text string | source string |
|---|---|
in a mixed- advantage formulation with the indicator vector w. ∇θJ(θ|λ) =Est,at∼π(st−1)[Aπθ λ(st, at)∇θlogπ θ(st, at)] where Aπθ λ(st, at) =NX i=1wiAπθ i(st, at) (19) E Experiment Details E.1 Datasets Details We utilize the following dataset for training and evaluation. For Helpful Assistant task, we utilize “hh-rlhf” ... | https://arxiv.org/abs/2505.20925v1 |
[5] and published it on our depository. Free-merging: we reproduce Free-merging according to https://github.com/Zhengsh123/FREE- Merging We faithfully reproduced the these baselines using their code, replicating their experimental setup and benchmarks as described in the original papers. For MORLHF, we only train on a ... | https://arxiv.org/abs/2505.20925v1 |
of the policy gradient ∇θk+1J(θk+1|λ)metioned in 3.2 aligns with the gradient of TCH scalarazation ∇θk+1L(θk+1|λ) L(θk+1|λ) =Ex∼D[Vπk+1 θ λ(x)]−Ex∼D[Vπ∗ λ(x)] (22) =Ex∼D[Vπk+1 θ λ(x)]−Ex∼D[Vπk θ λ(x)] +Ex∼D[Vπk θ λ(x)]−Ex∼D[Vπ∗ λ(x)] (23) =Ex∼D,τ∼πk+1(x)[∞X t=1γt(r(st, at) +γVπk θ λ(st+1))−Vπk θ λ(st))] +L(θk|λ)(24) =E... | https://arxiv.org/abs/2505.20925v1 |
arXiv:2505.20933v1 [cs.CL] 27 May 2025Information-Theoretic Complementary Prompts for Improved Continual Text Classification Duzhen Zhanga,∗, Yong Renc,∗, Chenxing Lic, Dong Yud, Tielin Zhangb,∗∗ aMohamed bin Zayed University of Artificial Intelligence, Abu Dhabi, UAE bCenter for Excellence in Brain Science and Intelli... | https://arxiv.org/abs/2505.20933v1 |
PT shifts the focus from modifying the model’s parameters to crafting prompts that guide the model’s learning for specific tasks. These prompts encapsulate task-specific knowledge, making PT more effective in utilizing frozen PLMs compared to traditional fine-tuning approaches. In the context of CTC, harnessing PT to a... | https://arxiv.org/abs/2505.20933v1 |
across sequential tasks. Our contributions can be summarized as follows: •We introduce InfoComp for CTC, which integrates P-Prompt and S- Prompt to effectively capture task-specific and task-invariant knowl- edge, respectively. This method is easy to implement and eliminates the need for data replay, making it particul... | https://arxiv.org/abs/2505.20933v1 |
addition, the prompt list remains constant in length during CL, and our information theory framework promotes the learning of more informative prompts. 2.3. Information Theory Approaches in Natural Language Processing Information theory, pioneered by Claude Shannon in the mid-20th cen- tury, provides a mathematical fra... | https://arxiv.org/abs/2505.20933v1 |
training on that task. As a result, the learning objective across all tasks is defined as: max ΘnX k=1mkX i=1logpΘ(Yk i|Xk i) (1) The simplest method for CL is finetuning, where the PLM sequentially minimizes the loss for each task Tk,k∈ {1..n}, by updating all its parameters Θ (including the parameters of PLM and clas... | https://arxiv.org/abs/2505.20933v1 |
CF. Meanwhile, a shared S-Prompt is utilized across all tasks to capture task-invariant knowledge, facilitating FKT. The training objective for the k-th task Tkis formulated as follows: Lk(θPk, θS, θhead k) =−mkX i=1logp(Yk i|[Pk, S, Xk i], θ, θhead k, θPk, θS), (4) 8 S-Prompt Classifier 1 P-Prompt1Input Text TokensTrm... | https://arxiv.org/abs/2505.20933v1 |
the MI between the encoded representations of the same input using both the current and previous S-Prompt states. This strategy ensures better preservation of task-invariant knowledge, enhances FKT, and supports the continuity of shared knowledge across sequential tasks. Let the PLM encoder be denoted as F(·). Given an... | https://arxiv.org/abs/2505.20933v1 |
Standard CTC Benchmark. Building on the methodology outlined in Prog- Prompt [13], we initially assess our proposed approach using the widely rec- ognized CTC benchmark, which comprises five text classification datasets introduced by [68]: AG News, Amazon reviews, DBpedia, Yelp reviews, and Yahoo Answers. In line with ... | https://arxiv.org/abs/2505.20933v1 |
)imdb ) sst2 )dbpedia )ag )yahoo )multirc )boolq )wic Table 2: Seven distinct task sequence orders were used for the CTC experiments. Orders 1 through 4 correspond to the standard CTC benchmark, while Orders 5 through 7 represent long-sequence settings involving 15 tasks. tion of the IMDB movie reviews dataset. The tas... | https://arxiv.org/abs/2505.20933v1 |
with prior works, including IDBR [3], MBPA++ [2], and ProgPrompt [13], we adopt the pre-trained bert-base-uncased model as the PLM backbone. Our implementation relies on PyTorch [77] along with the HuggingFace Transformers library [78]. For the standard CTC benchmark, we use the official datasets from [68]3, consistent... | https://arxiv.org/abs/2505.20933v1 |
2.3, 2.0, and 2 .3 over ProgPrompt, and 0 .4, 1.3, 0.9, and 1 .0 3http://goo.gl/JyCnZq 4https://github.com/huggingface/datasets 5The prompt length refers to the sequence length of tokens in a prompt. 17 1 2 3 4 5 Step80859095100Acc. Order3 ProgPrompt InfoComp 1 2 3 4 5 Step6570758085Acc. Order4 ProgPrompt InfoCompFigur... | https://arxiv.org/abs/2505.20933v1 |
of 57 .9, 69.6, and 71 .2 for the few-shot scenarios of 20, 200, and 1000 samples per class, re- spectively. This results in improvements of 3 .7, 2.7, and 1 .9 over the previous SOTA method, ProgPrompt. 6.3. Ablation Study This subsection assesses the effectiveness of the individual components of our InfoComp method i... | https://arxiv.org/abs/2505.20933v1 |
for further validation of its efficacy across a broader range of CTC benchmarks. References [1] J. Devlin, M.-W. Chang, K. Lee, K. Toutanova, BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding, in: Pro- ceedings of the 2019 Conference of the North American Chapter of the Association for Co... | https://arxiv.org/abs/2505.20933v1 |
O’Reilly, Why there are complementary learning systems in the hippocampus and neocortex: in- sights from the successes and failures of connectionist models of learning and memory., Psychological review 102 (3) (1995) 419. [16] D. Kumaran, D. Hassabis, J. L. McClelland, What learning systems do intelligent agents need? ... | https://arxiv.org/abs/2505.20933v1 |
gradient descent for continual learning, in: International Conference on Artificial Intelli- gence and Statistics, PMLR, 2020, pp. 3762–3773. [32] S. Hou, X. Pan, C. C. Loy, Z. Wang, D. Lin, Learning a unified clas- sifier incrementally via rebalancing, in: Proceedings of the IEEE/CVF Conference on Computer Vision and ... | https://arxiv.org/abs/2505.20933v1 |
M. Huang, Continual Prompt Tuning for Dialog State Tracking, in: Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2022, Dublin, Ireland, May 22-27, 2022, 2022, pp. 1124–1137. [48] Z. Wang, Y. Liu, T. Ji, X. Wang, Y. Wu, C. Jiang, Y. Chao, Z. Han, L. Wa... | https://arxiv.org/abs/2505.20933v1 |
Theo- retic Perspective, in: 9th International Conference on Learning Repre- sentations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021, 2021. 27 [63] Z. Chi, L. Dong, F. Wei, N. Yang, S. Singhal, W. Wang, X. Song, X. Mao, H. Huang, M. Zhou, InfoXLM: An Information-Theoretic Framework for Cross-Lingual Language Model... | https://arxiv.org/abs/2505.20933v1 |
arXiv:2505.20937v1 [cs.CL] 27 May 2025On VLMs for Diverse Tasks in Multimodal Meme Classification Deepesh Gavit, Debajyoti Mazumder, Samiran Das, Jasabanta Patro Department of Data Science and Engineering Indian Institute of Science Education and Research, Bhopal, India {gavit20, debajyoti22, samiran, jpatro}@iiserb.ac... | https://arxiv.org/abs/2505.20937v1 |
Meme Classification: We benchmark the performance of various VLMs using diverse prompting strategies. •Evaluating LoRA Fine-Tuning: We sys- tematically explore diverse prompting strate- gies such as Zero Shot (ZS), Zero Shot Chain-of-Thought (ZSC), Few Shot (FS), Few Shot Chain-of-Thought (FSC), and parameter- efficien... | https://arxiv.org/abs/2505.20937v1 |
giving VLM-generated meme understanding as input. We reported the details of the individual experiment in subsections. 3.1 Experiment-1 (Exp1): Prompting VLMs using various methods In this experiment, we evaluated various prompting strategies such as ZS,ZSC,FS, and FSC to clas- sify memes. While in the ZS setting we pr... | https://arxiv.org/abs/2505.20937v1 |
The highest F1 score for each task is in blue (column-wise). MDL Memotion MAMI Avg. HM SR OF SN MV MG MGT LV 30.21 17.58 11.57 29.83 49.16 64.16 24.53 32.43 QW 36.35 20.97 29.89 30.34 51.58 61.33 27.83 36.89 LM 33.95 23.17 26.34 31.72 49.33 58.98 22.13 35.08 IB 23.03 15.98 18.98 28.61 35.19 47.33 19.83 26.99 Avg. 30.89... | https://arxiv.org/abs/2505.20937v1 |
29.25 34.77 33.20 26.97 45.63 35.23 45.59 56.69 53.21 53.97 68.26 70.85 70.40 56.23 56.12 56.33 44.98 IBZS 25.15 16.24 21.47 24.62 18.94 29.34 23.20 22.20 17.87 29.63 30.78 25.42 49.35 36.75 42.90 60.15 62.81 60.86 52.34 53.82 52.10 36.00 ZSC 32.25 30.10 28.37 31.02 28.89 30.03 26.69 28.98 29.28 43.46 33.83 35.39 55.21... | https://arxiv.org/abs/2505.20937v1 |
relatively small num- ber of parameters from very few selective layers. Our proposed CoVExFiL approach, which inte- grates VLMs and LLMs, performs well in multiple tasks. These findings underscore the effectiveness of prompting and distillation for improving meme understanding. Our analysis shows that VLMs grasp meme c... | https://arxiv.org/abs/2505.20937v1 |
chatbot impressing gpt-4 with 90%* chatgpt quality. Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. 2024. Scaling instruction-finetuned language models. Journal of Machine Learning Research , 25(70):1–53. Shiv Shankar Das.... | https://arxiv.org/abs/2505.20937v1 |
2024a. Improved baselines with visual instruc- tion tuning. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition , pages 26296–26306. Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. 2024b. Lla- vanext: Improved reasoning, ocr, and world knowl- e... | https://arxiv.org/abs/2505.20937v1 |
cyberbullying , pages 32–41. Chris Van der Lee, Albert Gatt, Emiel Van Miltenburg, and Emiel Krahmer. 2021. Human evaluation of automatically generated text: Current trends and best practice guidelines. Computer Speech & Language , 67:101151. Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc... | https://arxiv.org/abs/2505.20937v1 |
pre- dictions (Afridi et al., 2021). Several specialized corpora have been released to capture the rich va- riety of figurative, humorous, and harmful content in memes. Xu et al. (Xu et al., 2022) introduced MET-meme, a collection targeted at metaphorical memes, demonstrating that models struggle when literal text mask... | https://arxiv.org/abs/2505.20937v1 |
learning (Lee and Shen, 2022), and external knowl- edge integration (Pramanick et al., 2021), have been shown to improve meme classification per- formance and interpretability. (Jha et al., 2024) proposed a MemeGuard framework fine-tuned with a dedicated VLM for harmful meme interpretation,applied a multi-modal knowled... | https://arxiv.org/abs/2505.20937v1 |
the type of misogynist content. The memes in this dataset were collected from two sources- social media platforms such as Twitter and Reddit, and meme-sharing websites like 9GAG, Imgur, and KnowYourMeme. These memes were collected by using specific hashtags such as #girl, #girlfriend, #women, #feminist, threads, and di... | https://arxiv.org/abs/2505.20937v1 |
shaming, stereotype, objectification, violence). Qwen2-VL: Qwen2-VL (Bai et al., 2024) is an advanced vision-language model, specifically de- signed to employ a mixed training regimen, both image and text. It utilized Vision Tr B.2.1 Configuration of LLMs To complement the vision-language understanding from VLMs, we us... | https://arxiv.org/abs/2505.20937v1 |
In Experiment 1, we explored a range of prompt- ing strategies to evaluate their effectiveness in meme understanding. These included Zero-Shot (ZS), Zero-Shot Chain-of-Thought (ZSC), Few- Shot (FS), and Few-Shot Chain-of-Thought (FSC) prompting. The specific prompts used in each strat- egy, along with their correspondi... | https://arxiv.org/abs/2505.20937v1 |
and v_proj layers in InstructBLIP, targeting its core multimodal fusion mechanism to steer cross-modal alignment while minimizing additional memory and compute overhead. MEME VLM Meme Understanding CoT Label: hateful_offensive Predicted: not_offensive Visual Content Analysis The meme consists of two images arranged ver... | https://arxiv.org/abs/2505.20937v1 |
singer than the pop star Justin Bieber. The humor comes from the absurdity of the comparison?given that Justin Bieber is a well-known and popular singer, it is unlikely that someone on a TV show would claim to be a better singer than him. The juxtaposition of a fictional character's claim with a real-world celebrity's ... | https://arxiv.org/abs/2505.20937v1 |
of the best models and prompting techniques using qualitative and quantitative methods. C.1 Quantitative analysis We evaluate the textual meme understanding gen- erated by the VLMs in Experiment-3 using stan- dard metrics—BLEU (Post, 2018), ROUGE (Lin, 2004), and BERTScore (Zhang et al., 2019)—fol- lowing best practice... | https://arxiv.org/abs/2505.20937v1 |
Our observations are as follows: In these cases, the explanations for incorrect classifications were often lexically similar to those of correct classifications, suggesting that models struggled to distinguish subtle linguistic and con- textual cues in SR and OF. The results indicate that SR and OF resulted in the lowe... | https://arxiv.org/abs/2505.20937v1 |
content, indicating a significant gap in detecting subtle hate cues. •Example 2: In the sarcasm task, a meme was labeled as very_twisted but misclassified as not_sarcastic. The meme features a historical Nazi rally with a man defiantly refusing to perform the Nazi salute, surrounded by oth- ers, including Hitler. The a... | https://arxiv.org/abs/2505.20937v1 |
Research Community Perspectives on “Intelligence” and Large Language Models Bertram Højer1, Terne Sasha Thorn Jakobsen2, Anna Rogers1, Stefan Heinrich1 1IT University of Copenhagen,2University of Copenhagen GitHub: bertramhojer/perspectives-on-intelligence Abstract Despite the widespread use of “artificial intelli- gen... | https://arxiv.org/abs/2505.20959v1 |
to its broader general knowledge and advanced reasoning capabil- ities." and the o1 class of models " think before they answer ".arXiv:2505.20959v1 [cs.CL] 27 May 2025 researchers mislead themselves by using terms like “understanding”, which have long been criticized as “wishful mnemonics” (McDermott, 1976). In partic-... | https://arxiv.org/abs/2505.20959v1 |
about the current state of AI, but these have mostly focused on impact predictions rather than beliefs about the actual tech- nology. For example, the recent Pew survey on how US adults and experts view AI (Pasquini et al., 2025) does not elicit how respondents define or use that term. Neither are definitions the focus... | https://arxiv.org/abs/2505.20959v1 |
no study attempting to elucidate the notion of “in- telligence” can likely ever claim to be exhaustive. This is due to the inherent complexity of the con- cept and the difficulty we observe for researchers in considering even a small set of criteria. 2.3 Relevant NLP Research Directions There are many research directio... | https://arxiv.org/abs/2505.20959v1 |
survey was distributed through social me- dia (Bluesky and LinkedIn) as well as the mailing lists seen in Table 8 following the approach used in similar surveys (Michael et al., 2023). The mailing lists were selected to collect diverse responses from researchers in multiple fields in which the term “in- telligence” is ... | https://arxiv.org/abs/2505.20959v1 |
Perspectives on the Field Inspired by Michael et al. (2023), we pose Q12: “To what extent is your notion of intelligence shared by other researchers in your field? ” This allows us to estimate to what degree real disagreement on “intelligence” matches the perceived disagree- ment. We further ask the participants which ... | https://arxiv.org/abs/2505.20959v1 |
compute the phi(ϕ) coefficients using 2x2 con- tingency tables for each pair of criteria. For binary data the ϕcoefficient is equivalent to the Pearson correlation coefficient. This analysis thus assesses whether respondents are likely to select a criterion A if they have selected a criterion B. Figure 1 illustrates th... | https://arxiv.org/abs/2505.20959v1 |
only 32-36% of respondents stated that the current sys- tems are lacking in goal achievement, problem solv- ing, & understanding - see Appendix A. The latter is surprising, given that many respondents come from NLP & Computational Linguistics communi- ties, where numerous studies and position papers have highlighted th... | https://arxiv.org/abs/2505.20959v1 |
are more likely to consider the current systems intelligent ( ≈48% of postdocs agree or strongly agree, for students that number is ≈56%). Senior researchers demonstrate greater skepticism, with ≈77% disagreeing or strongly disagreeing. Research Area Effects. We observe a significant effect of research area ( ϕc= 0.270... | https://arxiv.org/abs/2505.20959v1 |
ϕc=.216, p=.01). This is an interesting observation, given the above finding that the respondents identifying their primary field of research as “AI” are among the most skeptical about the current systems. This suggests a dissociation between “AI” as a research field and “intelligent technology” as a research goal. Per... | https://arxiv.org/abs/2505.20959v1 |
a similar skepticism regarding neural approaches being enough to achieve general-purpose intelligence matching that of humans (Rossi, 2025, p. 74). for the recent debate on the relation between the fields of “NLP” and “AI” (Bender, 2024; Goldberg, 2024; Mortensen, 2024). The majority of survey respondents do not see th... | https://arxiv.org/abs/2505.20959v1 |
proposal to consider LLMs as “a new type of intelligence” by McCoy et al. (2024)). This possibility necessitates further research. Definitions of intelligence criteria. Many of the criteria we listed, especially “reasoning” and “un- derstanding”, are themselves lacking clear defini- tions, and merit similar surveys. He... | https://arxiv.org/abs/2505.20959v1 |
Not an AI Conference. Emily M. Bender and Batya Friedman. 2018. Data Statements for Natural Language Processing: TowardMitigating System Bias and Enabling Better Science. Transactions of the Association for Computational Linguistics , 6:587–604. Emily M. Bender and Alexander Koller. 2020. Climbing towards NLU: On meani... | https://arxiv.org/abs/2505.20959v1 |
Crawford. 2020. Datasheets for Datasets. arXiv:1803.09010 [cs] . Yoav Goldberg. 2024. ACL is not an AI Conference (?). Suchin Gururangan, Swabha Swayamdipta, Omer Levy, Roy Schwartz, Samuel Bowman, and Noah A. Smith. 2018. Annotation artifacts in natural language infer- ence data. In Proceedings of the 2018 Conference ... | https://arxiv.org/abs/2505.20959v1 |
McMillan- Major, Ezinwanne Ozoani, Nazneen Rajani, Tristan Thrush, Yacine Jernite, and Douwe Kiela. 2023. Mea- suring data. Preprint , arXiv:2212.05129. Melanie Mitchell. 2021. Why AI is harder than we think. In Proceedings of the Genetic and Evolution- ary Computation Conference , GECCO ’21, page 3. Association for Co... | https://arxiv.org/abs/2505.20959v1 |
Research . Featured Certification. Joshua B Tenenbaum, Charles Kemp, Thomas L Grif- fiths, and Noah D Goodman. 2011. How to grow a mind: Statistics, structure, and abstraction. science , 331(6022):1279–1285. Alan M. Turing. 1950. I.—COMPUTING MACHIN- ERY AND INTELLIGENCE. LIX(236):433–460. Marcel Van Gerven. 2017. Comp... | https://arxiv.org/abs/2505.20959v1 |
survey questions to- gether with answer options. The survey is also available as a form in the supplementary materials. C Appendix: Mailing lists Table 8 provides a full list of mailing lists to which the survey was sent. D Appendix: Geographic Distribution of Respondents Figure 7 provides an overview of the geographic... | https://arxiv.org/abs/2505.20959v1 |
protein structure prediction), Current autonomous robotic systems (e.g. self-driving cars), Earlier chatbot system (e.g. customer support bots), Current LLM-based chatbot systems (e.g. ChatGPT), Current autonomous LLM-based agents (e.g. based on ChatGPT), None of the above Q12 To what extent, is your notion of intellig... | https://arxiv.org/abs/2505.20959v1 |
Carbon based emulation because other loose forms of consciousness becomes very unbounded and we don’t have sufficient tools or language yet to describe or understand it. C2 Abstraction: being able to make a new concept that abstracts the core similarities of the underlying group. If using LLMs as an example, then imagi... | https://arxiv.org/abs/2505.20959v1 |
arXiv:2505.20963v1 [cs.CL] 27 May 2025Context-Aware Content Moderation for German Newspaper Comments Felix Krejca[0009−0002−0549−0865]1 ,2, Tobias Kietreiber[0009−0003−4396−1135]1, Alexander Buchelt[0000−0003−3851−6320]1, and Sebastian Neumaier[0000−0002−9804−4882]1 1St. Pölten University of Applied Sciences, Austria f... | https://arxiv.org/abs/2505.20963v1 |
Turbo in terms of accuracy and F1-score. The remainder of this paper is structured as follows: Section 2 provides a re- view of the relevant literature and theoretical foundations, defining key concepts and presenting a taxonomy of existing methods in the field. Section 3 outlines the methodology, including data prepro... | https://arxiv.org/abs/2505.20963v1 |
study. 4Der Standard Community Guidelines, https://www.derstandard.at/ communityrichtlinien, last accessed 2025-02-06 4 F. Krejca, et al. The Rheinische Post dataset [2] consists of user comments from the Rheinis- chePostwebsite,5annotatedbymoderatorsandcrowdworkersusingacustomla- belingschemathatincludesoffensivecateg... | https://arxiv.org/abs/2505.20963v1 |
or abusive language, rather than addressing content moderation as a whole [2,26]. This distinction is crucial, as different definitions and conceptual scopes lead to variations in annotation guidelines and dataset labeling practices in this research field [5]. Paper Classifier Word Representation Schabus et al. 2017 [2... | https://arxiv.org/abs/2505.20963v1 |
“deleted”. The Path,Titleand Comment variables are text variables, so we will need further preprocessing on them. Following the suggestion of [11], we make use of encoding, cleaning, tokenization, stop word removal, and lemmatization. After importing, the data is further cleaned by converting all text to lowercase and ... | https://arxiv.org/abs/2505.20963v1 |
learning methods additionally use LSTM and CNN architectures. As embeddings we use pre-trained word embeddings from fasttext [10]. We use three models with simple LSTM architectures, three with more advanced LSTM architectures, and three with CNN architectures, which all differ in the inputs they take to make their pre... | https://arxiv.org/abs/2505.20963v1 |
0.725 0.802 0.725 0.695 0.757 / base_LSTM_title_path 0.703 0.777 0.698 0.678 0.720 / base_LSTM_title 0.699 0.772 0.687 0.683 0.693 / naive_bayes 0.678 0.743 0.679 0.677 0.681 / logistic_regression 0.663 0.726 0.656 0.670 0.643 / base_LSTM 0.661 0.728 0.667 0.629 0.710 / GPT_mod_title_path 0.635 / 0.651 0.609 0.700 75 G... | https://arxiv.org/abs/2505.20963v1 |
but are consistently outperformed by more advanced models such as base_LSTM_title andbase_LSTM_title_path . It is also worth noting that there are no significant differences in performance between the advanced LSTM and CNN architectures. Comparison with Previous Literature. When compared to previous work in the field, ... | https://arxiv.org/abs/2505.20963v1 |
9. Jose,M.,Anthony,J.,Joseph,J.V.,Thomas,J.,Thomas,S.B.:Areviewofmachine learning and deep learning approaches for offensive text detection. International Journal on Emerging Research Areas (IJERA) 04(02), 47–50 (2025). https://doi. org/10.5281/zenodo.14651005, https://doi.org/10.5281/zenodo.14651005 10. Joulin, A., Gr... | https://arxiv.org/abs/2505.20963v1 |
Trapp, M.: One million posts: A data set of german on- line discussions. In: Proceedings of the 40th International ACM SIGIR Conference on Research and Development in Information Retrieval. p. 1241–1244. SIGIR ’17, New York, NY, USA (2017). https://doi.org/10.1145/3077136.3080711 27. Spence, R., Bifulco, A., Bradbury, ... | https://arxiv.org/abs/2505.20963v1 |
arXiv:2505.20966v1 [cs.CL] 27 May 2025Personalized Query Auto-Completion for Long and Short-Term Interests with Adaptive Detoxification Generation Zhibo Wang Kuaishou Technology Beijing, China wangzhibo07@kuaishou.comXiaoze Jiang∗ Kuaishou Technology Beijing, China jiangxiaoze@kuaishou.comZhiheng Qin Independent Beijin... | https://arxiv.org/abs/2505.20966v1 |
V.2 (KDD ’25), Au- gust 3–7, 2025, Toronto, ON, Canada. ACM, New York, NY, USA, 11 pages. https://doi.org/10.1145/3711896.3737256 1 Introduction Query auto-completion (QAC) is a crucial feature in modern search systems. Generally speaking, given the query prefix (i.e. the sub- string of a word or a sentence), the searc... | https://arxiv.org/abs/2505.20966v1 |
methods involve user feedback, filter toxic contents by the rule or conduct genera- tion upon limited query templates [ 6,13,16,38]. These efforts are unsuccessful in achieving flexible generation and require constant maintenance. There are also some works that focus on non-toxic text generation for long texts by detox... | https://arxiv.org/abs/2505.20966v1 |
[Reject] token are entirely non-toxic. As a result, the model is able to adaptively complete query based on different inputs. There are four main contributions in our work: (1) We emphasize the two limitations of QAC generation task for online applications. The first is the lack of hierarchical interest representation,... | https://arxiv.org/abs/2505.20966v1 |
of user historical behavior information while meeting the requirements for online inference efficiency. Detoxification in QAC. Some studies have focused on detoxifica- tion in free-form generation by detoxifying clean datasets [ 15,30], refining the decoding process [ 5,24], or employing reinforcement learning-based ap... | https://arxiv.org/abs/2505.20966v1 |
past searched query𝑏𝑙𝑖from the𝐵𝑙, ande𝑏𝑙𝑖is the encoded long-term representa- tion of𝑏𝑙𝑖(𝑖=1,...,𝐿 ,𝐿is the length of 𝐵𝑙). Short-Term Interests Encoding Unlike long-term interests, short-term interests reflect the user’s immediate feedback. They precisely capture the user’s immediate search intentions, ... | https://arxiv.org/abs/2505.20966v1 |
after generation will incur extra latency for real-time online applications. As illustrated in Figure 2, to achieve end-to-end online detoxifi- cation generation, we first introduce a Detoxification Expert. This expert is implemented as a discriminative transformer model, al- lowing for the immediate evaluation of the ... | https://arxiv.org/abs/2505.20966v1 |
is capable of providing genuine online real-time inference while meeting time constraints. As illustrated in Figure 3, the inputs for online serving consist of three components: the user’s prefix, the Real-time GSU (which provides the user’s recently searched queries), and the Memory Bank (which stores the user’s long-... | https://arxiv.org/abs/2505.20966v1 |
is precisely calibrated to maintain both the precision and recall rates of our Detoxification Expert above 90 %. For each prefix, LaD generates up to 4 completions. For the KSQAC dataset, the vocabulary is constructed using Chinese characters, resulting in a total vocabulary size of 53,704. The experiments are conducte... | https://arxiv.org/abs/2505.20966v1 |
30.83 % 33.34 % 23.94 % 0.0727 9.17 % The compared models are shown in Table 2, which can be divided into two major categories. (1) The models in the first block of Table 2 are the models without considering of detoxification. In detail, we compare our model with statistical methods such as MPC [3], generative models t... | https://arxiv.org/abs/2505.20966v1 |
[47] 9.49% 21.70 % 6.04% 0.2650 37.25 % DAC [28] 7.14% 19.10 % 4.65% 0.2294 33.25 % LaD w/o AD 13.58 % 26.10 % 9.80% 0.3596 50.53 % LaD 4.18% 18.53 % 2.65% 0.1274 18.02 % Table 4: Ablation study of Long-Short Interests Hierarchical Capturing, where “S” denotes the length of short-term in- terests, “L” indicates the len... | https://arxiv.org/abs/2505.20966v1 |
the lowest performance in the metrics. This underscores the significance of personalized information in QAC tasks. (2)SL-10 ,SL-30 , and SL-50 are models that incorporate short- term interests. As the number of short-term interests increases (from 1 to 5), the R @4, BLEU, and MRR metrics improve by 2.05 %, 1.1%, and 1.... | https://arxiv.org/abs/2505.20966v1 |
detoxification tailored to various different inputs. (4)LDPO w Online Reject not only improves the R @4, BLEU and MRR metric but also significantly reduces the UAmaxT and UP- rob compared to LDPO w Offline Reject . It proves the superiority of the online strategy. (5)LaD achieves the best performance in UAmaxT and UPro... | https://arxiv.org/abs/2505.20966v1 |
higher value indicates a greater number of generated queries that are unrelated to the prefix. Duplicate : It assesses the presence of meaningless character repetitions in generated queries. A higher value suggests a larger number of queries containing such repetitions. The results are shown in Table 6, LaD significant... | https://arxiv.org/abs/2505.20966v1 |
across the entire app. As shown in the third block of Table 7, LaD improves the Search Penetration Rate by 0.438 %, indicating that it significantly expands the user base for search and plays a crucial role in cultivating users’ search habits. LaD (Gen + AD + LS) outperforms Gen + AD across all metrics, highlighting th... | https://arxiv.org/abs/2505.20966v1 |
the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 3 (Industry Papers) . 208–215. [9]Luciano Floridi and Massimo Chiriatti. 2020. GPT-3: Its Nature, Scope, Limits, and Consequences. Minds and Machines 30 (2020), 681–694. [10] Bruno M Fonseca, Paulo Golgher, ... | https://arxiv.org/abs/2505.20966v1 |
International Conference on Learning Representations. [24] Ben Krause, Akhilesh Deepak Gotmare, Bryan McCann, Nitish Shirish Keskar, Shafiq Joty, Richard Socher, and Nazneen Fatema Rajani. 2021. GeDi: Generative Discriminator Guided Sequence Generation. In Findings of the Association for Computational Linguistics: EMNL... | https://arxiv.org/abs/2505.20966v1 |
In Proceedings of the International AAAI Conference on Web and Social Media , Vol. 10. 687–690. [39] A Vaswani. 2017. Attention Is All You Need. Advances in Neural Information Processing Systems (2017). [40] Po Wei Wang, Huan Zhang, Vijai Mohan, Inderjit S Dhillon, and J Zico Kolter. 2018. Realtime query completion via... | https://arxiv.org/abs/2505.20966v1 |
toxic comment classification model, to calculate the toxicity score. A score greater than 0.5 indicates that a query is considered toxic. Based on this, the testset is split into two parts, 𝑇𝑡𝑒𝑠𝑡indicating toxic and𝑁𝑇𝑡𝑒𝑠𝑡indicating non-toxic, to evaluate the toxicity of the model’s generated outputs. We use ... | https://arxiv.org/abs/2505.20966v1 |
27.47 % 0.0310 1.93 % DAPT [15] 41.73 % 51.35 % 36.73 % 0.0206 0.69 % Quark [27] 31.12 % 39.26 % 24.51 % 0.0173 0.68 % PPO [47] 38.38 % 48.21 % 32.96 % 0.0221 0.92 % DAC [28] 30.42 % 41.77 % 24.74 % 0.0182 0.69 % LaD w/o AD 36.64 % 48.42 % 32.03 % 0.0238 1.14 % LaD 41.73 % 51.42 % 36.61 % 0.0167 0.23 % excellent genera... | https://arxiv.org/abs/2505.20966v1 |
arXiv:2505.20971v1 [cs.CL] 27 May 2025Reason-Align-Respond: Aligning LLM Reasoning with Knowledge Graphs for KGQA Xiangqing Shen, Fanfan Wang, and Rui Xia* School of Computer Science and Engineering, Nanjing University of Science and Technology, China {xiangqing.shen, ffwang, rxia}@njust.edu.cn Abstract LLMs have demon... | https://arxiv.org/abs/2505.20971v1 |
answering, e.g, knowledge graph question answer- ing (KGQA), has gained interest as an effective strategy to mitigate hallucinations and enhance in- terpretability. The existing LLM-based KGQA studies broadly fall into two main categories: Training-free Agent Exploration methods (Sun et al., 2024) and Training-based Pa... | https://arxiv.org/abs/2505.20971v1 |
turn promotes the generation of high-quality re- sponses, ultimately forming a stable closed loop. We conduct extensive evaluation across multi- ple benchmarks, including WebQSP, CWQ, Com- monsenseQA (CSQA), and MedQA, using Free- base, ConceptNet and a medical KG as knowl- edge graphs. The results demonstrate the effe... | https://arxiv.org/abs/2505.20971v1 |
path zp, working together to answer the question qbased on a KG G: •Reasoning Chain zrdenotes a chain of dis- crete reasoning steps expressed in natural lan- guage, working together to address the ques- tionq. •Knowledge Path zpdenotes an interconnected path of knowledge triples extracted from a KG G. Since neither zrn... | https://arxiv.org/abs/2505.20971v1 |
=X zpw(a|q, z)pψ(z|G, q),(2) where Gdenotes the KG, qthe question, and zthe unified latent variable (combining the Reasoning Chain and Knowledge Path). Learning Objective. We aim to learn the param- eters (w, ψ)by maximizing the log-likelihood of the training data with respect to Eq. (2), written as: max w,ψO(w, ψ) = l... | https://arxiv.org/abs/2505.20971v1 |
Karypis, 2022) 76.4 70.9 52.9 47.8 KG+LLMKD-CoT (Wang et al., 2023a) 68.6 52.5 55.7 - EWEK-QA (Dehghan et al., 2024) 71.3 - 52.5 - ToG (GPT-4) (Sun et al., 2024) 82.6 - 68.5 - EffiQA (Dong et al., 2025) 82.9 - 69.5 RoG (Llama-2-7B) (Luo et al., 2024b) 85.7 70.8 62.6 56.2 GNN-RAG+RA (Mavromatis and Karypis, 2024) 90.7 7... | https://arxiv.org/abs/2505.20971v1 |
recall. For CSQA, a multiple-choice QA dataset, we use accuracy as the evaluation metric. Implementations. Our implementation uses Llama-3.1-8B (Meta, 2024) as the backbone for Reasoner, Aligner, and Responser. To enhance question decomposition, we pretrain both Reasoner and Aligner using 2,000 exemplars demonstrating ... | https://arxiv.org/abs/2505.20971v1 |
minimal computational overhead. Particularly noteworthy is the compari- son with GCR, the previous state-of-the-art method using KG constraints. Our approach achieves a 15.2% improvement in answer accuracy over GCR, with only a marginal increase in runtime from Rea- soning Chain generation. This modest overhead is well... | https://arxiv.org/abs/2505.20971v1 |
Reasoning Chains and Knowledge Paths generated by RARunder different iteration steps. Model CSQA MedQA GPT-4o-mini 91 75 GCR 94 79 RAR 94 80 Table 5: Zero-shot transferability to unseen KG. the Knowledge Path in the KG. In contrast, the later- stage model generates both Reasoning Chains and Knowledge Paths effectively,... | https://arxiv.org/abs/2505.20971v1 |
that RARsuccessfully bridges the gap between LLM reasoning and structured knowledge, offering a promising direction for build- ing reliable and interpretable QA systems. Limitations One limitation of our framework lies in the compu- tational overhead introduced by Reasoner. In cer- tain cases, especially for complex qu... | https://arxiv.org/abs/2505.20971v1 |
Natural Language Processing , pages 7646–7663, Miami, Florida, USA. Association for Computational Linguistics. Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. 2025. Deepseek-r1: In- centivizing reasoning capability in llms via reinforce- ment... | https://arxiv.org/abs/2505.20971v1 |
Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024 . OpenReview.net. Linhao Luo, Yuan-Fang Li, Gholamreza Haffari, and Shirui Pan. 2024b. Reasoning on graphs: Faithful and interpretable large language model reasoning. In International Conference on Learning Representa- tions . Linhao Luo, Zicheng Zhao... | https://arxiv.org/abs/2505.20971v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.