.75; 95% CI, 0.52-4.98; p = 0.016), which means that, on average, having ILD increased the length of ICU stay in 2.75 days when compared with the length of ICU stay among cancer patients. In the logistic regression model, the authors found that patients with ILD, when compared with cancer patients, were less likely to have any documentation of their pain assessment in the last 24 h of life (OR = 0.43; 95% CI, 0.19-0.97; p = 0.042), which means that having ILD decreased the odds of documentation of pain assessment by more than half.
Academia.edu no longer supports Internet Explorer.
To browse Academia.edu and the wider internet faster and more securely, please take a few seconds to upgrade your browser .
Enter the email address you signed up with and we'll email you a reset link.
Jesus Navarrete
Gifted Child Quarterly
Francis Huang
Veterinary World
How to cite this article: Selim AM, Elhaig MM, Moawed SA, El-Nahas E (2018) Modeling the potential risk factors of bovine viral diarrhea prevalence in Egypt using univariable and multivariable logistic regression analyses, Veterinary World, 11(3): 259-267. Abstract Aim: The present cross-sectional study was conducted to determine the seroprevalence and potential risk factors associated with Bovine viral diarrhea virus (BVDV) disease in cattle and buffaloes in Egypt, to model the potential risk factors associated with the disease using logistic regression (LR) models, and to fit the best predictive model for the current data. Materials and Methods: A total of 740 blood samples were collected within November 2012-March 2013 from animals aged between 6 months and 3 years. The potential risk factors studied were species, age, sex, and herd location. All serum samples were examined with indirect ELIZA test for antibody detection. Data were analyzed with different statistical approaches such as Chi-square test, odds ratios (OR), univariable, and multivariable LR models. Results: Results revealed a non-significant association between being seropositive with BVDV and all risk factors, except for species of animal. Seroprevalence percentages were 40% and 23% for cattle and buffaloes, respectively. OR for all categories were close to one with the highest OR for cattle relative to buffaloes, which was 2.237. Likelihood ratio tests showed a significant drop of the −2LL from univariable LR to multivariable LR models. Conclusion: There was an evidence of high seroprevalence of BVDV among cattle as compared with buffaloes with the possibility of infection in different age groups of animals. In addition, multivariable LR model was proved to provide more information for association and prediction purposes relative to univariable LR models and Chi-square tests if we have more than one predictor.
Andres Sandoval-Hernandez
Factors and conditions that promote academic resilience: A cross-country perspective Objectives or purposes The main objective of this paper is to identify factors and conditions that could help socially disadvantaged students in different countries to become academically resilient. To do that four specific objectives have been set: i) to conceptualize and quantitatively operationalize the notion of academic resilience; then, ii) to estimate the proportion of resilient students across different school systems; iii) to identify the factors and conditions more consistently associated to a high likelihood of academic resilience; and finally, iv) to evaluate the existence of cross-country patterns in the findings listed above. Perspective(s) or theoretical framework Students from low socio-economic status (SES) families live and study in different contexts, and therefore have specific and different educational needs than their more socially advantaged peers. Although it is well documented that students from low SES families tend to perform worse at school, several studies have shown that in most countries there is a group of students who are academically successful despite their challenging backgrounds. These students are called resilient. There is a fairly large body of empirical research on resilience in education; however the discussion of the concept has very often lacked a sound theoretical basis. This paper proposes the Bronfenbrenner's Ecological Systems Theory as a framework to elaborate a theoretical concept and to explain the processes related to academic resilience. Bronfenbrenner suggests that human development processes (e.g. resilience) can be explained in terms of the relationships between individuals and their environment. In resemblance to the hierarchical structure of educational data, under this view the environment consists of different dimensions, or levels, that make up an individual’s context. Methods, techniques or modes of inquiry Then, in order to address our second objective, we operationalized academic resilience based upon the two sine qua non characteristics of a resilient student: a challenging social background and academic success. We used principal component analysis (PCA) to summarize six variables into a SES index (e.g. parent’s level of education, parent’s occupational status, subjective family financial status, home possessions). We then categorized students coming from “challenging backgrounds” as those who score at or below the 20th percentile of the SES index across each group of countries. Finally, we defined “academically successful” students as those who, while controlling for SES, achieved a reading score at or above the 20th percentile in their country. Finally, in order to identify the factors and conditions more consistently associated to a high likelihood of academic resilience, we used different specifications of hierarchical logistic regression models. While controlling for student and socio-demographic characteristics, these models evaluate the association of a theoretically relevant set of variables with the likelihood of academic resilience. Data sources or evidence The data used for the analyses stem from the Progress in International Reading Literacy Study (PIRLS) 2006. PIRLS assessed students’ reading literacy in a target population of 4th graders in 40 participant countries. Apart from reading literacy scores, PIRLS collects extensive information from the pupils, their teachers and head teachers, and their parents, to explore home, school and national influences on student achievement. Results and/or conclusions/points of view Preliminary results suggest that: i) The proportion of resilient students varies considerably across education systems. ii) While generally providing a good fit to the empirical data, from the four dimensions proposed by the Bronfenbrenner’s Model (i.e. personal, family, school and community), the first one seemed to be the most important in predicting a high likelihood of academic resilience; specifically through students’ reading self-concept, and positive attitudes towards reading. In a lesser degree, school and family resources were also among the variables more consistently associated to resilience across countries. iii) Descriptive analyses reveal a faint pattern indicating that variables identified as important in the previous analyses show a stronger association with resilience in less-developed and more unequal countries. Interactions among dimensions, possible policy implications and suggested further research are discussed in the full paper. Educational importance of this study Resilience is important because delivering quality education to all students is a major goal for every education system. As preliminary results suggest, education can indeed play a catalytic role in encouraging that children’s future is not pre-determined by their SES. Understanding the processes involved in academic resilience could provide conceptual and theoretical tools for breaking the intergenerational cycle of poor academic achievement, poor job prospects and poverty. Moreover, an international comparative study of this kind can contribute to establishing a basis for the development of effective policies and practices for promoting resilience across countries. Connection to the themes of the congress Finally, as it is implicit in the theoretical model and in the analytical approach described above, the implications of this work are closely related to the interplay between policy, research and practice. The information used to fit the statistical models was collected from students, teachers, head-teachers, parents and national coordinators; consequently, the discussion and recommendations drawn from the analysis call for coordinated actions of all of them.
Lực Trần Thế
Desalegne Mesa
ABSTRACT Back ground:The term malnutrition, generally, refers both to under nutrition and over nutrition, but in this study theterm is used to refer solely to a deficiency of nutrition.Nutritional status is the result of complex interactions between food consumption and the overall status of health and health care practices. In Ethiopia, 44% of under-five children are stunted while 29% are underweight(EDHS, 2011).Although studies shows good progression in declining proportion of malnourished children in the recent past, there are still problems to be addressed. Objective:Generally, the focus of the study is to identify and examine the correlates of malnutrition of under-five children in the SNNPR regional state. Method:Based on the nature of the response variable, the statistical method employed in this study is ordinal logistic regression model. In other words, the response variable, nutritional status of children under-five years of age, possess the characteristic of ordinal in addition to its multilevel. Therefore, the writer used the ordinal logistic regression method for analysis of data. Results: Results from descriptive statistics show that 37.4% of children in SNNPR state are severely malnourished while 24.2% are moderately malnourished.The findings of the study show that size of child at birth, use of vitamin A during the six months prior to the survey, prenatal treatment of mothers by iron tablet/syrup, education level of mother/father/partner, mothers’ age at first birth, preceding birth interval, sex and age of a child have statistically significant effect on the nutritional status of children under-five years of age. Conclusion: Based on the findings of the studyuse of micronutrients for children andmothers, prevention of early marriage for females, stretched birth spacing, giving due attention for children aged below 11 months and education for mothers are recommended in order totackle the problems related tomalnutrition of under-five children. Key words: Nutritional status, ordinal logistic regression, proportional odds model.
Educational Review
This study reviews the international literature of empirical educational research to examine the application of logistic regression. The aim is to examine common practices of the report and interpretation of logistic regression results, and to discuss the implications for educational research. A review of 130 studies suggests that: (a) the majority of studies report statistical significance and sign of predictors but do not interpret relationship magnitude in terms of probabilities; (b) odds ratio is the most commonly reported effect size, and it tends to be incorrectly interpreted as relative risk, which leads to significant exaggeration of the association magnitude and misleading conclusions; and (c) marginal effects and predicted probabilities are reported by only 10.7% of reviewed studies, and the specification of independent variables’ values is frequently missing. It is suggested that marginal effects and predicted probabilities be reported more frequently to fully utilise the information provided by logistic regression results.
Dale Steele
Research in Higher Education
Alicia C. Dowd , Tarek Coury
Sheryl L Hendriks
The lack of a “gold standard” to determine and predict household food insecurity is well documented. While a considerable volume of research continues to explore universally applicable measurement approaches, robust statistical techniques have not been applied in food security monitoring and early warning systems, especially in countries where food insecurity is chronic. This study explored the application of various Ordinal Logistic Regression techniques in the analysis of national data from South Sudan. Five Link Functions of the Ordinal Regression model were tested. Of these techniques, the Probit Model was found to be the most efficient for predicting food security using ordered categorical outcomes (Food Consumption Scores). The study presents the first rigorous analysis of national food security levels in post conflict South Sudan and shows the power of the model in identifying significant predictors of food insecurity, surveillance, monitoring and early warning.
Loading Preview
Sorry, preview is currently unavailable. You can download the paper by clicking the button above.
Journal of Applied Research in the Community College
Keith Wurtz
Clifford Adelman
Science and Education Development Institute (SEDInst)
Kwesi Nsoah
Sociology Mind
Vernon Loke , Gina Chowa
Dove Medical Press
Venanzio Vella
Texila International Journal , Rajab Muchiri
Educational Psychology Review
DeLeon Gray
Mulugeta Mengstu
Marian Hickendorff
Thomas Debray
WILBERT M G A N G A MTESSIGWA
UNILAG Journal of Medicine, Science and Technology
Ademola Ogunniran
International Review of Education/ …
Journal of College Student Retention Research Theory and Practice
Dheeraj Raju
IAEME Publication
Shlomo S. Sawilowsky
Sunil Bhougal
Youth & Society
D. Shifrer , Chandra Muller
IJESRT Journal
Texila International Journal
Psychometrika
Norman Verhelst
Kennedy Gachigi
Journal of Consumer Psychology
Joseph Verducci
Dissertation
Darby Southgate
Etika Permata
Irene Peniche Ayora
Indira Adhikari
Journal of Women and Minorities in Science and Engineering
Amaury Nora
Dawit G Ayele
Journal of Educational and Behavioral Statistics
Educational Measurement: Issues and Practice
Michael Chajewski , Krista Mattern
Journal of Postgraduate Medicine
S. Kuril , Sunil Karande
Adesupo Akinrefon
Elizabeth Crawford , Gary Wilkerson , David Rausch
International Journal of Pediatric Obesity
Annemien Haveman-Nies
Click through the PLOS taxonomy to find articles in your field.
For more information about PLOS Subject Areas, click here .
Loading metrics
Open Access
Peer-reviewed
Research Article
Roles Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Project administration, Resources, Software, Supervision, Writing – original draft
Affiliation Department of Human Nutrition, Institute of Public Health, University of Gondar, Gondar, Ethiopia
Roles Conceptualization, Formal analysis, Methodology, Software, Writing – original draft, Writing – review & editing
Roles Formal analysis, Software, Writing – review & editing
Roles Conceptualization, Data curation, Formal analysis, Methodology, Software, Writing – original draft, Writing – review & editing
* E-mail: [email protected]
Poor maternal nutrition during pregnancy creates a stressful environment that can lead to long-term effects on tissue development. Understanding the food consumption score can be used to prevent problems associated with poor dietary intake of pregnant mothers. In Ethiopia, the food consumption score ranges from 54% to 81.5%, which is far below the World Food Program (WFP) recommendation. Thus, this study aimed to assess food consumption score and associated factors among pregnant women attending antenatal care services in health centers of Addis Ababa, Ethiopia.
This study has used institution based cross sectional study. Overall, 999 pregnant women were selected for this study. A multistage sampling technique followed by systematic random sampling was used to include pregnant women coming for antenatal care services in the selected health centers of Addis Ababa from June 07 to July 08, 2022. We used interviewer administered questionnaire using the Kobo toolbox. Food consumption score (FCS) was assessed after collecting data on frequency of eight food groups consumed over the previous seven days, which were weighted according to their relative nutritional value. STATA 14 was used to analyse the data. Ordinal logistic regression was used to identify independent predictors of food consumption score. Those variables having p value < 0.25 in the bivariable ordinal logistic regression were considered for the final model. Crude and Adjusted Odds Ratio were used to assess the strength of the association. In the final model, p value < 0.05 at 95% confidence interval was used to declare statistical significance.
From the total of 949 pregnant women a little over half (51.20% (95%CI: 48.00%-54.40%) had acceptable food consumption score, while just over two fifth (42.60% (95% CI: 39.40%-45.70%)) and a small proportion (6.2% (95%CI: 4.84%-7.94%)) of the study participants had borderline and poor food consumption score, respectively. No meal skip (AOR = 1.37, 95% CI:1.03–1.81), able to read and write (AOR = 3.99, 95% CI: 1.33–11.96), poorest wealth status (AOR = 0.52, 95% CI: 0.34–0.78), positive attitude towards consumption of a diversified diet (AOR = 1.52,95% CI: 1.17–1.98) were independent predictors of acceptable food consumption score.
In this study, considerably low level of acceptable food consumption score among the study participants was observed. Besides, not skipping meal, having better educational status, wealth status and attitude towards consumption of a diversified diet were associated with acceptable food consumption score. Therefore, nutritional education considering important dietary modifications should be intensified targeting vulnerable groups.
Citation: Belay JK, Abebe SM, Baffa LD, Mengistu B (2024) Food consumption score and predictors among pregnant women attending antenatal care services in health centers of Addis Ababa, Ethiopia: Using ordinal logistic regression model. PLoS ONE 19(6): e0306169. https://doi.org/10.1371/journal.pone.0306169
Editor: Girma Beressa, Madda Walabu University, ETHIOPIA
Received: February 8, 2023; Accepted: June 12, 2024; Published: June 26, 2024
Copyright: © 2024 Belay et al. This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: all relevant data are within the manuscript and its Supporting Information files.
Funding: The authors received no specific funding for this work.
Competing interests: The authors have declared that no competing interests exist.
Abbreviations: FCS, Food Consumption Score; WFP, World Food Program; ANC, Antenatal Care; COR, Crude Odds Ratio; SD, Standard Deviation; AOR, Adjusted Odds Ratio; VIF, Variance Inflation Factor
A woman’s nutritional requirements vary during pregnancy as she is now feeding both her unborn child and herself. Although prenatal nutrition has an impact on how a pregnancy develops, there is never a wrong moment to start eating healthily. Therefore, it is imperative to have a sound nutrition during the period of gestation for both the mother and her growing foetus [ 1 – 3 ].
However, poor maternal nutrition during pregnancy that is either due to decreased intake or quality results a range of problems [ 4 ]. It affects the general growth and development of the offspring. These changes can have a significant impact on the overall health and production performance of the offspring [ 5 , 6 ]. Along with its negative impacts on the offspring’s nutritional quality, it also produces a stressful environment that may have long-term or permanent repercussions on tissue development, as seen by the emergence of chronic non-communicable diseases later in life [ 7 – 9 ]. Understanding the food consumption score (FCS) of a pregnant woman will help to prevent the issues linked to poor dietary intake during the period of gestation [ 10 ].
Nutritional needs during pregnancy can be satisfied by eating foods from a variety of food groups including fruits, vegetables, dairy products, carbohydrates, fats, and vitamins [ 11 ]. However, poor dietary diversity and FCS have been reported during pregnancy. For example, in Bangladesh acceptable FCS among pregnant women was found to be 58%, different studies in Ethiopia have also revealed a similar figure of FCS among pregnant women:81.5% in East Gojam Zone [ 12 ], and 54% in rural Eastern Ethiopia [ 13 ], which were far below the World Food Program (WFP) recommendations (90%) [ 1 ].
A number of studies have shown the following as independent predictors of having an acceptable FCS during pregnancy: religion [ 12 ], residence [ 12 ], maternal educational status [ 14 ], educational status of the father [ 10 ], wealth status [ 13 , 14 ], attitude [ 13 ], antenatal care (ANC) visit [ 13 ], skipping meal [ 15 ] and consumption of animal source food [ 13 ].
In recent years, introduction of western lifestyles in the big cities of Ethiopia like Addis Ababa has brought a drastic change in food consumption pattern of pregnant women [ 16 ], which runs counter to unrelenting efforts that is outlined in different policies and programmes enacted by the government [ 17 , 18 ]. Socio-cultural factors such as women’s education and employment, food preference, recent epidemics like COVID-19 and cultural practices have also been reported as driving forces for this change [ 19 , 20 ]. In cognizant of this, findings from this study can be used to provide an evidence-based decision to determine factors that influence FCS of pregnant women [ 21 ].
Even though there are a handful of researches that focused on FCS among pregnant women, our study employed a different method-ordinal logistic regression to better understand predictors of FCS among pregnant women [ 22 ]. Thus, this study aimed to assess the food consumption pattern and associated factors among pregnant women attending ANC services in health centers of Addis Ababa, Ethiopia. The goal of this study is to improve the dietary practice of pregnant women, thereby preventing long term ramifications of malnutrition.
The study was conducted in the capital city of Ethiopia, Addis Ababa, it is among the fastest growing cities in Africa. It was estimated that 5,228,000 people reside in the ten sub-cities of Addis Ababa in the study period [ 23 ]. The city has a sub-tropical highland climate, and is populated by people from the different regions of Ethiopia. The magnitude of food insecurity among productive safety net program beneficiaries of the city was 77.10% [ 24 ]. There were six publicly owned general hospitals and one hundred two (102) health centers, and eleven privately owned hospitals and 882 clinics in the city. By using cross-sectional study, pregnant mothers who came for ANC follow up from June 07 to July 08, 2022 at the selected health centers were approached to participate in this study. In these health centers, there were 2478 mothers who came for ANC services.
Sample size was estimated for each specific objective, and the highest was taken for this study. For the first specific objective, by assuming 54.46% proportion of FCS from previous study [ 13 ], 5% margin of error, 1.96 Z value at 95% confidence interval (CI) and by adding 10% non-response rate at 1.5 design effect and it was estimated to be 629. However, the highest sample size was obtained using the second specific objective. Accordingly, epi-info version 7.2.2 was used to estimate the sample size by considering the following assumptions: crude odds ratio of having acceptable FCS among pregnant women who had positive attitude towards consumption a diversified diet, which was 1.6 from a previous study [ 13 ], 80% power and 95% CI, 1.5 designs effect. Therefore, 999 was the final sample size after adding 10% non-response rate.
Pregnant women coming for antenatal care services at the selected health centers were included. However, pregnant women who were seriously ill during the data collection period were excluded in the study. Multistage sampling technique followed by systematic random sampling technique was employed to select the study participants. Out of the ten sub-cities in Addis Ababa, four sub-cities were selected randomly (30%): Nifas silk lafto sub-city, Kolfe keraniyo sub-city, Lideta sub-city and Akaki kality sub-city. In the selected sub-cities, there were 28 health centers. First, nine health centers (one from Nifas silk lafto sub-city, two from Kolfe keraniyo sub-city, three from Lideta sub-city and three Akaki kality sub-city) were selected randomly using a lottery method. Then, the required sample size was proportionated to the selected health centers, and every three (k≈ 2478/999) pregnant woman who was coming for ANC follow up was selected.
Data was collected using pretested interviewer administered questionnaire that comprises socio-demographic data, dietary habits, attitude towards consumption of a diversified diet, obstetric history, and food consumption score (FCS). The questionnaire was first prepared in English and then translated into Amharic (Local language). We used kobo toolbox to collect the data. Nine B.Sc. nurses and four public health officers were the data collectors and supervisors, respectively. The questionnaire was pretested at 5% of the final estimated sample size at Arada sub-city. After the pre-test, the question that assessed participants’ residence was excluded as all the study participants were urban residents. On the food frequency questionnaire, necessary modification was made by including foods that were not previously included.
The outcome variable of this study was food consumption score (FCS), information on foods which were consumed in the last seven days prior to the data collection time was gathered. Food consumption score (FCS) is a composite variable that is constructed based on the following criteria: food frequency, diet diversity and relative nutritional value of each food item [ 1 ]. Food consumption score (FCS) was computed after asking the study participants about the frequency and consumption of eight food groups over the period of seven days prior to the data collection period. In the questionnaire, there were 70 food items which were commonly consumed in the study area. The Cronbach’s alpha value (internal consistency) was observed to be 0.82.
Then, the consumption frequencies were summed and multiplied by the standardized food group weight. Finally, it was categorized into three categories; poor food consumption score(FCS)(0–21), borderline food consumption score(FCS) (21.5–35), and acceptable food consumption score(FCS) (>35) [ 1 , 25 ]. The wealth status was determined using principal component analysis which contained 15 items, and it was later categorized into five categories (Poorest to the richest) [ 23 ]. The attitude of the study participants towards the consumption diversified diet was measured using 4 item Likert-scale questions, the response ranges from strongly disagree to strongly agree. It was considered positive attitude when respondents score above the median. The internal consistency of the questionnaire was checked using Cronbach’s alpha (0.78). The trimesters were defined as first trimester (less than 14 weeks), second trimester (14–27 complete weeks) and third trimester (28 complete weeks until delivery). Finally, birth interval was categorized as recommended birth interval when interpregnancy interval was more 24 months otherwise it was categorized as not recommended birth interval [ 26 ].
The collected data using Kobo toolbox was exported to STATA 14 for analysis. A descriptive data was reported as frequencies, percentage, mean(±SD) and presented in tables. Ordinal logistic regression was used to identify predictors of FCS. Multicollinearity was checked using Variance inflation factor (VIF<10). Brant test of parallel regression assumption (p value = 0.66) conferred proportion of odds assumption. After checking the assumptions of ordinal logistic regression, COR and AOR at 95% was used to ascertain predictors of the outcome variable in both bivariable (p value <0.25) and multivariable ordinal logistic regression respectively. Finally, P value < 0.05 was used to determine level of significance in the final model. The final model reached after checking adequacy of the data using the Hosmer and Lemeshow test.
The study was conducted according to the guidelines of the 1964 Declaration of Helsinki and following amendments. Ethical clearance was obtained from University of Gondar Institutional Review Board of Institute of Public Health (Ref. No IPH/2119/2014). Permission letter was obtained from Addis Ababa Health Office. Written informed consent was obtained from all study participants. Study participants who were unable to read and write signed by fingerprints, while doing so there were two literate witnesses. Data collectors have strictly followed COVID-19 prevention protocols. Confidentiality of the study participants was ensured; no person identifiers were used and the kobo account was password protected-only authorized user was able to access the data.
In this study, 949 pregnant women consented to participate in the study period, yielding 95% response rate. The vast majority of the study participants (96.80%) were married. The mean (±Standard deviation (SD)) age of the study participants in years was 27.16(±4.46SD), about two fifth (39.10%) of the study participants were in the age range 25–29 years. Regarding educational status, half of (50.10%) the study participates had accomplished primary education. More than two fifth (43.90%) of the study participants were housewives. Almost a quarter (24.50%) of the participants were from poor households. More than half (57.60%) of pregnant women have positive attitude towards consumption of variety of food ( Table 1 ).
https://doi.org/10.1371/journal.pone.0306169.t001
As to the maternal characteristics the study participants, more than half (57.6%) were multigravida, almost two third (61.5%) were in the second trimester pregnancy, more than two third (67.02%) had at least one ANC visit, and 69.6% had received nutritional counselling when they came for ANC visit ( Table 2 ).
https://doi.org/10.1371/journal.pone.0306169.t002
Of the study participants, less than half (45.3%) ate three times a day, whereas over half (56.5%) regularly ate snacks. Nearly two thirds (59.2%) skipped meals, with the most common reasons being fatigued or preoccupied with work (19.6%), not wanting to gain weight (19.6%), and other (31.3%) causes such as loss of appetite, vomiting, and discomfort. Likewise, nearly one-third (31.1%) reported a history of food taboos. Lastly, more than a quarter (26.7%) of study participants reported having a history of food cravings ( Table 3 ).
https://doi.org/10.1371/journal.pone.0306169.t003
In this study, practically all of the study participants had consumed common staples, and nearly three quarters (73.2%) of the participants had consumed animal-source food, such as meat ( Table 4 ).
https://doi.org/10.1371/journal.pone.0306169.t004
This study has revealed that a little over half [51.20% (95%CI: 48.00%-54.40)] had acceptable food consumption score. More than two fifth [42.60% (95% CI: 39.45%-45.74%)] had borderline food consumption, and the small proportion [6.2% (95%CI: 4.84%-7.94%) ( Table 5 ).
https://doi.org/10.1371/journal.pone.0306169.t005
Ordinal logistic regression was used to identify factors associated with food consumption score. The following variables which were significant in the bivariable analysis (p value<0.25): age, husband educational status, husband occupation, maternal education, attitude, wealth status, family size, meal skip, food avoid, food craving, taking supplements, still birth, ANC visit, and nutrition counselling during ANC follow-up were fitted in the final model. However, only meal skip, maternal education, attitude and wealth status were found to be the independent predictors of food consumption score.
The odds of having acceptable food consumption score among study participants who can read and write was 3.99 (Relative to borderline and poor food consumption score) times higher than study participants who were unable to read and write [AOR = 3.99,95%CI: 1.33–11.96]. The odds of having acceptable food consumption score were 49% (Relative to borderline and poor food consumption score) lower among study participants who came from the poorest households when compared to participants who came from the richest households [AOR = 0.52, 95%CI: 0.24–0.78]. The odds of having acceptable food consumption score were 1.36 times higher among study participants who did not skip meal (Versus borderline and poor food consumption score) compared to participants who skipped meal [AOR = 1.36, 95%CI: 1.03–1.81]. Finally, among study participants with positive attitude towards consumption of diversified diet there was 52% increased odds to have acceptable food consumption score (Relative to borderline and poor food consumption score) [AOR = 1.52,95%CI: 1.17–1.98] ( Table 6 ).
https://doi.org/10.1371/journal.pone.0306169.t006
This study sought to examine FCS and associated factors among pregnant women who were having ANC follow up in health centers of Addis Ababa, Ethiopia. The results of this study have showed a little over half (51.20%, 95% CI: 48.00%-54.30%) of the study participants had acceptable FCS, and the small proportion of the study participants had poor FCS (6.20%).
Our report was far below the WFP recommendation [ 1 ]. Furthermore, the finding has showed that the percentage of acceptable FCS was comparatively lower than studies from Bangladesh(58%) [ 27 ], Nigeria (80.3%) [ 28 ] and pocket studies from Ethiopia (81.5% and 54.6% at Shegaw Motta and Eastern Ethiopia, respectively) [ 12 , 13 ]. The study period could explain the decreased rate FCS, for example, the study at Shegaw Motta was conducted in the main harvest season while our study was conducted in fasting season when there is a decreased consumption animal source food [ 29 ]. Methodologically, the use of larger sample size in the current study and difference in outcome ascertainment might explain decreased rate of acceptable FCS in this study. In Ethiopia, pregnant women avoid foods due to cultural and religious reasons, and this might explain the discrepancy between the current the study and study from Nigeria where religion and culture has lesser influence over their food choice [ 30 ].
As to the associated factors of FCS, our study has showed maternal educational status-able to read and write, not being in the poorest wealth status, positive attitude towards dietary diversity, and skipping meal were independent predictors of FCS. Those mothers who were able to read and write had higher odds of having acceptable FCS compared to mothers who were unable to read and write, emphasizing the importance of nutritional educational during pregnancy. This was supported by other similar studies conducted in Nigeria [ 30 ], Ghana [ 31 ], and other studies in Ethiopia [ 32 ]. It is evident that increasing level of literacy is crucial to mitigate the problem even in the poorest households [ 33 ]. Besides, mothers who are able to read and write will have a better access to nutritional information from internet, brochures, newspapers and magazines [ 34 – 36 ]. In the affluents, where the toll of non-communicable disease is spiralling- enhancing level of literacy will play a pivotal role for an appropriate food selection and consumption too [ 37 ].
Being in the poorest wealth status decreases the odds of having acceptable FCS by 49% when compared mothers from the richest wealth status. This was also observed in previous studies conducted in Bishoftu, Oromia [ 10 ]. Pregnant mothers from the poorest households have limited economical accesses to procure and buy a diversified diet. On top of this, different studies have pinpointed that being in the lowest wealth status is associated with decreased consumption of animal source food [ 38 ], which in turn results lower FCS. Mothers who did not skip meal had also higher odds of having acceptable FCS when compared to their counterparts. A similar finding was observed from a study in Eastern Ethiopia [ 15 ]. During the period of gestation, meal patterning is highly important since pregnant women who sustain prolonged periods of time without food by skipping meals or snacks may be inducing a physiologic stress in their pregnancy [ 39 ]. Even though accidentally skipping a meal is not going to be harmful, skipping meals regularly for different reasons is not advisable to have a better pregnancy outcome [ 40 , 41 ]. Moreover, from different studies, it has been seen that skipping meals during this period is associated decreased dietary quality [ 15 ].
The study has also revealed, study participants who had positive attitude towards consumption diversified diet had an increased odds of having acceptable FCS than their counter parts. A similar finding was observed from a study conducted in Eastern Ethiopia [ 13 ]. Different researches have supported that pregnant women with increased level of attitude have a better practice of consuming a diversified diet [ 42 ]. Women with positive feeling towards a diversified diet are also motivated to consume foods from different food groups [ 43 , 44 ].
It should be mentioned that the present study has provided greater evidence on the dietary quality and predictors among pregnant women using ordinal logistic regression [ 22 ]. However, methodological limitation of the study cannot go unnoticed. Despite the use of probes like photographs-to recite memory of the study participants-problem of recall bias cannot be ignored which in turn might overestimate or underestimate the result. On top of that, cross- sectional nature of the study limits detection of causal association between the outcome and predicator variables. Even though FCS is a validated tool to asses calorie intake, the tool has not been validated to measure adequacy of macronutrients and micronutrients. The use of a 4 item Likert questionnaire is another limitation of the current study, while recommending the use of a questionnaire with sufficient numbers questions.
The findings of this study can be used to implement public health policies and programmes that strive to bring a better pregnancy outcome by promoting a balanced diet to vulnerable groups of the population. Therefore, to meet the WFP recommendation of having 90% acceptable FCS, interventions need to give a due attention to mothers with lower educational status who are from a lower socio-economic status. The implications of this study can be linked to the importance nutritional educations that target to bring a positive attitude towards consumption of a diversified diet. Moreover, findings of the study imply the importance of provision of a diversified diet in deterring the sequala of malnutrition.
The findings of this study have showed that only half of the study participants had acceptable FCS which is far below the WFP recommendation. Besides, able to read and write, not skipping meal, positive attitude towards the consumption variety of foods, not being from the poorest household were significantly associated with having acceptable FCS relative to borderline and poor FCS. Therefore, it is important to give a special attention to pregnant mothers with low socioeconomic status, and mothers who skip meals in order to enhance their food variety score and improve their nutritional intake.
Future researches are encouraged to investigate nutrient adequacy among pregnant women. Finally, future studies triangulated with qualitative research that investigate behavioural factors such as food taboos and norms that influence FCS among pregnant women are also encouraged.
S1 file. fcs plos one..
https://doi.org/10.1371/journal.pone.0306169.s001
The authors of this article are grateful for the study participants without whom this would not be possible.
Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript.
Nature volume 630 , pages 625–630 ( 2024 ) Cite this article
70k Accesses
1 Citations
1458 Altmetric
Metrics details
Large language model (LLM) systems, such as ChatGPT 1 or Gemini 2 , can show impressive reasoning and question-answering capabilities but often ‘hallucinate’ false outputs and unsubstantiated answers 3 , 4 . Answering unreliably or without the necessary information prevents adoption in diverse fields, with problems including fabrication of legal precedents 5 or untrue facts in news articles 6 and even posing a risk to human life in medical domains such as radiology 7 . Encouraging truthfulness through supervision or reinforcement has been only partially successful 8 . Researchers need a general method for detecting hallucinations in LLMs that works even with new and unseen questions to which humans might not know the answer. Here we develop new methods grounded in statistics, proposing entropy-based uncertainty estimators for LLMs to detect a subset of hallucinations—confabulations—which are arbitrary and incorrect generations. Our method addresses the fact that one idea can be expressed in many ways by computing uncertainty at the level of meaning rather than specific sequences of words. Our method works across datasets and tasks without a priori knowledge of the task, requires no task-specific data and robustly generalizes to new tasks not seen before. By detecting when a prompt is likely to produce a confabulation, our method helps users understand when they must take extra care with LLMs and opens up new possibilities for using LLMs that are otherwise prevented by their unreliability.
‘Hallucinations’ are a critical problem 9 for natural language generation systems using large language models (LLMs), such as ChatGPT 1 or Gemini 2 , because users cannot trust that any given output is correct.
Hallucinations are often defined as LLMs generating “content that is nonsensical or unfaithful to the provided source content” 9 , 10 , 11 but they have come to include a vast array of failures of faithfulness and factuality. We focus on a subset of hallucinations which we call ‘confabulations’ 12 for which LLMs fluently make claims that are both wrong and arbitrary—by which we mean that the answer is sensitive to irrelevant details such as random seed. For example, when asked a medical question “What is the target of Sotorasib?” an LLM confabulates by sometimes answering KRASG12 ‘C’ (correct) and other times KRASG12 ‘D’ (incorrect) despite identical instructions. We distinguish this from cases in which a similar ‘symptom’ is caused by the following different mechanisms: when LLMs are consistently wrong as a result of being trained on erroneous data such as common misconceptions 13 ; when the LLM ‘lies’ in pursuit of a reward 14 ; or systematic failures of reasoning or generalization. We believe that combining these distinct mechanisms in the broad category hallucination is unhelpful. Our method makes progress on a portion of the problem of providing scalable oversight 15 by detecting confabulations that people might otherwise find plausible. However, it does not guarantee factuality because it does not help when LLM outputs are systematically bad. Nevertheless, we significantly improve question-answering accuracy for state-of-the-art LLMs, revealing that confabulations are a great source of error at present.
We show how to detect confabulations by developing a quantitative measure of when an input is likely to cause an LLM to generate arbitrary and ungrounded answers. Detecting confabulations allows systems built on LLMs to avoid answering questions likely to cause confabulations, to make users aware of the unreliability of answers to a question or to supplement the LLM with more grounded search or retrieval. This is essential for the critical emerging field of free-form generation in which naive approaches, suited to closed vocabulary and multiple choice, fail. Past work on uncertainty for LLMs has focused on simpler settings, such as classifiers 16 , 17 and regressors 18 , 19 , whereas the most exciting applications of LLMs relate to free-form generations.
The term hallucination in the context of machine learning originally comes from filling in ungrounded details, either as a deliberate strategy 20 or as a reliability problem 4 . The appropriateness of the metaphor has been questioned as promoting undue anthropomorphism 21 . Although we agree that metaphor must be used carefully with LLMs 22 , the widespread adoption of the term hallucination reflects the fact that it points to an important phenomenon. This work represents a step towards making that phenomenon more precise.
To detect confabulations, we use probabilistic tools to define and then measure the ‘semantic’ entropy of the generations of an LLM—an entropy that is computed over meanings of sentences. High entropy corresponds to high uncertainty 23 , 24 , 25 —so semantic entropy is one way to estimate semantic uncertainties. Semantic uncertainty, the broader category of measures we introduce, could be operationalized with other measures of uncertainty, such as mutual information, instead. Entropy in free-form generation is normally hard to measure because answers might mean the same thing (be semantically equivalent) despite being expressed differently (being syntactically or lexically distinct). This causes naive estimates of entropy or other lexical variation scores 26 to be misleadingly high when the same correct answer might be written in many ways without changing its meaning.
By contrast, our semantic entropy moves towards estimating the entropy of the distribution of meanings of free-form answers to questions, insofar as that is possible, rather than the distribution over the ‘tokens’ (words or word-pieces) which LLMs natively represent. This can be seen as a kind of semantic consistency check 27 for random seed variation. An overview of our approach is provided in Fig. 1 and a worked example in Supplementary Table 1 .
a , Naive entropy-based uncertainty measures variation in the exact answers, treating ‘Paris’, ‘It’s Paris’ and ‘France’s capital Paris’ as different. But this is unsuitable for language tasks for which sometimes different answers mean the same things. Our semantic entropy clusters answers which share meanings before computing the entropy. A low semantic entropy shows that the LLM is confident about the meaning. b , Semantic entropy can also detect confabulations in longer passages. We automatically decompose a long generated answer into factoids. For each factoid, an LLM generates questions to which that factoid might have been the answer. The original LLM then samples M possible answers to these questions. Finally, we compute the semantic entropy over the answers to each specific question, including the original factoid. Confabulations are indicated by high average semantic entropy for questions associated with that factoid. Here, semantic entropy classifies Fact 1 as probably not a confabulation because generations often mean the same thing, despite very different wordings, which a naive entropy would have missed.
Intuitively, our method works by sampling several possible answers to each question and clustering them algorithmically into answers that have similar meanings, which we determine on the basis of whether answers in the same cluster entail each other bidirectionally 28 . That is, if sentence A entails that sentence B is true and vice versa, then we consider them to be in the same semantic cluster. We measure entailment using both general-purpose LLMs and natural language inference (NLI) tools developed specifically for detecting entailment for which we show direct evaluations in Supplementary Tables 2 and 3 and Supplementary Fig. 1 . Textual entailment has previously been shown to correlate with faithfulness 10 in the context of factual consistency 29 as well as being used to measure factuality in abstractive summarization 30 , especially when applied at the right granularity 31 .
Semantic entropy detects confabulations in free-form text generation across a range of language models and domains, without previous domain knowledge. Our evaluations cover question answering in trivia knowledge (TriviaQA 32 ), general knowledge (SQuAD 1.1; ref. 33 ), life sciences (BioASQ 34 ) and open-domain natural questions (NQ-Open 35 ) derived from actual queries to Google Search 36 . In addition, semantic entropy detects confabulations in mathematical word problems (SVAMP 37 ) and in a biography-generation dataset, FactualBio, accompanying this paper.
Our results for TriviaQA, SQuAD, BioASQ, NQ-Open and SVAMP are all evaluated context-free and involve sentence-length answers (96 ± 70 characters, mean ± s.d.) and use LLaMA 2 Chat (7B, 13B and 70B parameters) 38 , Falcon Instruct (7B and 40B) 39 and Mistral Instruct (7B) 40 . In the Supplementary Information , we further consider short-phrase-length answers. Results for FactualBio (442 ± 122 characters) use GPT-4 (ref. 1 ). At the time of writing, GPT-4 (ref. 1 ) did not expose output probabilities 41 or hidden states, although it does now. As a result, we propose a discrete approximation of our estimator for semantic entropy which allows us to run experiments without access to output probabilities, which we use for all GPT-4 results in this paper and which performs similarly well.
Our confabulation detection with semantic entropy is more robust to user inputs from previously unseen domains than methods which aim to ‘learn’ how to detect confabulations from a set of example demonstrations. Our method is unsupervised, meaning that we do not need labelled examples of confabulations. By contrast, supervised methods detect confabulations by learning patterns behind examples of confabulations, assuming that future questions preserve these patterns. But this assumption is often untrue in new situations or with confabulations that human overseers are unable to identify (compare Fig. 17 of ref. 24 ). As a strong supervised baseline, we compare to an embedding regression method inspired by ref. 24 which trains a logistic regression classifier to predict whether the model correctly answered a question on the basis of the final ‘embedding’ (hidden state) of the LLM. We also use the P (True) method 24 which looks at the probability with which an LLM predicts that the next token is ‘True’ when few-shot prompted to compare a main answer with ‘brainstormed’ alternatives.
Confabulations contribute substantially to incorrect answers given by language models. We show that semantic entropy can be used to predict many incorrect model answers and to improve question-answering accuracy by refusing to answer those questions the model is uncertain about. Corresponding to these two uses, we evaluate two main metrics. First, the widely used area under the receiver operating characteristic (AUROC) curve for the binary event that a given answer is incorrect. This measure captures both precision and recall and ranges from 0 to 1, with 1 representing a perfect classifier and 0.5 representing an un-informative classifier. We also show a new measure, the area under the ‘rejection accuracy’ curve (AURAC). This studies the case in which the confabulation detection score is used to refuse to answer the questions judged most likely to cause confabulations. Rejection accuracy is the accuracy of the answers of the model on the remaining questions and the area under this curve is a summary statistic over many thresholds (representative threshold accuracies are provided in Supplementary Material ). The AURAC captures the accuracy improvement which users would experience if semantic entropy was used to filter out questions causing the highest entropy.
In Fig. 2 , we show that both semantic entropy and its discrete approximation outperform our best baselines for sentence-length generations. These results are averaged across datasets and provide the actual scores on the held-out evaluation dataset. We report the raw average score across held-out evaluation datasets without standard error because the distributional characteristics are more a property of the models and datasets selected than the method. Consistency of relative results across different datasets is a stronger indicator of variation in this case.
Semantic entropy outperforms leading baselines and naive entropy. AUROC (scored on the y -axes) measures how well methods predict LLM mistakes, which correlate with confabulations. AURAC (likewise scored on the y -axes) measures the performance improvement of a system that refuses to answer questions which are judged likely to cause confabulations. Results are an average over five datasets, with individual metrics provided in the Supplementary Information .
Semantic entropy greatly outperforms the naive estimation of uncertainty using entropy: computing the entropy of the length-normalized joint probability of the token sequences. Naive entropy estimation ignores the fact that token probabilities also express the uncertainty of the model over phrasings that do not change the meaning of an output.
Our methods also outperform the supervised embedding regression method both in- and out-of-distribution. In pale-yellow bars we show that embedding regression performance deteriorates when its training data do not match the deployment distribution—which mirrors the common real-world case in which there is a distribution shift between training and deployment 42 —the plotted value is the average metric for embedding regression trained on one of the four ‘off-distribution’ datasets for that evaluation. This is critical because reliable uncertainty is most important when the data distribution shifts. Semantic entropy also outperforms P (True) which is supervised ‘in-context’; that is, it is adapted to the deployment task with a few training examples provided in the LLM prompt itself. The discrete variant of semantic entropy performs similarly to our standard estimator, despite not requiring exact output probabilities.
Averaged across the 30 combinations of tasks and models we study, semantic entropy achieves the best AUROC value of 0.790 whereas naive entropy (0.691), P (True) (0.698) and the embedding regression baseline (0.687) lag behind it. Semantic entropy performs well consistently, with stable performance (between 0.78 and 0.81 AUROC) across the different model families (LLaMA, Falcon and Mistral) and scales (from 7B to 70B parameters) which we study (we report summary statistics for each dataset and model as before). Although semantic entropy outperforms the baselines across all model sizes, P (True) seems to improve with model size, suggesting that it might become more competitive for very capable honest models in settings that the model understands well (which are, however, not the most important cases to have good uncertainty). We use ten generations to compute entropy, selected using analysis in Supplementary Fig. 2 . Further results for short-phrase generations are described in Supplementary Figs. 7 – 10 .
The results in Fig. 2 offer a lower bound on the effectiveness of semantic entropy at detecting confabulations. These evaluations determine whether semantic entropy and baseline methods can detect when the answers of the model are incorrect (which we validate against human correctness evaluations in Supplementary Table 4 ). In addition to errors from confabulations (arbitrary incorrectness), this also includes other types of mistakes for which semantic entropy is not suited, such as consistent errors learned from the training data. The fact that methods such as embedding regression are able to spot other kinds of errors, not just confabulations, but still are outperformed by semantic entropy, suggests that confabulations are a principal category of errors for actual generations.
Examples of questions and answers from TriviaQA, SQuAD and BioASQ, for LLaMA 2 Chat 70B, are shown in Table 1 . These illustrate how only semantic entropy detects when the meaning is constant but the form varies (the first row of the table) whereas semantic entropy and naive entropy both correctly predict the presence of confabulations when the form and meaning vary together (second row) and predict the absence of confabulations when the form and meaning are both constant across several resampled generations (third row). In the final row, we give an example in which semantic entropy is erroneously high as a result of overly sensitive semantic clustering relative to the reference answer. Our clustering method distinguishes the answers which provide a precise date from those which only provide a year. For some contexts that would have been correct but in this context the distinction between the specific day and the year is probably irrelevant. This highlights the importance of context and judgement in clustering, especially in subtle cases, as well as the shortcomings of evaluating against fixed reference answers which do not capture the open-ended flexibility of conversational deployments of LLMs.
Semantic entropy is most natural for sentences that express a single proposition but the idea of semantic equivalence is trickier to apply to longer passages which express many propositions which might only agree partially 43 . Nevertheless, we can use semantic entropy to detect confabulations in longer generations, such as entire paragraphs of text. To show this, we develop a dataset of biographical generations from GPT-4 (v.0613) for 21 individuals notable enough to have their own Wikipedia page but without extensive online biographies. From each biography generated by GPT-4, we automatically extract propositional factual claims about the individual (150 factual claims in total), which we manually label as true or false.
Applying semantic entropy to this problem is challenging. Naively, one might simply regenerate each sentence (conditioned on the text so far) and then compute semantic entropy over these regenerations. However, the resampled sentences often target different aspects of the biography: for example, one time describing family and the next time profession. This is analogous to the original problem semantic entropy was designed to resolve: the model is uncertain about the right ordering of facts, not about the facts themselves. To address this, we break down the entire paragraph into factual claims and reconstruct questions which might have been answered by those claims. Only then do we apply semantic entropy (Fig. 1 ) by generating three new answers to each question (selected with analysis in Supplementary Figs. 3 and 4 ) and computing the semantic entropy over those generations plus the original factual claim. We aggregate these by averaging the semantic entropy over all the questions to get an uncertainty score for each proposition, which we use to detect confabulations. Unaggregated results are shown in Supplementary Figs. 5 and 6 .
As GPT-4 did not allow access to the probability of the generation at the time of writing, we use a discrete variant of semantic entropy which makes the further approximation that we can infer a discrete empirical distribution over semantic meaning clusters from only the generations ( Methods ). This allows us to compute semantic entropy using only the black-box outputs of an LLM. However, we were unable to compute the naive entropy baseline, the standard semantic entropy estimator or the embedding regression baseline for GPT-4 without output probabilities and embeddings.
In Fig. 3 we show that the discrete variant of semantic entropy effectively detects confabulations on this dataset. Its AUROC and AURAC are higher than either a simple ‘self-check’ baseline—which just asks the LLM whether the factoid is likely to be true—or a variant of P (True) which has been adapted to work for the paragraph-length setting. Discrete semantic entropy has better rejection accuracy performance until 20% of the questions have been rejected at which point P (True) has a narrow edge. This indicates that the questions predicted to cause confabulations are indeed more likely to be wrong.
The discrete variant of our semantic entropy estimator outperforms baselines both when measured by AUROC and AURAC metrics (scored on the y -axis). The AUROC and AURAC are substantially higher than for both baselines. At above 80% of questions being answered, semantic entropy has the highest accuracy. Only when the top 20% of answers judged most likely to be confabulations are rejected does the answer accuracy on the remainder for the P (True) baseline exceed semantic entropy.
Our probabilistic approach, accounting for semantic equivalence, detects an important class of hallucinations: those that are caused by a lack of LLM knowledge. These are a substantial portion of the failures at present and will continue even as models grow in capabilities because situations and cases that humans cannot reliably supervise will persist. Confabulations are a particularly noteworthy failure mode for question answering but appear in other domains too. Semantic entropy needs no previous domain knowledge and we expect that algorithmic adaptations to other problems will allow similar advances in, for example, abstractive summarization. In addition, extensions to alternative input variations such as rephrasing or counterfactual scenarios would allow a similar method to act as a form of cross-examination 44 for scalable oversight through debate 45 .
The success of semantic entropy at detecting errors suggests that LLMs are even better at “knowing what they don’t know” than was argued by ref. 24 —they just don’t know they know what they don’t know. Our method explicitly does not directly address situations in which LLMs are confidently wrong because they have been trained with objectives that systematically produce dangerous behaviour, cause systematic reasoning errors or are systematically misleading the user. We believe that these represent different underlying mechanisms—despite similar ‘symptoms’—and need to be handled separately.
One exciting aspect of our approach is the way it makes use of classical probabilistic machine learning methods and adapts them to the unique properties of modern LLMs and free-form language generation. We hope to inspire a fruitful exchange of well-studied methods and emerging new problems by highlighting the importance of meaning when addressing language-based machine learning problems.
Semantic entropy as a strategy for overcoming confabulation builds on probabilistic tools for uncertainty estimation. It can be applied directly to any LLM or similar foundation model without requiring any modifications to the architecture. Our ‘discrete’ variant of semantic uncertainty can be applied even when the predicted probabilities for the generations are not available, for example, because access to the internals of the model is limited.
In this section we introduce background on probabilistic methods and uncertainty in machine learning, discuss how it applies to language models and then discuss our contribution, semantic entropy, in detail.
We aim to detect confabulations in LLMs, using the principle that the model will be uncertain about generations for which its output is going to be arbitrary.
One measure of uncertainty is the predictive entropy of the output distribution, which measures the information one has about the output given the input 25 . The predictive entropy (PE) for an input sentence x is the conditional entropy ( H ) of the output random variable Y with realization y given x ,
A low predictive entropy indicates an output distribution which is heavily concentrated whereas a high predictive entropy indicates that many possible outputs are similarly likely.
We do not distinguish between aleatoric and epistemic uncertainty in our analysis. Researchers sometimes separate aleatoric uncertainty (uncertainty in the underlying data distribution) from epistemic uncertainty (caused by having only limited information) 46 . Further advances in uncertainty estimation which separate these kinds of uncertainty would enhance the potential for our semantic uncertainty approach by allowing extensions beyond entropy.
Generative LLMs produce strings of text by selecting tokens in sequence. Each token is a wordpiece that often represents three or four characters (though especially common sequences and important words such as numbers typically get their own token). To compute entropies, we need access to the probabilities the LLM assigns to the generated sequence of tokens. The probability of the entire sequence, s , conditioned on the context, x , is the product of the conditional probabilities of new tokens given past tokens, whose resulting log-probability is \(\log P({\bf{s}}| {\boldsymbol{x}})={\sum }_{i}\log P({s}_{i}| {{\bf{s}}}_{ < i},{\boldsymbol{x}})\) , where s i is the i th output token and s < i denotes the set of previous tokens.
When comparing the log-probabilities of generated sequences, we use ‘length normalization’, that is, we use an arithmetic mean log-probability, \(\frac{1}{N}{\sum }_{i}^{N}\log P({s}_{i}| {{\bf{s}}}_{ < i},{\boldsymbol{x}})\) , instead of the sum. In expectation, longer sequences have lower joint likelihoods because of the conditional independence of the token probabilities 47 . The joint likelihood of a sequence of length N shrinks exponentially in N . Its negative log-probability therefore grows linearly in N , so longer sentences tend to contribute more to entropy. We therefore interpret length-normalizing the log-probabilities when estimating the entropy as asserting that the expected uncertainty of generations is independent of sentence length. Length normalization has some empirical success 48 , including in our own preliminary experiments, but little theoretical justification in the literature.
If we naively calculate the predictive entropy directly from the probabilities of the generated sequence of tokens, we conflate the uncertainty of the model over the meaning of its answer with the uncertainty over the exact tokens used to express that meaning. For example, even if the model is confident in the meaning of a generation, there are still usually many different ways for phrasing that generation without changing its meaning. For the purposes of detecting confabulations, the uncertainty of the LLM over meanings is more important than the uncertainty over the exact tokens used to express those meanings.
Our semantic uncertainty method therefore seeks to estimate only the uncertainty the LLM has over the meaning of its generation, not the choice of words. To do this, we introduce an algorithm that clusters model generations by meaning and subsequently calculates semantic uncertainty. At a high level this involves three steps:
Generation: sample output sequences of tokens from the predictive distribution of a LLM given a context x .
Clustering: cluster sequences by their meaning using our clustering algorithm based on bidirectional entailment.
Entropy estimation: estimate semantic entropy by summing probabilities of sequences that share a meaning following equation ( 2 ) and compute their entropy.
Given some context x as input to the LLM, we sample M sequences, { s (1) , …, s ( M ) } and record their token probabilities, { P ( s (1) ∣ x ), …, P ( s ( M ) ∣ x )}. We sample all our generations from a single model, varying only the random seed used for sampling from the token probabilities. We do not observe the method to be particularly sensitive to details of the sampling scheme. In our implementation, we sample at temperature 1 using nucleus sampling ( P = 0.9) (ref. 49 ) and top- K sampling ( K = 50) (ref. 50 ). We also sample a single generation at low temperature (0.1) as an estimate of the ‘best generation’ of the model to the context, which we use to assess the accuracy of the model. (A lower sampling temperature increases the probability of sampling the most likely tokens).
To estimate semantic entropy we need to cluster generated outputs from the model into groups of outputs that mean the same thing as each other.
This can be described using ‘semantic equivalence’ which is the relation that holds between two sentences when they mean the same thing. We can formalize semantic equivalence mathematically. Let the space of tokens in a language be \({\mathcal{T}}\) . The space of all possible sequences of tokens of length N is then \({{\mathcal{S}}}_{N}\equiv {{\mathcal{T}}}^{N}\) . Note that N can be made arbitrarily large to accommodate whatever size of sentence one can imagine and one of the tokens can be a ‘padding’ token which occurs with certainty for each token after the end-of-sequence token. For some sentence \({\bf{s}}\in {{\mathcal{S}}}_{N}\) , composed of a sequence of tokens, \({s}_{i}\in {\mathcal{T}}\) , there is an associated meaning. Theories of meaning are contested 51 . However, for specific models and deployment contexts many considerations can be set aside. Care should be taken comparing very different models and contexts.
Let us introduce a semantic equivalence relation, E ( ⋅ , ⋅ ), which holds for any two sentences that mean the same thing—we will operationalize this presently. Recall that an equivalence relation is any reflexive, symmetric and transitive relation and that any equivalence relation on a set corresponds to a set of equivalence classes. Each semantic equivalence class captures outputs that can be considered to express the same meaning. That is, for the space of semantic equivalence classes \({\mathcal{C}}\) the sentences in the set \(c\in {\mathcal{C}}\) can be regarded in many settings as expressing a similar meaning such that \(\forall {\bf{s}},{{\bf{s}}}^{{\prime} }\in c:E({\bf{s}},{{\bf{s}}}^{{\prime} })\) . So we can build up these classes of semantically equivalent sentences by checking if new sentences share a meaning with any sentences we have already clustered and, if so, adding them into that class.
We operationalize E ( ⋅ , ⋅ ) using the idea of bidirectional entailment, which has a long history in linguistics 52 and natural language processing 28 , 53 , 54 . A sequence, s , means the same thing as a second sequence, s ′, only if the sequences entail (that is, logically imply) each other. For example, ‘The capital of France is Paris’ entails ‘Paris is the capital of France’ and vice versa because they mean the same thing. (See later for a discussion of soft equivalence and cases in which bidirectional entailment does not guarantee equivalent meanings).
Importantly, we require that the sequences mean the same thing with respect to the context—key meaning is sometimes contained in the context. For example, ‘Paris’ does not entail ‘The capital of France is Paris’ because ‘Paris’ is not a declarative sentence without context. But in the context of the question ‘What is the capital of France?’, the one-word answer does entail the longer answer.
Detecting entailment has been the object of study of a great deal of research in NLI 55 . We rely on language models to predict entailment, such as DeBERTa-Large-MNLI 56 , which has been trained to predict entailment, or general-purpose LLMs such as GPT-3.5 (ref. 57 ), which can predict entailment given suitable prompts.
We then cluster sentences according to whether they bidirectionally entail each other using the algorithm presented in Extended Data Fig. 1 . Note that, to check if a sequence should be added to an existing cluster, it is sufficient to check if the sequence bidirectionally entails any of the existing sequences in that cluster (we arbitrarily pick the first one), given the transitivity of semantic equivalence. If a sequence does not share meaning with any existing cluster, we assign it its own cluster.
Having determined the classes of generated sequences that mean the same thing, we can estimate the likelihood that a sequence generated by the LLM belongs to a given class by computing the sum of the probabilities of all the possible sequences of tokens which can be considered to express the same meaning as
Formally, this treats the output as a random variable whose event-space is the space of all possible meaning-classes, C , a sub- σ -algebra of the standard event-space S . We can then estimate the semantic entropy (SE) as the entropy over the meaning-distribution,
There is a complication which prevents direct computation: we do not have access to every possible meaning-class c . Instead, we can only sample c from the sequence-generating distribution induced by the model. To handle this, we estimate the expectation in equation ( 3 ) using a Rao–Blackwellized Monte Carlo integration over the semantic equivalence classes C ,
where \(P({C}_{i}| {\boldsymbol{x}})=\frac{P({c}_{i}| {\boldsymbol{x}})}{{\sum }_{c}P(c| {\boldsymbol{x}})}\) estimates a categorical distribution over the cluster meanings, that is, ∑ i P ( C i ∣ x ) = 1. Without this normalization step cluster ‘probabilities’ could exceed one because of length normalization, resulting in degeneracies. Equation ( 5 ) is the estimator giving our main method that we refer to as semantic entropy throughout the text.
For scenarios in which the sequence probabilities are not available, we propose a variant of semantic entropy which we call ‘discrete’ semantic entropy. Discrete semantic entropy approximates P ( C i ∣ x ) directly from the number of generations in each cluster, disregarding the token probabilities. That is, we approximate P ( C i ∣ x ) as \({\sum }_{1}^{M}\frac{{I}_{c={C}_{i}}}{M}\) , the proportion of all the sampled answers which belong to that cluster. Effectively, this just assumes that each output that was actually generated was equally probable—estimating the underlying distribution as the categorical empirical distribution. In the limit of M the estimator converges to equation ( 5 ) by the law of large numbers. We find that discrete semantic entropy results in similar performance empirically.
We provide a worked example of the computation of semantic entropy in Supplementary Note 1 .
Semantic entropy is designed to detect confabulations, that is, model outputs with arbitrary meaning. In our experiments, we use semantic uncertainty to predict model accuracy, demonstrating that confabulations make up a notable fraction of model mistakes. We further show that semantic uncertainty can be used to improve model accuracy by refusing to answer questions when semantic uncertainty is high. Last, semantic uncertainty can be used to give users a way to know when model generations are probably unreliable.
We use the datasets BioASQ 34 , SQuAD 33 , TriviaQA 32 , SVAMP 37 and NQ-Open 35 . BioASQ is a life-sciences question-answering dataset based on the annual challenge of the same name. The specific dataset we use is based on the QA dataset from Task B of the 2023 BioASQ challenge (11B). SQuAD is a reading comprehension dataset whose context passages are drawn from Wikipedia and for which the answers to questions can be found in these passages. We use SQuAD 1.1 which excludes the unanswerable questions added in v.2.0 that are deliberately constructed to induce mistakes so they do not in practice cause confabulations to occur. TriviaQA is a trivia question-answering dataset. SVAMP is a word-problem maths dataset containing elementary-school mathematical reasoning tasks. NQ-Open is a dataset of realistic questions aggregated from Google Search which have been chosen to be answerable without reference to a source text. For each dataset, we use 400 train examples and 400 test examples randomly sampled from the original larger dataset. Note that only some of the methods require training, for example semantic entropy does not use the training data. If the datasets themselves are already split into train and test (or validation) samples, we sample our examples from within the corresponding split.
All these datasets are free-form, rather than multiple choice, because this better captures the opportunities created by LLMs to produce free-form sentences as answers. We refer to this default scenario as our ‘sentence-length’ experiments. In Supplementary Note 7 , we also present results for confabulation detection in a ‘short-phrase’ scenario, in which we constrain model answers on these datasets to be as concise as possible.
To make the problems more difficult and induce confabulations, we do not provide the context passages for any of the datasets. When the context passages are provided, the accuracy rate is too high for these datasets for the latest generations of models to meaningfully study confabulations.
For sentence-length generations we use: Falcon 39 Instruct (7B and 40B), LLaMA 2 Chat 38 (7B, 13B and 70B) and Mistral 40 Instruct (7B).
In addition to reporting results for semantic entropy, discrete semantic entropy and naive entropy, we consider two strong baselines.
Embedding regression is a supervised baseline inspired by the P (IK) method 24 . In that paper, the authors fine-tune their proprietary LLM on a dataset of questions to predict whether the model would have been correct. This requires access to a dataset of ground-truth answers to the questions. Rather than fine-tuning the entire LLM in this way, we simply take the final hidden units and train a logistic regression classifier to make the same prediction. By contrast to their method, this is much simpler because it does not require fine-tuning the entire language model, as well as being more reproducible because the solution to the logistic regression optimization problem is not as seed-dependent as the fine-tuning procedure. As expected, this supervised approach performs well in-distribution but fails when the distribution of questions is different from that on which the classifier is trained.
The second baseline we consider is the P (True) method 24 , in which the model first samples M answers (identically to our semantic entropy approach) and then is prompted with the list of all answers generated followed by the highest probability answer and a question whether this answer is “(a) True” or “(b) False”. The confidence score is then taken to be the probability with which the LLM responds with ‘a’ to the multiple-choice question. The performance of this method is boosted with a few-shot prompt, in which up to 20 examples from the training set are randomly chosen, filled in as above, but then provided with the actual ground truth of whether the proposed answer was true or false. In this way, the method can be considered as supervised ‘in-context’ because it makes use of some ground-truth training labels but can be used without retraining the model. Because of context-size constraints, this method cannot fit a full 20 few-shot examples in the context when input questions are long or large numbers of generations are used. As a result, we sometimes have to reduce the number of few-shot examples to suit the context size and we note this in the Supplementary Material .
Any NLI classification system could be used for our bidirectional entailment clustering algorithm. We consider two different kinds of entailment detector.
One option is to use an instruction-tuned LLM such as LLaMA 2, GPT-3.5 (Turbo 1106) or GPT-4 to predict entailment between generations. We use the following prompt:
We are evaluating answers to the question {question} Here are two possible answers: Possible Answer 1: {text1} Possible Answer 2: {text2} Does Possible Answer 1 semantically entail Possible Answer 2? Respond with entailment, contradiction, or neutral.
Alternatively, we consider using a language model trained for entailment prediction, specifically the DeBERTa-large model 56 fine-tuned on the NLI dataset MNLI 58 . This builds on past work towards paraphrase identification based on embedding similarity 59 , 60 and BERT-style models 61 , 62 . We template more simply, checking if DeBERTa predicts entailment between the concatenation of the question and one answer and the concatenation of the question and another answer. Note that DeBERTa-large is a relatively lightweight model with only 1.5B parameters which is much less powerful than most of the LLMs under study.
In Supplementary Note 2 , we carefully evaluate the benefits and drawbacks of these methods for entailment prediction. We settle on using GPT-3.5 with the above prompt, as its entailment predictions agree well with human raters and lead to good confabulation detection performance.
In Supplementary Note 3 , we provide a discussion of the computational cost and choosing the number of generations for reliable clustering.
We use a simple generation template for all sentence-length answer datasets:
Answer the following question in a single brief but complete sentence. Question: {question} Answer:
We use three main metrics to evaluate our method: AUROC, rejection accuracy and AURAC. Each of these is grounded in an automated factuality estimation measurement relative to the reference answers provided by the datasets that we use.
First, we use the AUROC curve, which measures the reliability of a classifier accounting for both precision and recall. The AUROC can be interpreted as the probability that a randomly chosen correct answer has been assigned a higher confidence score than a randomly chosen incorrect answer. For a perfect classifier, this is 1.
Second, we compute the ‘rejection accuracy at X %’, which is the question-answering accuracy of the model on the most-confident X % of the inputs as identified by the respective uncertainty method. If an uncertainty method works well, predictions on the confident subset should be more accurate than predictions on the excluded subset and the rejection accuracy should increase as we reject more inputs.
To summarize this statistic we compute the AURAC—the total area enclosed by the accuracies at all cut-off percentages X %. This should increase towards 1 as given uncertainty method becomes more accurate and better at detecting likely-inaccurate responses but it is more sensitive to the overall accuracy of the model than the AUROC metric.
In Supplementary Note 5 , we provide the unaggregated rejection accuracies for sentence-length generations.
For the short-phrase-length generation setting presented in Supplementary Note 7 , we simply assess the accuracy of the generations by checking if the F1 score of the commonly used SQuAD metric exceeds 0.5. There are limitations to such simple scoring rules 63 but this method is widely used in practice and its error is comparatively small on these standard datasets.
For our default scenario, the longer sentence-length generations, this measure fails, as the overlap between the short reference answer and our long model answer is invariably too small. For sentence-length generations, we therefore automatically determine whether an answer to the question is correct or incorrect by using GPT-4 to compare the given answer to the reference answer. We use the template:
We are assessing the quality of answers to the following question: {question} The expected answer is: {reference answer} The proposed answer is: {predicted answer} Within the context of the question, does the proposed answer mean the same as the expected answer? Respond only with yes or no.
We make a small modification for datasets with several reference answers: line two becomes “The following are expected answers to this question:” and the final line asks “does the proposed answer mean the same as any of the expected answers?”.
In Supplementary Note 6 , we check the quality of our automated ground-truth evaluations against human judgement by hand. We find that GPT-4 gives the best results for determining model accuracy and thus use it in all our sentence-length experiments.
In this section we describe the application of semantic entropy to confabulation detection in longer model generations, specifically paragraph-length biographies.
We introduce a biography-generation dataset—FactualBio—available alongside this paper. FactualBio is a collection of biographies of individuals who are notable enough to have Wikipedia pages but not notable enough to have large amounts of detailed coverage, generated by GPT-4 (v.0613). To generate the dataset, we randomly sampled 21 individuals from the WikiBio dataset 64 . For each biography, we generated a list of factual claims contained in each biography using GPT-4, with 150 total factual claims (the total number is only coincidentally a round number). For each of these factual claims, we manually determined whether the claim was correct or incorrect. Out of 150 claims, 45 were incorrect. As before, we apply confabulation detection to detect incorrect model predictions, even though there may be model errors which are not confabulations.
Given a paragraph-length piece of LLM-generated text, we apply the following sequence of steps:
Automatically decompose the paragraph into specific factual claims using an LLM (not necessarily the same as the original).
For each factual claim, use an LLM to automatically construct Q questions which might have produced that claim.
For each question, prompt the original LLM to generate M answers.
For each question, compute the semantic entropy of the answers, including the original factual claim.
Average the semantic entropies over the questions to arrive at a score for the original factual claim.
We pursue this slightly indirect way of generating answers because we find that simply resampling each sentence creates variation unrelated to the uncertainty of the model about the factual claim, such as differences in paragraph structure.
We decompose the paragraph into factual claims using the following prompt:
Please list the specific factual propositions included in the answer above. Be complete and do not leave any factual claims out. Provide each claim as a separate sentence in a separate bullet point.
We found that we agreed with the decompositions in all cases in the dataset.
We then generate six questions for each of the facts from the decomposition. We generate these questions by prompting the model twice with the following:
Following this text: {text so far} You see the sentence: {proposition} Generate a list of three questions, that might have generated the sentence in the context of the preceding original text, as well as their answers. Please do not use specific facts that appear in the follow-up sentence when formulating the question. Make the questions and answers diverse. Avoid yes-no questions. The answers should not be a full sentence and as short as possible, e.g. only a name, place, or thing. Use the format “1. {question} – {answer}”.
These questions are not necessarily well-targeted and the difficulty of this step is the main source of errors in the procedure. We generate three questions with each prompt, as this encourages diversity of the questions, each question targeting a different aspect of the fact. However, we observed that the generated questions will sometimes miss obvious aspects of the fact. Executing the above prompt twice (for a total of six questions) can improve coverage. We also ask for brief answers because the current version of GPT-4 tends to give long, convoluted and highly hedged answers unless explicitly told not to.
Then, for each question, we generate three new answers using the following prompt:
We are writing an answer to the question “{user question}”. So far we have written: {text so far} The next sentence should be the answer to the following question: {question} Please answer this question. Do not answer in a full sentence. Answer with as few words as possible, e.g. only a name, place, or thing.
We then compute the semantic entropy over these answers plus the original factual claim. Including the original fact ensures that the estimator remains grounded in the original claim and helps detect situations in which the question has been interpreted completely differently from the original context. We make a small modification to handle the fact that GPT-4 generations often include refusals to answer questions. These refusals were not something we commonly observe in our experiments with LLaMA 2, Falcon or Mistral models. If more than half of the answers include one of the strings ‘not available’, ‘not provided’, ‘unknown’ or ‘unclear’ then we treat the semantic uncertainty as maximal.
We then average the semantic entropies for each question corresponding to the factual claim to get an entropy for this factual claim.
Despite the extra assumptions and complexity, we find that this method greatly outperforms the baselines.
To compute semantic entailment between the original claim and regenerated answers, we rely on the DeBERTa entailment prediction model as we find empirically that DeBERTa predictions result in higher train-set AUROC than other methods. Because DeBERTa has slightly lower recall than GPT-3.5/4, we use a modified set-up for which we say the answers mean the same as each other if at least one of them entails the other and neither is seen to contradict the other—a kind of ‘non-defeating’ bidirectional entailment check rather than true bidirectional entailment. The good performance of DeBERTa in this scenario is not surprising as both factual claims and regenerated answers are relatively short. We refer to Supplementary Notes 2 and 3 for ablations and experiments regarding our choice of entailment estimator for paragraph-length generations.
We implement two baselines. First, we implement a variant of the P (True) method, which is adapted to the new setting. For each factoid, we generate a question with answers in the same way as for semantic entropy. We then use the following prompt:
Question: {question} Here are some brainstormed ideas: {list of regenerated answers} Possible answer: {original answer} Is the possible answer true? Respond with “yes” or “no”.
As we cannot access the probabilities GPT-4 assigns to predicting ‘yes’ and ‘no’ as the next token, we approximate this using Monte Carlo samples. Concretely, we execute the above prompt ten times (at temperature 1) and then take the fraction of answers which was ‘yes’ as our unbiased Monte Carlo estimate of the token probability GPT-4 assigns to ‘yes’.
As a second, simpler, baseline we check if the model thinks the answer is true. We simply ask:
Following this text: {text so far} You see this statement: {proposition} Is it likely that the statement is true? Respond with ‘yes’ or ‘no’.
It is interesting that this method ought to perform very well if we think that the model has good ‘self-knowledge’ (that is, if “models mostly know what they don’t know” 24 ) but in fact semantic entropy is much better at detecting confabulations.
The data used for the short-phrase and sentence-length generations are publicly available and the released code details how to access it. We release a public version of the FactualBio dataset as part of the code base for reproducing the paragraph-length experiments.
We release all code used to produce the main experiments. The code for short-phrase and sentence-length experiments can be found at github.com/jlko/semantic_uncertainty and https://doi.org/10.5281/zenodo.10964366 (ref. 65 ). The code for paragraph-length experiments can be found at github.com/jlko/long_hallucinations and https://doi.org/10.5281/zenodo.10964366 (ref. 65 ).
GPT-4 technical report. Preprint at https://arxiv.org/abs/2303.08774 (2023).
Gemini: a family of highly capable multimodal models. Preprint at https://arxiv.org/abs/2312.11805 (2023).
Xiao, Y. & Wang, W. Y. On hallucination and predictive uncertainty in conditional language generation. In Proc. 16th Conference of the European Chapter of the Association for Computational Linguistics 2734–2744 (Association for Computational Linguistics, 2021).
Rohrbach, A., Hendricks, L. A., Burns, K., Darrell, T. & Saenko, K. Object hallucination in image captioning. In Proc. 2018 Conference on Empirical Methods in Natural Language Processing (eds Riloff, E., Chiang, D., Hockenmaier, J. & Tsujii, J.) 4035–4045 (Association for Computational Linguistics, 2018).
Weiser, B. Lawyer who used ChatGPT faces penalty for made up citations. The New York Times (8 Jun 2023).
Opdahl, A. L. et al. Trustworthy journalism through AI. Data Knowl. Eng . 146 , 102182 (2023).
Shen, Y. et al. ChatGPT and other large language models are double-edged swords. Radiology 307 , e230163 (2023).
Article PubMed Google Scholar
Schulman, J. Reinforcement learning from human feedback: progress and challenges. Presented at the Berkeley EECS Colloquium. YouTube www.youtube.com/watch?v=hhiLw5Q_UFg (2023).
Ji, Z. et al. Survey of hallucination in natural language generation. ACM Comput. Surv. 55 , 248 (2023).
Maynez, J., Narayan, S., Bohnet, B. & McDonald, R. On faithfulness and factuality in abstractive summarization. In Proc. 58th Annual Meeting of the Association for Computational Linguistics (eds Jurafsky, D., Chai, J., Schluter, N. & Tetreault, J.) 1906–1919 (Association for Computational Linguistics, 2020).
Filippova, K. Controlled hallucinations: learning to generate faithfully from noisy data. In Findings of the Association for Computational Linguistics: EMNLP 2020 (eds Webber, B., Cohn, T., He, Y. & Liu, Y.) 864–870 (Association for Computational Linguistics, 2020).
Berrios, G. Confabulations: a conceptual history. J. Hist. Neurosci. 7 , 225–241 (1998).
Article CAS PubMed Google Scholar
Lin, S., Hilton, J. & Evans, O. Teaching models to express their uncertainty in words. Transact. Mach. Learn. Res. (2022).
Evans, O. et al. Truthful AI: developing and governing AI that does not lie. Preprint at https://arxiv.org/abs/2110.06674 (2021).
Amodei, D. et al. Concrete problems in AI safety. Preprint at https://arxiv.org/abs/1606.06565 (2016).
Jiang, Z., Araki, J., Ding, H. & Neubig, G. How can we know when language models know? On the calibration of language models for question answering. Transact. Assoc. Comput. Linguist. 9 , 962–977 (2021).
Article Google Scholar
Desai, S. & Durrett, G. Calibration of pre-trained transformers. In Proc. 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP) (eds Webber, B., Cohn, T., He, Y. & Liu, Y.) 295–302 (Association for Computational Linguistics, 2020).
Glushkova, T., Zerva, C., Rei, R. & Martins, A. F. Uncertainty-aware machine translation evaluation. In Findings of the Association for Computational Linguistics: EMNLP 2021 (eds Moens, M-F., Huang, X., Specia, L. & Yih, S.) 3920–3938 (Association for Computational Linguistics, 2021).
Wang, Y., Beck, D., Baldwin, T. & Verspoor, K. Uncertainty estimation and reduction of pre-trained models for text regression. Transact. Assoc. Comput. Linguist. 10 , 680–696 (2022).
Baker, S. & Kanade, T. Hallucinating faces. In Proc. Fourth IEEE International Conference on Automatic Face and Gesture Recognition . 83–88 (IEEE, Catalogue no PR00580, 2002).
Eliot, L. AI ethics lucidly questioning this whole hallucinating AI popularized trend that has got to stop. Forbes Magazine (24 August 2022).
Shanahan, M. Talking about large language models. Commun. Assoc. Comp. Machinery 67 , 68–79 (2024).
MacKay, D. J. C. Information-based objective functions for active data selection. Neural Comput. 4 , 590–604 (1992).
Kadavath, S. et al. Language models (mostly) know what they know. Preprint at https://arxiv.org/abs/2207.05221 (2022).
Lindley, D. V. On a measure of the information provided by an experiment. Ann. Math. Stat. 27 , 986–1005 (1956).
Article MathSciNet Google Scholar
Xiao, T. Z., Gomez, A. N. & Gal, Y. Wat zei je? Detecting out-of-distribution translations with variational transformers. In Workshop on Bayesian Deep Learning at the Conference on Neural Information Processing Systems (NeurIPS, Vancouver, 2019).
Christiano, P., Cotra, A. & Xu, M. Eliciting Latent Knowledge (Alignment Research Center, 2021); https://docs.google.com/document/d/1WwsnJQstPq91_Yh-Ch2XRL8H_EpsnjrC1dwZXR37PC8/edit .
Negri, M., Bentivogli, L., Mehdad, Y., Giampiccolo, D. & Marchetti, A. Divide and conquer: crowdsourcing the creation of cross-lingual textual entailment corpora. In Proc. 2011 Conference on Empirical Methods in Natural Language Processing 670–679 (Association for Computational Linguistics, 2011).
Honovich, O. et al. TRUE: Re-evaluating factual consistency evaluation. In Proc. Second DialDoc Workshop on Document-grounded Dialogue and Conversational Question Answering 161–175 (Association for Computational Linguistics, 2022).
Falke, T., Ribeiro, L. F. R., Utama, P. A., Dagan, I. & Gurevych, I. Ranking generated summaries by correctness: an interesting but challenging application for natural language inference. In Proc. 57th Annual Meeting of the Association for Computational Linguistics 2214–2220 (Association for Computational Linguistics, 2019).
Laban, P., Schnabel, T., Bennett, P. N. & Hearst, M. A. SummaC: re-visiting NLI-based models for inconsistency detection in summarization. Trans. Assoc. Comput. Linguist. 10 , 163–177 (2022).
Joshi, M., Choi, E., Weld, D. S. & Zettlemoyer, L. TriviaQA: a large scale distantly supervised challenge dataset for reading comprehension. In Proc. 55th Annual Meeting of the Association for Computational Linguistics 1601–1611 (Association for Computational Linguistics. 2017).
Rajpurkar, P., Zhang, J., Lopyrev, K. & Liang, P. SQuAD: 100,000+ questions for machine compression of text. In Proc. 2016 Conference on Empirical Methods in Natural Language Processing (eds Su, J., Duh, K. & Carreras, X.) 2383–2392 (Association for Computational Linguistics, 2016).
Tsatsaronis, G. et al. An overview of the BIOASQ large-scale biomedical semantic indexing and question answering competition. BMC Bioinformatics 16 , 138 (2015).
Article PubMed PubMed Central Google Scholar
Lee, K., Chang, M.-W. & Toutanova, K. Latent retrieval for weakly supervised open domain question answering. In Proc. 57th Annual Meeting of the Association for Computational Linguistics 6086–6096 (Association for Computational Linguistics, 2019).
Kwiatkowski, T. et al. Natural questions: a benchmark for question answering research. Transact. Assoc. Comput. Linguist. 7 , 452–466 (2019).
Patel, A., Bhattamishra, S. & Goyal, N. Are NLP models really able to solve simple math word problems? In Proc. 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (eds Toutanova, K. et al.) 2080–2094 (Assoc. Comp. Linguistics, 2021).
Touvron, H. et al. Llama 2: open foundation and fine-tuned chat models. Preprint at https://arxiv.org/abs/2307.09288 (2023).
Penedo, G. et al. The RefinedWeb dataset for Falcon LLM: outperforming curated corpora with web data, and web data only. In Proc. 36th Conference on Neural Information Processing Systems (eds Oh, A. et al.) 79155–79172 (Curran Associates, 2023)
Jiang, A. Q. et al. Mistral 7B. Preprint at https://arxiv.org/abs/2310.06825 (2023).
Manakul, P., Liusie, A. & Gales, M. J. F. SelfCheckGPT: Zero-Resource Black-Box hallucination detection for generative large language models. In Findings of the Association for Computational Linguistics: EMNLP 2023 (eds Bouamor, H., Pino, J. & Bali, K.) 9004–9017 (Assoc. Comp. Linguistics, 2023).
Mukhoti, J., Kirsch, A., van Amersfoort, J., Torr, P. H. & Gal, Y. Deep deterministic uncertainty: a new simple baseline. In IEEE/CVF Conference on Computer Vision and Pattern Recognition 24384–24394 (Computer Vision Foundation, 2023).
Schuster, T., Chen, S., Buthpitiya, S., Fabrikant, A. & Metzler, D. Stretching sentence-pair NLI models to reason over long documents and clusters. In Findings of the Association for Computational Linguistics: EMNLP 2022 (eds Goldberg, Y. et al.) 394–412 (Association for Computational Linguistics, 2022).
Barnes, B. & Christiano, P. Progress on AI Safety via Debate. AI Alignment Forum www.alignmentforum.org/posts/Br4xDbYu4Frwrb64a/writeup-progress-on-ai-safety-via-debate-1 (2020).
Irving, G., Christiano, P. & Amodei, D. AI safety via debate. Preprint at https://arxiv.org/abs/1805.00899 (2018).
Der Kiureghian, A. & Ditlevsen, O. Aleatory or epistemic? Does it matter? Struct. Saf. 31 , 105–112 (2009).
Malinin, A. & Gales, M. Uncertainty estimation in autoregressive structured prediction. In Proceedings of the International Conference on Learning Representations https://openreview.net/forum?id=jN5y-zb5Q7m (2021).
Murray, K. & Chiang, D. Correcting length bias in neural machine translation. In Proc. Third Conference on Machine Translation (eds Bojar, O. et al.) 212–223 (Assoc. Comp. Linguistics, 2018).
Holtzman, A., Buys, J., Du, L., Forbes, M. & Choi, Y. The curious case of neural text degeneration. In Proceedings of the International Conference on Learning Representations https://openreview.net/forum?id=rygGQyrFvH (2020).
Fan, A., Lewis, M. & Dauphin, Y. Hierarchical neural story generation. In Proc. 56th Annual Meeting of the Association for Computational Linguistics (eds Gurevych, I. & Miyao, Y.) 889–898 (Association for Computational Linguistics, 2018).
Speaks, J. in The Stanford Encyclopedia of Philosophy (ed. Zalta, E. N.) (Metaphysics Research Lab, Stanford Univ., 2021).
Culicover, P. W. Paraphrase generation and information retrieval from stored text. Mech. Transl. Comput. Linguist. 11 , 78–88 (1968).
Google Scholar
Padó, S., Cer, D., Galley, M., Jurafsky, D. & Manning, C. D. Measuring machine translation quality as semantic equivalence: a metric based on entailment features. Mach. Transl. 23 , 181–193 (2009).
Androutsopoulos, I. & Malakasiotis, P. A survey of paraphrasing and textual entailment methods. J. Artif. Intell. Res. 38 , 135–187 (2010).
MacCartney, B. Natural Language Inference (Stanford Univ., 2009).
He, P., Liu, X., Gao, J. & Chen, W. Deberta: decoding-enhanced BERT with disentangled attention. In International Conference on Learning Representations https://openreview.net/forum?id=XPZIaotutsD (2021).
Brown, T. et al. Language models are few-shot learners. Adv. Neural Inf. Process. Syst. 33 , 1877–1901 (2020).
Williams, A., Nangia, N. & Bowman, S. R. A broad-coverage challenge corpus for sentence understanding through inference. In Proc. 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (eds Walker, M. et al.) 1112–1122 (Assoc. Comp. Linguistics, 2018).
Yu, L., Hermann, K. M., Blunsom, P. & Pulman, S. Deep learning for answer sentence selection. Preprint at https://arxiv.org/abs/1412.1632 (2014).
Socher, R., Huang, E., Pennin, J., Manning, C. D. & Ng, A. Dynamic pooling and unfolding recursive autoencoders for paraphrase detection. In Proceedings of the 24th Conference on Neural Information Processing Systems (eds Shawe-Taylor, J. et al.) (2011)
He, R., Ravula, A., Kanagal, B. & Ainslie, J. Realformer: Transformer likes residual attention. In Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021 (eds Zhong, C., et al.) 929–943 (Assoc. Comp. Linguistics, 2021).
Tay, Y. et al. Charformer: fast character transformers via gradient-based subword tokenization. In Proceedings of the International Conference on Learning Representations https://openreview.net/forum?id=JtBRnrlOEFN (2022).
Kane, H., Kocyigit, Y., Abdalla, A., Ajanoh, P. & Coulibali, M. Towards neural similarity evaluators. In Workshop on Document Intelligence at the 32nd conference on Neural Information Processing (2019).
Lebret, R., Grangier, D. & Auli, M. Neural text generation from structured data with application to the biography domain. In Proc. 2016 Conference on Empirical Methods in Natural Language Processing (eds Su, J. et al.) 1203–1213 (Association for Computational Linguistics, 2016).
Kossen, J., jlko/semantic_uncertainty: Initial release v.1.0.0. Zenodo https://doi.org/10.5281/zenodo.10964366 (2024).
Download references
We thank G. Irving, K. Perlin, J. Richens, L. Rimell and M. Turpin for their comments or discussion related to this work. We thank K. Handa for his help with the human evaluation of our automated accuracy assessment. We thank F. Bickford Smith and L. Melo for their code review. Y.G. is supported by a Turing AI Fellowship funded by the UK government’s Office for AI, through UK Research and Innovation (grant reference EP/V030302/1), and delivered by the Alan Turing Institute.
These authors contributed equally: Sebastian Farquhar, Jannik Kossen, Lorenz Kuhn
OATML, Department of Computer Science, University of Oxford, Oxford, UK
Sebastian Farquhar, Jannik Kossen, Lorenz Kuhn & Yarin Gal
You can also search for this author in PubMed Google Scholar
S.F. led the work from conception to completion and proposed using bidirectional entailment to cluster generations as a way of computing entropy in LLMs. He wrote the main text, most of the Methods and Supplementary Information and prepared most of the figures. J.K. improved the mathematical formalization of semantic entropy; led the extension of semantic entropy to sentence- and paragraph-length generations; wrote the code for, and carried out, all the experiments and evaluations; wrote much of the Methods and Supplementary Information and prepared drafts of many figures; and gave critical feedback on the main text. L.K. developed the initial mathematical formalization of semantic entropy; wrote code for, and carried out, the initial experiments around semantic entropy and its variants which demonstrated the promise of the idea and helped narrow down possible research avenues to explore; and gave critical feedback on the main text. Y.G. ideated the project, proposing the idea to differentiate semantic and syntactic diversity as a tool for detecting hallucinations, provided high-level guidance on the research and gave critical feedback on the main text; he runs the research laboratory in which the work was carried out.
Correspondence to Sebastian Farquhar .
Competing interests.
S.F. is currently employed by Google DeepMind and L.K. by OpenAI. For both, this paper was written under their University of Oxford affiliation. The remaining authors declare no competing interests.
Peer review information.
Nature thanks Mirella Lapata and the other, anonymous, reviewer(s) for their contribution to the peer review of this work.
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Extended data fig. 1 algorithm outline for bidirectional entailment clustering..
Given a set of outputs in response to a context, the bidirectional entailment answer returns a set of sets of outputs which have been classified as sharing a meaning.
Supplementary information.
Supplementary Notes 1–7, Figs. 1–10, Tables 1–4 and references. Includes, worked example for semantic entropy calculation, discussion of limitations and computational cost of entailment clustering, ablation of entailment prediction and clustering methods, discussion of automated accuracy assessment, unaggregated results for sentence-length generations and further results for short-phrase generations.
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/ .
Reprints and permissions
Cite this article.
Farquhar, S., Kossen, J., Kuhn, L. et al. Detecting hallucinations in large language models using semantic entropy. Nature 630 , 625–630 (2024). https://doi.org/10.1038/s41586-024-07421-0
Download citation
Received : 17 July 2023
Accepted : 12 April 2024
Published : 19 June 2024
Issue Date : 20 June 2024
DOI : https://doi.org/10.1038/s41586-024-07421-0
Anyone you share the following link with will be able to read this content:
Sorry, a shareable link is not currently available for this article.
Provided by the Springer Nature SharedIt content-sharing initiative
By submitting a comment you agree to abide by our Terms and Community Guidelines . If you find something abusive or that does not comply with our terms or guidelines please flag it as inappropriate.
Sign up for the Nature Briefing: AI and Robotics newsletter — what matters in AI and robotics research, free to your inbox weekly.
IMAGES
VIDEO
COMMENTS
2 The Logistic Regression Model. Let X∈Rn×dbe a data matrix where nis the number of instances (examples) and dis the number of features (parameters or attributes), and ybe a binary. outcomes ...
Abstract. Logistic regression is used to obtain odds ratio in the presence of more than one explanatory variable. The procedure is quite similar to multiple linear regression, with the exception that the response variable is binomial. The result is the impact of each variable on the odds ratio of the observed event of interest.
Logistic regression is used to estimate the association of one or more independent (predictor) variables with a binary dependent (outcome) variable. 2 A binary (or dichotomous) variable is a categorical variable that can only take 2 different values or levels, such as "positive for hypoxemia versus negative for hypoxemia" or "dead versus ...
Binary logistic regression is one method that is particularly appropriate for analysing survey data in the widely used cross-sectional and case-control research designs. 7-9 In the Family Medicine and Community Health (FMCH) journal, 35 out of the 142 (24.6%) peer-reviewed published original research papers between 2013 and 2020 reported ...
Logistic regression is a common classification method when the response variable is binary. Given a response vector yn×1, a model matrix X =[X1,..., X n]∈Rn×p, and regression coefficients β ∈Rp×1,the logistic regression model assumes log(P(yi =1 |xi)/ P(yi =0 |xi))=β xi. Logistic regression minimizes the negative log-likelihood of ...
Logistic regression accomplishes this by using a link function to generalize the linear model for non-continuous outcomes. You may be wondering why linear regression cannot be implemented when the categorical outcome is dummy coded as outlined in chapter "Data Preparation". In a binary case, in which the categorical response has been coded ...
Regression analysis is a valuable research method because of its versatile application to different study contexts. For instance, one may wish to examine associations between an outcome and several independent variables (also commonly referred to as covariates, predictors, and explanatory variables), 1 or one might want to determine how well an outcome is predicted from a set of independent ...
For example, if the research sample is 50 persons and the Logistic Regression analysis contains 50 Independent Variable, the outcome is an overfit (and hence unstable) model. In general, the beta coefficients of Independent Variable in an overfit model are significantly higher than they otherwise would be, and the standard errors are larger ...
Illustration of Logistic Regression Analysis. and Reporting. For the sake of illustration, we constructed a hypothetical. data set to which logistic regression was applied, and we. interpreted its results. The hypothetical data consisted of reading scores and genders of 189 inner city school children.
Regression techniques are versatile in their application to medical research because they can measure associations, predict outcomes, and control for confounding variable effects. As one such technique, logistic regression is an efficient and powerful way to analyze the effect of a group of independent vari-ables on a binary outcome by ...
gistic regression its name. The sigmoid has the following equation, sh. 1. =1+e1=z 1+exp( z)(5.4)(For the rest of the book, we'll use the. otation exp(x) to mean ex.) The sigmoid has a number of advantages; it takes a real-valued number and maps it into the range (0;1), which is just wha.
The authors evaluated the use and interpretation of logistic regression presented in 8 articles published in The Journal of Educational Research between 1990 and 2000. They found that all 8 ...
Logistic regression models the log odds ratio as a linear combination of the independent variables. For our example, height ( H) is the independent variable, the logistic fit parameters are β0 ...
Logistic regression sometimes called the logistic model or logit model, analyzes the relationship between multiple independent variables and a categorical dependent variable, and estimates the probability of occur-rence of an event by fitting data to a logistic curve. There are two models of logistic regression, binary logistic regression and ...
Logistic regression is used to obtain the odds ratio in the presence of more than one explanatory variable. This procedure is quite similar to multiple linear regression, with the only exception that the response variable is binomial. The result is the impact of each variable on the odds ratio of the observed event of interest.
These results extend the research base regarding the relationship between the LA program and positive student outcomes. ... we use logistic regression with pre-existing institutional data to investigate the relationship between exposure to LA support in large introductory STEM courses and general failure rates in these same and other ...
This paper provides an introduction to Linear Logistic Regression and its value in semiconductor yield and reliability analysis. The reliability community has become well experienced in fitting of survival distributions, the use of design of experiments (DOE) and the associated general linear model (linear regression and analysis of variance methods) approach to analysis. This method provides ...
Linear and logistic regressions are widely used statistical methods to assess the association between variables in medical research. These methods estimate if there is an association between the independent variable (also called predictor, exposure, or risk factor) and the dependent variable (outcome). 2. The association between two variables ...
The logistic regression model allows to examine the influence of many independent variables ð '‹ð '‹1 ,… ,ð '‹ð '‹ð '˜ð '˜the dependent variable Y. The variable Y takes only two values and is dichotomous. ... .†Research Papers of WrocÅ‚aw University of Economics/ Prace Naukowe Uniwersytetu Ekonomicznego ...
RESEARCH PAPER APPROVAL PREDICTING STUDENT SUCCESS: A LOGISTIC REGRESSION ANALYSIS OF DATA FROM MULTIPLE SIU-C COURSES By Patrick B. Soule A Research Paper Submitted in Partial Ful llment of the Requirements for the Degree of Master of Science in the eld of Mathematics Approved by: Dr. B. Bhattacharya, Chair Dr. M. Wright Dr. R. Habib Graduate ...
Summary In this paper, we demonstrate that logistic regression can be a powerful analytical technique for use when the outcome variable is dichotomous. The effectiveness of the logistic model was shown to be supported by (a) significance tests of the model against the null model, (b) the significance test of each predictor, (c) descriptive and ...
ABSTRACT OF DISSERTATION. TOPICS IN LOGISTIC REGRESSION ANALYSIS. Discrete-time Markov chains have been used to analyze the transition of subjects from intact cognition to dementia with mild cognitive impairment and global impair-ment as intervening transient states, and death as competing risk.
tion of logistic regression applied to a data set in testing a research hypothesis. Recommendations are also offered for appropriate reporting formats of logistic regression results and the minimum observation-to-predictor ratio. The authors evaluated the use and interpretation of logistic regression pre-
STATA 14 was used to analyse the data. Ordinal logistic regression was used to identify independent predictors of food consumption score. Those variables having p value < 0.25 in the bivariable ordinal logistic regression were considered for the final model. Crude and Adjusted Odds Ratio were used to assess the strength of the association.
As a strong supervised baseline, we compare to an embedding regression method inspired by ref. 24 which trains a logistic regression classifier to predict whether the model correctly answered a ...