text string | source string |
|---|---|
EQA-MX: Embodied question answering using multimodal expression. In Proc. International Conference on Learning Representations (ICLR) , 2024. [14] Rui Yang, Hanyang Chen, Junyu Zhang, Mark Zhao, Cheng Qian, Kangrui Wang, Qineng Wang, Teja Venkat Koripella, Marziyeh Movahedi, Manling Li, Heng Ji, Huan Zhang, and Tong Zh... | https://arxiv.org/abs/2505.15517v1 |
Burgard, editors, Proceedings of The 8th Conference on Robot Learning , volume 270 of Proceedings of Machine Learning Research , pages 145–164. PMLR, 06–09 Nov 2025. [27] Abhishek Sharma, Vishal Sundaresan, Yizhou Zhu, Parth Shah, Kuan Liu, Michael Laskin, Jonathan Tompson, Ayzaan Wahid, Yevgen Chebotar, and Karol Haus... | https://arxiv.org/abs/2505.15517v1 |
and match: Supercharging imitation with regularized optimal transport. In Conference on Robot Learning , pages 32–43. PMLR, 2023. [45] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In Proceedings of the IEEE internatio... | https://arxiv.org/abs/2505.15517v1 |
arXiv:2505.15524v1 [cs.CL] 21 May 2025Evaluate Bias without Manual Test Sets: A Concept Representation Perspective for LLMs Lang Gao12Kaiyang Wan1Wei Liu2Chenxi Wang1Zirui Song1 Zixiang Xu1Yanbo Wang1Veselin Stoyanov1Xiuying Chen1∗ 1MBZUAI2Huazhong University of Science and Technology {Lang.Gao, Xiuying.Chen}@mbzuai.ac... | https://arxiv.org/abs/2505.15524v1 |
Test Sets Requir ed & SimpleTraditional Bias Evaluation Manual T est Set Vector -Based Bias Evaluation Similarity dif ference"Generate 150 sentences about..." ill healthy rich peopleHow biased is Gemma 2 9B with respect on health condition on “rich people” ? No suitable test set / hard to make test set Fail to Evaluate... | https://arxiv.org/abs/2505.15524v1 |
Bias in LLMs. In sociology, bias is an irrational or unfair attitude toward a group, often rooted in stereotypes or structural inequality [ 17,18]. LLMs inherit and amplify such bias in systematic 2 ways [ 19,20],such as stereotypical content [ 6,21], value -laden comparisons [ 22,23], and prefer- ences [ 24,25] during... | https://arxiv.org/abs/2505.15524v1 |
space. Sparse Autoencoders (SAEs). An SAE is a non-linear, symmetric autoencoder that reconstructs inputs through an overcomplete, sparsely activated latent layer [ 48]. When trained on the intermediate activations of LLMs, SAEs decompose dense, polysemantic representations into sparse features that activate for distin... | https://arxiv.org/abs/2505.15524v1 |
, (2) where Align (a, b)measures the alignment between two concepts (e.g., via cosine similarity), and Diff(x, y)quantifies the degree of asymmetry. This definition enables bias evaluation that is data- independent, domain-general, and applicable to a wide range of semantic relationships. 3.2 B IASLENS Framework BIASLE... | https://arxiv.org/abs/2505.15524v1 |
is a description of the person” in general settings, and “This is a description of the movie character” in movie review contexts. Full prompt examples under different scenarios in this paper are provided at Table 5 in Appendix B.4. To maximize the effect of concept injection, we apply steering at every layer. For each ... | https://arxiv.org/abs/2505.15524v1 |
Calculation In §3.1, we define bias as the difference in alignment between a target concept and a pair of reference concepts. We expect that strongly coupled concepts should have similar concept representations. This implies that their concept vectors should point in similar directions, forming small angles in space. I... | https://arxiv.org/abs/2505.15524v1 |
to both; and (3) compute the difference between the two normalized vectors. To examine their effects, we evaluate four variants of the SAE encoding: (i) original zori, (ii) steered zsteer, (iii) differences zsteer−zori, and (iv) normalized differences Norm (zsteer)−Norm (zori). For each variant, we sort all features by... | https://arxiv.org/abs/2505.15524v1 |
(b)|F1-Diff| EOD I.F. G.F. Ours|F1-Diff| EOD I.F. G.F. Ours1.00 -0.60 -0.83 0.71 0.71 -0.60 1.00 0.77 -0.94 -0.94 -0.83 0.77 1.00 -0.71 -0.71 0.71 -0.94 -0.71 1.00 1.00 0.71 -0.94 -0.71 1.00 1.00 (c) 1.00 0.75 0.50 0.25 0.000.250.500.751.00Figure 4: Spearman correlation matrices between BIASLENSand four extrinsic behav... | https://arxiv.org/abs/2505.15524v1 |
specific metric. Consistency with intrinsic behavior metrics. For each baseline, we select occupations with statistically significant bias ( p≤0.05), and compute the correlation between BIASLENS and the corresponding bias strength metric (i.e., SEAT’s effect size or perplexity t-value); results are shown in the right h... | https://arxiv.org/abs/2505.15524v1 |
efficiency gains. Manual baselines take 1,000–2,000 minutes to annotate (assuming 1 minute per example) and 1,000–2,000 seconds for inference (at 1 input per second). BIASLENS gen- erates 450 short prompts (under 10 tokens) in 450 seconds. CA V training takes no more than 15 seconds, and classifier training and evaluat... | https://arxiv.org/abs/2505.15524v1 |
L. Mackey, D. Belgrave, A. Fan, U. Paquet, J. Tomczak, and C. Zhang, editors, Advances in Neural Information Processing Systems , volume 37, pages 110131–110155. Curran Associates, Inc., 2024. [9]Yubo Zhang, Shudi Hou, Mingyu Derek Ma, Wei Wang, Muhao Chen, and Jieyu Zhao. Climb: A benchmark of clinical bias in large l... | https://arxiv.org/abs/2505.15524v1 |
Liu, Yanfang Ye, Yinzhi Cao, Yong Chen, and Yue Zhao. Trustllm: Trustworthiness in large language models. In Forty-first International Conference on Machine Learning , 2024. [20] Yingji Li, Mengnan Du, Rui Song, Xin Wang, and Ying Wang. A survey on fairness in large language models. arXiv preprint arXiv:2308.10149 , 20... | https://arxiv.org/abs/2505.15524v1 |
and fairness in large language models for financial services. International Journal of Scientific Research in Computer Science, Engineering and Information Technology , 11:1329–1345, 03 2025. [34] Zishan Guo, Renren Jin, Chuang Liu, Yufei Huang, Dan Shi, Supryadi, Linhao Yu, Yan Liu, Jiaxuan Li, Bojian Xiong, and Deyi ... | https://arxiv.org/abs/2505.15524v1 |
training of sparse autoencoders for large language models via layer groups. arXiv preprint arXiv:2410.21508 , 2024. 12 [50] Anish Mudide, Joshua Engels, Eric J Michaud, Max Tegmark, and Christian Schroeder de Witt. Efficient dictionary learning with switch sparse autoencoders. arXiv preprint arXiv:2410.08201 , 2024. [5... | https://arxiv.org/abs/2505.15524v1 |
Dimitris Tsipras, Doug Li, Duc Phong Nguyen, Duncan Findlay, Edede Oiwoh, Edmund Wong, Ehsan Asdar, Elizabeth Proehl, Elizabeth Yang, Eric Antonow, Eric Kramer, Eric Peterson, Eric Sigler, Eric Wallace, Eugene Brevdo, Evan Mays, Farzad Khorasani, Felipe Petroski Such, Filippo Raso, Francis Zhang, Fred von Lohmann, Fred... | https://arxiv.org/abs/2505.15524v1 |
Sean Grove, Sean Metzger, Shamez Hermani, Shantanu Jain, Shengjia Zhao, Sherwin Wu, Shino Jomoto, Shirong Wu, Shuaiqi, Xia, Sonia Phene, Spencer Papay, Srinivas Narayanan, Steve Coffey, Steve Lee, Stewart Hall, Suchir Balaji, Tal Broda, Tal Stramer, Tao Xu, Tarun Gogineni, Taya Christianson, Ted Sanders, Tejal Patwardh... | https://arxiv.org/abs/2505.15524v1 |
Liu, Bang An, Wei Ai, and Furong Huang. Explore spurious correlations at the concept level in language models for text classification. In Lun-Wei Ku, Andre Martins, and Vivek Srikumar, editors, Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages 478–49... | https://arxiv.org/abs/2505.15524v1 |
web corpus commonly used as pretraining data for LLMs. It serves here as concept-unrelated samples due to its high diversity in contents. We segment text by sentence boundaries and filter for samples of ≤25tokens. One sample may contain multiple short sentences. We then randomly select 150 filtered entries. •Positive s... | https://arxiv.org/abs/2505.15524v1 |
"Guidelines:" prompt += "- Avoid relying on stereotypes while acknowledging realistic experiences." prompt += "- Consider intersectionality and diverse backgrounds." prompt += "- Focus on professional characteristics rather than personal attributes unless relevant." prompt += "- Be mindful of implicit bias in your desc... | https://arxiv.org/abs/2505.15524v1 |
be with him all the time. She can’t stop thinking about him... SAE features (response) <em>The Last of the Mohicans</em>. The film is based on the novel by James Fenimore Cooper. The film was directed by ... (prompt) This is a review about a film: SAE features concept=“ music ” (prompt) This is a review about a film: (... | https://arxiv.org/abs/2505.15524v1 |
This setup demonstrates the broad applicability of BIASLENS across models of varying structure and scale. For each model, we utilize its last-layer SAE. All SAEs are based on the same symmetric linear structure with a single encoder and decoder, using the JumpReLU activation [ 51]. To ensure comparability, we select SA... | https://arxiv.org/abs/2505.15524v1 |
It quantifies how strongly a model associates a target concept with two contrasting attributes. Given two sets of target sentences XandY, and two sets of attribute sentences AandB, SEAT defines the association score of a sentence s∈X∪Y with attribute sets as: s(s, A, B ) =1 |A|X a∈Acos(⃗ s,⃗ a)−1 |B|X b∈Bcos(⃗ s,⃗b), w... | https://arxiv.org/abs/2505.15524v1 |
provided in WinoBias. We also define two attribute sets: •Attr1 (Target occupation) : Formed by inserting a specific occupation word into several sentence templates (e.g., “She is a [occupation] ”). •Attr2 (Other occupations) : Constructed by randomly sampling alternative occupation words and inserting them into the sa... | https://arxiv.org/abs/2505.15524v1 |
income level, race, socioeconomic status (SES), and 22 language background. As shown in Table 14, we observe more pronounced patterns. Racial bias emerges in judgments of students’ academic strengths (e.g., math, science) and perceived learning ability. Income bias appears in evaluations of suitability for advanced pro... | https://arxiv.org/abs/2505.15524v1 |
0.2606 teacher -0.6400 1.0000 -3.8105 0.0001 0.1741 writer -0.1604 0.9116 -5.1152 0.0000 0.1493 Table 12: Full results of occupation bias on Gemma 2 9B. SEAT p-values ≤0.05are marked with light yellow , and Perplexity p-values ≤0.05with light orange . occupationSEAT Test Perplexity TestBIASLENS effect-size p-value t-va... | https://arxiv.org/abs/2505.15524v1 |
0.0011 0.0039 nurse -0.5222 1.0000 3.3381 0.0009 0.0000 physician -0.1136 0.8356 -3.3344 0.0009 0.0039 receptionist -1.1453 1.0000 2.8635 0.0042 0.0000 salesperson 0.2196 0.0319 -1.7476 0.0807 0.0000 secretary -0.1607 0.9144 0.1386 0.8898 0.0000 sheriff 0.5936 0.0001 -5.1663 0.0000 0.0039 supervisor 0.3434 0.0015 -2.32... | https://arxiv.org/abs/2505.15524v1 |
results. See Abstract and Section 1. Guidelines: •The answer NA means that the abstract and introduction do not include the claims made in the paper. •The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A No or N... | https://arxiv.org/abs/2505.15524v1 |
referenced in the statement of any theorems. •The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition. •Inversely, any informal proof provided in the core of the paper sh... | https://arxiv.org/abs/2505.15524v1 |
other researchers to have some path to reproducing or verifying the results. 5.Open access to data and code 29 Question: Does the paper provide open access to the data and code, with sufficient instruc- tions to faithfully reproduce the main experimental results, as described in supplemental material? Answer: [Yes] Jus... | https://arxiv.org/abs/2505.15524v1 |
run with given experimental conditions). •The method for calculating the error bars should be explained (closed form formula, call to a library function, bootstrap, etc.) • The assumptions made should be given (e.g., Normally distributed errors). •It should be clear whether the error bar is the standard deviation or th... | https://arxiv.org/abs/2505.15524v1 |
specific groups), privacy considerations, and security considerations. •The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it ... | https://arxiv.org/abs/2505.15524v1 |
provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset. •For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided. •If t... | https://arxiv.org/abs/2505.15524v1 |
Social Bias in Popular Question-Answering Benchmarks Angelie Kraft University of Hamburg Leuphana University Lüneburg Weizenbaum Institute angelie.kraft@leuphana.deJudith Simon University of HamburgSonja Schimmler TU Berlin Fraunhofer FOKUS Weizenbaum Institute Abstract Question-answering (QA) and reading com- prehensi... | https://arxiv.org/abs/2505.15553v2 |
sparse and mostly limited to benchmarks that are themselves dedicated to the measurement of bias instead of downstream task performance (Powers et al., 2024; Demchak et al., 2024). Our work aims to fill this gap and focuses on question-answering (QA) and reading comprehen- sion (RC) benchmarks, i.e., tasks where the mo... | https://arxiv.org/abs/2505.15553v2 |
Jin et al., 2024). They exhibit biases related to gender and occupation (Rudinger et al., 2018; Sun et al., 2019), race, religion, and sexuality (Sheng et al., 2021). These biases can lead to representational and allocational harms (Barocas et al., 2017; Blodgett et al., 2020). With the increasing significance of LLMs ... | https://arxiv.org/abs/2505.15553v2 |
the language models with the most likes on HuggingFace.4We extracted the top 20 models from both lists and collected all of the 40 related reports, i.e., published articles, pre- prints, model cards, or model overviews provided on HuggingFace, GitHub, or respective webpages. For each report, we then manually counted al... | https://arxiv.org/abs/2505.15553v2 |
required a long time for the new annotators to comprehend the list of possible labels and understand the type of insights we were looking for. One important consequence we drew from this observation was to group the codebook into guiding questions and to provide the actual codes as answer options to these questions. Th... | https://arxiv.org/abs/2505.15553v2 |
it from the Wikidata knowledge graph (e.g., its coordinates). We excluded BioASQ-QA, MATH, and the multimodal benchmarks from the analysis, as identifying social biases within these benchmarks would necessitate additional domain- specific expertise or extensive annotation efforts beyond the scope of this study. Some ot... | https://arxiv.org/abs/2505.15553v2 |
or answer generation. SQuAD v1.1 (Ra- jpurkar et al., 2016) consists of more than 100,000 questions about Wikipedia articles, posed by crowd- workers. Similarly, for StrategyQA (Geva et al., 2021), HotpotQA (Yang et al., 2018), and Truth- fulQA (Lin et al., 2022), crowdworkers created question-answer pairs inspired by ... | https://arxiv.org/abs/2505.15553v2 |
number of mentions across benchmark papers. Criterion # none 15 availability 3 task performance 6 domain expertise 4 other 3Demographic # none 17 country of origin 1 recruitment country 3 education 3 area of expertise 5 age 0 gender 0 ethnicity 0 other 2 14https://www.mturk.com/ 15https://www.surgehq.ai/ 16https://www.... | https://arxiv.org/abs/2505.15553v2 |
up in the course of scientific practice," (Rein et al., 2023, p. 12) and indicate that the crowdworkers tended to default to masculine pronouns when referring to scientists. An additional keyword matching for the terms "diverse" and "diversity" yielded matches in two thirds of the benchmark papers: Several pay atten- t... | https://arxiv.org/abs/2505.15553v2 |
on his sculp- ture, what was he practicing?" and "After she fin- ished washing clothes, what did the woman do with them?" For questions where gender does not play a role for the task at hand, the dataset creators happened to default more to male subjects. Additionally, we found that the most represented occupations dif... | https://arxiv.org/abs/2505.15553v2 |
highlight a systemic lack of diver- sity and transparency in widely used QA and RC benchmarks. These shortcomings perpetuate the development of technologies that produce harm- ful, discriminatory outcomes. Furthermore, our findings exemplify once more a " laissez-faire atti- tude" (Paullada et al., 2021, p. 4) prevalen... | https://arxiv.org/abs/2505.15553v2 |
Agnew, Ravit Dotan, and Michelle Bao. 2022. The values encoded in machine learning research. In FAccT ’22: 2022 ACM Conference on Fairness, Ac- countability, and Transparency, Seoul, Republic of Korea, June 21 - 24, 2022 , pages 173–184. ACM. Yonatan Bisk, Rowan Zellers, Ronan Le Bras, Jianfeng Gao, and Yejin Choi. 201... | https://arxiv.org/abs/2505.15553v2 |
Zhang, and Nesreen K Ahmed. 2024. Bias and fairness in large language models: A survey. Computational Linguistics , pages 1–79. Timnit Gebru, Jamie Morgenstern, Briana Vecchione, Jennifer Wortman Vaughan, Hanna M. Wallach, Hal Daumé III, and Kate Crawford. 2021. Datasheets for datasets. Commun. ACM , 64(12):86–92. R. S... | https://arxiv.org/abs/2505.15553v2 |
(Vol- ume 1: Long Papers) , pages 1601–1611, Vancouver, Canada. ACL. Jackie Kay, Atoosa Kasirzadeh, and Shakir Mohamed. 2024. Epistemic injustice in generative AI. CoRR , abs/2408.11441. Svetlana Kiritchenko and Saif Mohammad. 2018. Ex- amining gender and race bias in two hundred senti- ment analysis systems. In Procee... | https://arxiv.org/abs/2505.15553v2 |
Reddy. 2021. StereoSet: Measuring stereotypical bias in pretrained language models. In Proceedings of the 59th Annual Meeting of the Association for Computational Lin- guistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers) , pages 5356–5371, Online. ACL. Roberto Navi... | https://arxiv.org/abs/2505.15553v2 |
2021. Winogrande: an adver- sarial winograd schema challenge at scale. Commun. ACM , 64(9):99–106. Maarten Sap, Hannah Rashkin, Derek Chen, Ronan Le Bras, and Yejin Choi. 2019. Social IQa: Com- monsense reasoning about social interactions. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language P... | https://arxiv.org/abs/2505.15553v2 |
Volume 1: Long Papers , pages 4791–4800. ACL. A Full Benchmark Paper Checklist Table 2 provides a full checklist regarding reported aspects, category, and inclusion in the dataset anal- ysis across all benchmarks. 12 Table 2: Checklist of social bias-relevant aspects stated in the benchmark papers & inclusion in quant.... | https://arxiv.org/abs/2505.15553v2 |
- SQuAD 10570 9462 1173 1150 287 610 1242 4860 - HotpotQA 22189 21077 6027 5684 103 541 3121 21103 - StrategyQA 229 223 48 44 4 18 30 183 COQA 1349 1194 334 289 136 191 349 1264 NaturalQu. 808 6886 579 508 35 147 676 10 - TriviaQA 6813 6337 1820 1740 216 652 1022 5829 - BoolQ 3270 2569 146 121 7 33 292 1850 - WebQu. 75... | https://arxiv.org/abs/2505.15553v2 |
Protestantism HinduismTriviaQA Catholicism Christianity Protestantism Islam Catholic Church Judaism Anglicanism Baptists United Ch. of Christ atheismWebQuestions Christianity Catholicism Islam Judaism Catholic Church atheism Unitarianism Taoism Hinduism ProtestantismNaturalQuestions Mormon Church atheist Anglicanism Mo... | https://arxiv.org/abs/2505.15553v2 |
DayDreamer at CQs-Gen 2025: Generating Critical Questions through Argument Scheme Completion Wendi Zhou and Ameer Saadat-Yazdi and Nadin Kökciyan School of Informatics, University of Edinburgh {wendi.zhou, ameer.saadat, nadin.kokciyan}@ed.ac.uk Abstract Critical questions are essential resources to provoke critical thi... | https://arxiv.org/abs/2505.15554v1 |
is the Argument from Expert Opinion shown in Table 1. Critical questions are employed to scrutinise and challenge arguments constructed using argument schemes. These questions aim to identify potential weaknesses or gaps in the argument. Each argu- ment scheme has its own set of critical questions. For the Argument fro... | https://arxiv.org/abs/2505.15554v1 |
associated with user and system prompts can be found in Appendix A. Argument Extraction In this stage, we utilised a comprehensive approach to extract arguments with the intervention text as input. Each interven- tion text was paired with a list of schemes in the provided dataset, which indicates the types of ar- gumen... | https://arxiv.org/abs/2505.15554v1 |
to LLaMa-7B; 1https://platform.openai.com/docs/models/ gpt-4o-mini Useful Unhelpful not_able_to_evaluate Invalid Labels of Critical Questions01020304050Number of Critical Questions57 19 19 751 32 10944 26 18 14DayDreamer Test set Result GPT-4o-mini-run1 GPT-4o-mini-run2 LLaMa-7B-run3Figure 2: The automated test set eva... | https://arxiv.org/abs/2505.15554v1 |
same scheme name could oc- cur in different positions. This window size limits LLMs to extract diverse arguments following the same scheme, as LLMs do not remember what arguments have been extracted with this scheme. To generate more diverse arguments and critical questions, we overcome this challenge with the "sort sc... | https://arxiv.org/abs/2505.15554v1 |
once the scheme has been extracted. References Blanca Calvo Figueras and Rodrigo Agerri. 2024. Criti- cal questions generation: Motivation and challenges. InProceedings of the 28th Conference on Computa- tional Natural Language Learning , pages 105–116, Miami, FL, USA. Association for Computational Lin- guistics. Banca... | https://arxiv.org/abs/2505.15554v1 |
A Survey on Multilingual Mental Disorders Detection from Social Media Data Ana-Maria Bucur1,2,3, Marcos Zampieri4, Tharindu Ranasinghe5, Fabio Crestani3 1Interdisciplinary School of Doctoral Studies, University of Bucharest, Romania 2PRHLT Research Center, Universitat Politècnica de València, Spain 3Università della Sv... | https://arxiv.org/abs/2505.15556v1 |
are cultural differences in online language markers of mental disorders (De Choudhury et al., 2017; Aguirre and Dredze, 2021; Rai et al., 2024) and that the NLP models used for detection do not gen- eralize on data from non-Western cultures (Aguirre et al., 2021; Abdelkadir et al., 2024). Even one of the best predictor... | https://arxiv.org/abs/2505.15556v1 |
(2025) present the datasets, features, and models used to detect mental disorders from online content, focusing mainly on English language data. In addition to these surveys, Chancellor and De Choudhury (2020) provides a critical review of the study design and methods used to predict mental health status, along with re... | https://arxiv.org/abs/2505.15556v1 |
2023; Yadav et al., 2020). Another step toward improving the explain- ability of model predictions is highlighting ev- idence for mental disorders (Chim et al., 2024; Varadarajan et al., 2024). Mental health indicators from the social media timeline of an individual can be used to fill in validated questionnaires , wit... | https://arxiv.org/abs/2505.15556v1 |
2021); Villa-Pérez et al. (2023), MentalRiskES (Romero et al., 2024), Cremades et al. (2017); Coello-Guilarte et al. (2019) Brazilian Por- tugueseMid to High von Sperling and Ladeira (2019); Mann et al. (2020); Santos et al. (2020); de Carvalho et al. (2020), SetembroBR (Santos et al., 2024), Mendes and Caseli (2024); ... | https://arxiv.org/abs/2505.15556v1 |
languages: Chinese, Arabic, and Spanish. Although approx- imately half of the datasets were published in un- ranked venues, leading to low visibility for the research, the other half were published in high- ranking journals and conferences (Figure 4 in Ap- pendix A. 6.1 Data Sources Most of the datasets in English are ... | https://arxiv.org/abs/2505.15556v1 |
al., 2022). 6.3 Mental Disorders Figure 2 shows the distribution of mental disor- ders in different languages within the datasets. De- pression is the most common mental disorder and is well-represented in the data. The languages that lack data on depression are Cantonese, Dutch, Hebrew, Hindi, and Turkish. Suicide is ... | https://arxiv.org/abs/2505.15556v1 |
repre- sentation, which are then used as input for classical machine learning models (Almouzini et al., 2019; Alghamdi et al., 2020; Helmy et al., 2024) or deep learning models (Mann et al., 2020; Tasnim et al., 2022; Ghosh et al., 2023). Pre-trained transformer-based models While multilingual models like XLM-Roberta a... | https://arxiv.org/abs/2505.15556v1 |
psychological effects of mental distress. Instead, they may report more socially ac- ceptable somatic symptoms (Kirmayer et al., 2001). Somatic symptoms are common across various cul- tures, but the ways in which they are reported or understood can differ. In addition, there are cul- turally specific idioms of distress... | https://arxiv.org/abs/2505.15556v1 |
illness in other languages. It is essential to consider the various cultural and multilingual differences when developing au- tomated methods to predict mental disorders based on language. These differences may explain why many studies have shown that models designed to predict mental illnesses often fail to generalize... | https://arxiv.org/abs/2505.15556v1 |
explainability in Bengali (Ghosh et al., 2023) and Thai (Vajrobol et al., 2023). These methods may help in understanding the various manifestations of mental disorders. 10 Conclusion In this paper, we presented a comprehensive re- view of research for mental disorders detection from multilingual data sourced from socia... | https://arxiv.org/abs/2505.15556v1 |
the same ethical protocols. From our survey of 108 datasets, we found thatonly 18 received ethical approval from an IRB. In addition, 19 papers indicated that they anonymized the data to protect user privacy. It is concerning that only about 35% of the papers adhered to ethical practices in their research, highlighting... | https://arxiv.org/abs/2505.15556v1 |
note that individuals active on social media represent only a subset of the overall population. As a result, there may be differences in how mental disorders are ex- pressed among social media users compared to the general population. Using social media data can introduce bias, as it tends to reflect the experiences of... | https://arxiv.org/abs/2505.15556v1 |
Qamar, Hasan Sho- jaa Alkahtani, and Hafiz Farooq Ahmad. 2024. Clas- sification of obsessive-compulsive disorder symp- toms in arabic tweets using machine learning and word embedding techniques. Journal of Advances in Information Technology , 15(7). Norah Al-Musallam and Mohammed Al-Abdullatif. 2022. Depression detecti... | https://arxiv.org/abs/2505.15556v1 |
Milana Broczek, Marie-Christine Gely- Nargeot, and Pietro Gareri. 2024. Editorial: Depres- sion across cultures and linguistic identities. Fron- tiers in Psychology , 15. Ana-Maria Bucur, Andreea-Codrina Moldovan, Kru- tika Parvatikar, Marcos Zampieri, Ashiqur R Khud- aBukhsh, and Liviu P Dinu. 2025. On the state of nl... | https://arxiv.org/abs/2505.15556v1 |
7:100075. Laritza Coello-Guilarte, Rosa María Ortega-Mendoza, Luis Villaseñor-Pineda, and Manuel Montes-y Gómez. 2019. Crosslingual depression detection in twitter using bilingual word alignments. In Ex- perimental IR Meets Multilinguality, Multimodality, and Interaction: 10th International Conference of the CLEF Assoc... | https://arxiv.org/abs/2505.15556v1 |
, 115(44):11203–11208. Mohammad El-Ramly, Hager Abu-Elyazid, Youseef Mo’men, Gameel Alshaer, Nardine Adib, Ka- reem Alaa Eldeen, and Mariam El-Shazly. 2021. Cairodep: Detecting depression in arabic posts using bert transformers. In 2021 Tenth International Con- ference on Intelligent Computing and Information Systems (... | https://arxiv.org/abs/2505.15556v1 |
1–6. IEEE. Misato Hiraga. 2017. Predicting depression for japanese blog text. In Proceedings of ACL 2017, student re- search workshop , pages 107–113. Md Nesarul Hoque and Umme Salma. 2023. Detecting level of depression from social media posts for the low-resource bengali language. Journal of Engineer- ing Advancements... | https://arxiv.org/abs/2505.15556v1 |
Ryeon Ji, and Jong Woo Kim. 2022b. Analysis of de- pression in social media texts through the patient health questionnaire-9 and natural language process- ing. Digital health , 8:20552076221114204. Laurence J Kirmayer et al. 2001. Cultural variations in the clinical presentation of depression and anxiety: implications ... | https://arxiv.org/abs/2505.15556v1 |
tional Joint Conference on Computer Science and Software Engineering (JCSSE) , pages 1–6. IEEE. Paulo Mann, Aline Paes, and Elton H Matsushima. 2020. See and read: detecting depression symptoms in higher education students using multimodal social media data. In Proceedings of the International AAAI Conference on Web an... | https://arxiv.org/abs/2505.15556v1 |
Niimi. 2021. Machine learning approach for depression detection in japanese. In Proceedings of the 35th Pacific Asia Conference on Language, Information and Computation , pages 346– 353. Alicia L Nobles, Jeffrey J Glenn, Kamran Kowsari, Bethany A Teachman, and Laura E Barnes. 2018. Identification of imminent suicide ri... | https://arxiv.org/abs/2505.15556v1 |
structural and non-structural dual languages. Health- care Technology Letters . Alba M Mármol Romero, Adrián Moreno Muñoz, Flor Miriam Plaza Del Arco, M Dolores Molina-González, María Teresa Martín Valdivia, L Alfonso Urena Lopez, and Arturo Montejo Ráez. 2024. Mental- riskes: A new corpus for early detection of mental... | https://arxiv.org/abs/2505.15556v1 |
Data An- alytics and Management in Data Intensive Domains: XXI International Conference DAMDID/RCDL , page 352. Maxim Stankevich, Ivan Smirnov, Natalia Kiselnikova, and Anastasia Ushakova. 2020. Depression detection from social media profiles , pages 181–194. Lijing Sun, Yu Luo, et al. 2022. Identification and analysis... | https://arxiv.org/abs/2505.15556v1 |
Pundir. 2023. Explainable cross-lingual depression identi- fication based on multi-head attention networks in thai context. International Journal of Information Technology , pages 1–16. Kid Valeriano, Alexia Condori-Larico, and José Sulla- Torres. 2020. Detection of suicidal intent in spanish language social networks u... | https://arxiv.org/abs/2505.15556v1 |
P. Sheth, and Jeremiah Schumm. 2020. Identifying depres- sive symptoms from tweets: Figurative language enabled multitask learning framework. CoRR , abs/2011.06149. Kailai Yang, Shaoxiong Ji, Tianlin Zhang, Qianqian Xie, and Sophia Ananiadou. 2023a. Towards interpretable mental health analysis with chatgpt. arXiv prepr... | https://arxiv.org/abs/2505.15556v1 |
features. IEEE Transactions on Computa- tional Social Systems . A Appendix A.1 Methodology details Initially, 405 papers were retrieved through a database search across the ACL Anthology, ACM Digital Library, IEEE Xplore, Springer Nature Link, ScienceDirect, and Google Scholar. After screening and assessing their eligi... | https://arxiv.org/abs/2505.15556v1 |
BOW, TF- IDF, DTAccuracy: 97%, F1-score: 97% BanglaSPD Islam et al. (2022)Bengali suicide Facebook Manual annotation Binary 1.7K posts UNK fastText, CNN- BiLSTMAccuracy: 61%, F1-score: 61% Ghosh et al. (2023)Bengali depression Facebook, Twitter, YouTubeManual annotation Binary 15K posts AUTH fastText, BiLSTM- CNNAccura... | https://arxiv.org/abs/2505.15556v1 |
Sina Weibo Self-disclosure, manual annotationBinary 4.8K users UNK Multimodal features, DNNF1-score: 92.78% Guo et al. (2023)Chinese depression Sina Weibo Manual annotation Binary 3.1K users UNK Lexicon- based, XGBoostF1-score: 93.22% Wu et al. (2023)Chinese suicide Dcard and PTTManual annotation Risk levels 2K posts U... | https://arxiv.org/abs/2505.15556v1 |
Online forumsPHQ-9, Manual an- notationPHQ-9 score, PHQ-9 symptoms60 users, 28K postsUNK BERT- basedAccuracy: 68.3% Jung et al. (2023)Korean suicide Twitter Manual annotation Binary 20k posts UNK Metadata, word count, XGBoostF1-score: 83.57% Cha et al. (2022)Korean, Japanese, Englishdepression Twitter, Ev- erytimeLexic... | https://arxiv.org/abs/2505.15556v1 |
arXiv:2505.15561v1 [cs.CL] 21 May 2025Do RAG Systems Suffer From Positional Bias? Florin Cuconasu1,2*†, Simone Filice2‡ Guy Horowitz2, Yoelle Maarek2, Fabrizio Silvestri1 1Sapienza University of Rome,2Technology Innovation Institute Abstract Retrieval Augmented Generation enhances LLM accuracy by adding passages re- tr... | https://arxiv.org/abs/2505.15561v1 |
neous presence of relevant and highly distracting passages near the top of the retrieval ranking dras- tically reduces the impact of the positional bias, since it penalizes, in turn, both passage types. Following these findings, we empirically demon- strate that strategies to rearrange the passages in the prompt based ... | https://arxiv.org/abs/2505.15561v1 |
compress the prompt. 3 Experimental Setup Benchmarks and Models. We run experi- ments using the following commonly used public question-answering benchmarks: PopQA (Mallen et al., 2023) and the KILT version (Petroni et al., 2021a) of Natural Questions (NQ) (Kwiatkowski et al., 2019), and TriviaQA (Joshi et al., 2017). ... | https://arxiv.org/abs/2505.15561v1 |
retrieval pipeline, we compute the distract- ing effect of the retrieved irrelevant passages and assume DE=0 for relevant passages. Fig. 1c re- ports the DE of the most distracting passage among the top- kpositions (MaxDE), while Fig. 1d re- ports the mean DE considering the top- kpositions 3Exact prompts are provided ... | https://arxiv.org/abs/2505.15561v1 |
the model favors certain positions regardless of passage relevance. Table 1 further validates this point by showing accuracy when placing a hard distractor at position 3 (lowest DE) versus position 5 (highest DE). We observe an average decrease of about 6 accuracy points compared to using only weak distractors (first r... | https://arxiv.org/abs/2505.15561v1 |
positions is compensated by the unintended tendency to put in the same slots highly distracting passages. 6 Conclusions Our work demonstrates that while positional bias exists in current LLMs, its impact is minimal in re- alistic RAG settings: random ordering of retrieved passages yields statistically equivalent accura... | https://arxiv.org/abs/2505.15561v1 |
Alan Schel- ten, Alex Vaughan, Amy Yang, Angela Fan, Anirudh Goyal, Anthony Hartshorn, Aobo Yang, Archi Mi- tra, Archie Sravankumar, Artem Korenev, Arthur Hinsvark, and 542 others. 2024. The llama 3 herd of models. Preprint , arXiv:2407.21783. Jiawei Gu, Xuhui Jiang, Zhichao Shi, Hexiang Tan, Xuehao Zhai, Chengjin Xu, ... | https://arxiv.org/abs/2505.15561v1 |
The Twelfth International Conference on Learning Repre- sentations . Nelson F. Liu, Kevin Lin, John Hewitt, Ashwin Paran- jape, Michele Bevilacqua, Fabio Petroni, and Percy Liang. 2024. Lost in the middle: How language mod- els use long contexts. Transactions of the Association for Computational Linguistics , 12:157–17... | https://arxiv.org/abs/2505.15561v1 |
for Compu- tational Linguistics: EMNLP 2024 , pages 12545– 12556, Miami, Florida, USA. Association for Com- putational Linguistics. Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, Hao Zhang, Joseph E. Gonzalez, and Ion Stoica. 2023. Judgi... | https://arxiv.org/abs/2505.15561v1 |
penalizing correct answers that use different phrasing than the reference, en- suring our effectiveness metrics genuinely reflect the model’s ability to extract and utilize informa- tion rather than simply mimic exact answer formats. For example, if the ground truth answer to “What is the population of Yokyo?” is “14 m... | https://arxiv.org/abs/2505.15561v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.