text string | source string |
|---|---|
factuality annotations for the Falcon 3B model. Each matrix is based on 76 randomly selected claims, comparing annotations produced by the model with those from human annotators. faithful unfaithful-contra unfaithful-neutral GPT Predicted Labelsfaithful unfaithful-contra unfaithful-neutralHuman Annotated Labels40 1 8 1... | https://arxiv.org/abs/2505.21072v2 |
12.50 13.00 12.25 .602 .512 .230 (b) Only claims with AlignScore < 0.5 Table 10: Results averaged across 4 QA datasets for Llama 3B considering only claims with high and low AlignScore. MethodLlama3b ROC-AUC ↑PR-AUC ↑PRR↑ MCP .538 .115 .028 PTrue .463 .112 .002 Perplexity .480 .092 -.068 MTE .580 .167 .122 CCP .585 .13... | https://arxiv.org/abs/2505.21072v2 |
.478 .460 .385 .504 .531 .516 .537 .436 .529 .432 .412 .517 .367 .369 .404 .505 .490 .514 .488 .514 .530 .525 .539 .497 .530 .486 .479 .503 .459 .436 .434 .456 .424 .446 .149 .467 .492 .465 .503 .277 .496 .401 .360 .481 .110 .120 .275 .454 .421 .431 .156 .454 .474 .469 .475 .298 .471 .390 .358 .472 .151 .201 .276 .462 ... | https://arxiv.org/abs/2505.21072v2 |
it might be more effective to use AlignScore without faithfulness calibration. 19 D.3 Impact of train size on FRANQ 100 150 200 250 300 350 400 450 500 Train Size0.000.050.100.150.200.25PRR PRR vs Train Size for Various UQ Methodss FRANQ no calibration FRANQ calibrated FRANQ condition-calibrated XGBoost (all UQ feature... | https://arxiv.org/abs/2505.21072v2 |
XGBoost (all UQ features) .05 XGBoost ( FRANQ features) .06 FRANQ no calibration .44 FRANQ calibrated .02 FRANQ condition-calibrated .03 (a) Long-form QA Llama 3B dataset.Method Mean ECE ↓ Max Sequence Probability .46 CCP .23 P(True) .71 Lexical Similarity .07 Degree Matrix .14 Eccentricity .46 Sum of Eigenvalues .54 N... | https://arxiv.org/abs/2505.21072v2 |
faithful to r) = AlignScore(c, r) = 0.98 2. MaxNLI(c) = max(0.44, 0.34, 0.99) = 0.99 3. ParametricKnowledge(c) = = = 0.52 · 0.66 · … · 0.32 = 3.5·10-15 = 0.44 = 0.34 = 0.99 52%66% 33% 44% 22% 2%83%3%69%87%13%100%0.1%32% 0.2% 0.6% 32%Token probabilities from parametric knowledge FRANQno calibration(c) = = 0.98 · 0.99 + ... | https://arxiv.org/abs/2505.21072v2 |
it just doesn't work well. LLM Answer: Determining which type of diabetes is worse is a complex task … Type 1 diabetes is a condition where the body either res ists the effects of insulin or doesn 't produce enough insulin to maintain a normal glucose level … Claim: Type 1 diabetes is a condition where the body either ... | https://arxiv.org/abs/2505.21072v2 |
LLMs Think, But Not In Your Flow: Reasoning-Level Personalization for Black-Box Large Language Models Jieyong Kim1∗Tongyoung Kim1∗Soojin Yoon1Jaehyung Kim1Dongha Lee1† 1Yonsei University {jieyong99,dykim,soojiny,donalee}@yonsei.ac.kr Abstract Large language models (LLMs) have recently achieved impressive performance ac... | https://arxiv.org/abs/2505.21082v2 |
e o f b r e a d I ' v e h a d s i n c e . . .Quer yFinal R esponse Final R esponse “ D i e t s h i f t ” - “ G r o w t h ” ( 4 . 6 7 )P r o a c t i v e a p p r o a c h t o h e a l t h .R easoning 2“ T a s t e ” - “ I n s i g h t ” ( 4 . 5 1 )F l a v o r q u a l i t y i s a s t r o n g i n f l u e n c e o n r a t i n g ... | https://arxiv.org/abs/2505.21082v2 |
.4 . 6 84 . 6 74 . 2 74 . 5 1Final R easoning. . . T h e r e v i e w e m p h a s i z e s t h e t a s t e o f t h e p r o d u c t , w h i c h i s r a t e d p o s i t i v e l y u n d e r t h e ' I n s i g h t ' f a c t o r , w h e r e t h e u s e r h a s a h i g h a v e r a g e s c o r e o f 4 . 5 1 . T h e r e v i e w e... | https://arxiv.org/abs/2505.21082v2 |
a g e s c o r e o f 4 . 4 6 a n d t h e o v e r w h e l m i n g l y p o s i t i v e n a t u r e o f t h e r e v i e w , a s c o r e o f 5 i s j u s t i f i e d .Figure 1: Comparison of response-level (Top) and reasoning-level (Bottom) personalization in a rating prediction task with scores from 1 to 5. Our approach gen... | https://arxiv.org/abs/2505.21082v2 |
modeling user-specific reasoning processes, demonstrating the effectiveness of reasoning-level personalization. In addition, human evaluation shows that structured compo- nents and personalized reasoning paths not only improve interpretability, but also increase users’ trust in the model’s outputs across diverse user s... | https://arxiv.org/abs/2505.21082v2 |
aligns model inference with user-specific reasoning logic by constructing and leveraging personalized reasoning paths from history. The framework consists of three key components: (1) personalized factor construction , which extracts and groups response-relevant features into statistical user-level factors (Section 3.2... | https://arxiv.org/abs/2505.21082v2 |
we prompt the LLM with the feature extraction instruction for the query qi, resulting in a set of features Gqi={fj}|Gqi| j=1defined as follows: Gqi=M(qi), f j= (name j,context j,factor j), (1) where fjdenotes a j-th feature extracted from qi, with name jrepresenting the semantic label of the feature, context jspecifyin... | https://arxiv.org/abs/2505.21082v2 |
which any feature from F(m)appears: Coverage (F(m)) =X (qi,ai)∈HuI[∃fj∈F(m)∩ Gqi]. (4) Among the covered instances, we compute the number of cases where at least one feature in F(m)is judged to have influenced the response, yielding the influence count: Influence (F(m)) =X (qi,ai)∈HuI[∃fj∈F(m)∩ Gqi:IsInfl fj→ai=True ].... | https://arxiv.org/abs/2505.21082v2 |
rq, a)provides structured logics on how similar user-specific information was processed in the past. Reasoning example-augmented generation. We then guide the black-box LLM Musing the retrieved examples. Standard few-shot prompting often provides relevant examples but lacks guidance on how the given information should ... | https://arxiv.org/abs/2505.21082v2 |
underlined values indicate the best and second-best scores for each metric, respectively. Dataset LaMP-2 LaMP-3 LaMP-5 GOQA Method +CoT Acc.↑ F1↑ MAE↓RMSE ↓R-1↑ R-L↑ Acc.↑ Zero-shot0.430 0.360 0.361 0.680 0.446 0.364 0.562 ✓ 0.411 0.337 0.323 0.630 0.434 0.376 0.557 ICL0.495 0.412 0.333 0.638 0.455 0.395 0.695 ✓ 0.471 ... | https://arxiv.org/abs/2505.21082v2 |
A simple CoT reasoning example (without using features and factors) is denoted by rCoT qi=M(preason , qi;ai). Bold and underlined values indicate the best and second-best scores for each metric, respectively. Dataset LaMP-2 LaMP-3 LaMP-5 GOQA Method Input Output Acc.↑F1↑ MAE↓RMSE ↓R-1↑R-L↑Acc.↑ Zero-shot q′a′0.430 0.36... | https://arxiv.org/abs/2505.21082v2 |
(Figure 3 Left). This suggests that the structured input 8 Table 3: Comparison of retrieval strategies using different sources and methods. Examples are retrieved from either user history ( Hu) or reasoning-augmented history ( Su). For example retrieval, the target query q′or its extracted features Gq′can serve as an i... | https://arxiv.org/abs/2505.21082v2 |
examine the impact of the number of examples on the performance of the personalization, we vary the number of retrieved user information which are provided to the model as a user-specific target query context. As shown in Figure 4, increasing the number of examples leads to consistent performance improvements, particul... | https://arxiv.org/abs/2505.21082v2 |
Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530 , 2024. [12] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Su... | https://arxiv.org/abs/2505.21082v2 |
Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yixin Dai, Jiawei Sun, Haofen Wang, and Haofen Wang. Retrieval-augmented generation for large language models: A survey. arXiv preprint arXiv:2312.10997 , 2:1, 2023. [28] Cheng Li, Mingyang Zhang, Qiaozhu Mei, Yaqing Wang, Spurthi Amba Hombaiah, Yi Liang, and Michael Bende... | https://arxiv.org/abs/2505.21082v2 |
, 2024. [41] Gautier Izacard, Mathilde Caron, Lucas Hosseini, Sebastian Riedel, Piotr Bojanowski, Armand Joulin, and Edouard Grave. Unsupervised dense information retrieval with contrastive learning. arXiv preprint arXiv:2112.09118 , 2021. [42] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Qu... | https://arxiv.org/abs/2505.21082v2 |
corresponding to 30% of the entire feature pool, restricted to features that have not yet been covered. This subset is used solely in thePropose stage to generate a diverse set of candidate factors while keeping the prompt length manageable. The subsequent Assign andSelect stages operate on the entire uncovered feature... | https://arxiv.org/abs/2505.21082v2 |
predicting a single user- assigned tag for a movie based on its description. Each user is associated with a history of previously tagged movies, which serves as their profile. Only the 15 most popular tags from the MovieLens dataset are used as labels. We subsample 50 users from the original time-based validation split... | https://arxiv.org/abs/2505.21082v2 |
(2.80GHz, 64 cores, 128 threads), using Python 3.10.13. The black-box language model that serves as the backbone across all experiments is gpt-4o-mini (gpt-4o-mini-2024-07-18 ), accessed via the OpenAI API using the LangChain framework3. All components of RPM—including feature extraction, factor construction, and reaso... | https://arxiv.org/abs/2505.21082v2 |
0.395 ±0.007 0.659 ±0.021 RPM (w/o Reasoning) 0.510±0.013 0.398 ±0.019 0.305 ±0.005 0.599 ±0.007 0.466 ±0.001 0.388 ±0.002 0.820 ±0.011 RPM 0.561±0.012 0.463±0.014 0.259±0.009 0.548±0.008 0.492±0.003 0.416±0.003 0.852±0.017 •Faithfulness : Whether the reasoning accurately reflects the input information. •Interpretabili... | https://arxiv.org/abs/2505.21082v2 |
statements) or explicit evaluative expres- sions associated with the feature. • w/o context : only the name and associated factor are retained, omitting the context -field. •context (ours): a clarifying phrase that grounds the feature in its surrounding query, providing a disambiguated interpretation of its intended me... | https://arxiv.org/abs/2505.21082v2 |
candidates achieving the maximum Jaccard score; if fewer than 3×Kcandidates remain, iteratively add the next-best scored groups until exactly 3×Kcandidates are collected (truncating any surplus). •Stage 2 (feature scoring). Apply the same sample-level cosine similarity as in our default method to this reduced pool and ... | https://arxiv.org/abs/2505.21082v2 |
0.594 0.744 0.242 0.534 0.913 0.257 0.568 0.837 RPM w/ transfer 0.306 0.614 0.771 0.235 0.539 0.886 0.258 0.563 0.831 lets each backbone reach, and occasionally exceed, the scores obtained when its own memory is freshly constructed. E Limitations Modeling User Decision-Making. Our framework constructs personalized reas... | https://arxiv.org/abs/2505.21082v2 |
useful knowledge of the surrounding environment, which makes our life more predictable. However, these applications often require people to contribute, consciously or unconsciously, location-related data for analysis, and this gravely encroaches users’ location privacy. Aggregate processing is a feasible way for preser... | https://arxiv.org/abs/2505.21082v2 |
the retrieval method being based on sketches. Combined line rendering technique : This feature indicates the technical approach taken in the retrieval process. Descriptor based on orientation of feature lines : This feature emphasizes the analytical aspect of the method, which is crucial for matching. Offline and onlin... | https://arxiv.org/abs/2505.21082v2 |
notes the unique aspect of the proposed scheme being ’participant-density-aware,’ which directly addresses the challenges posed by low participant density in crowd-sensing scenarios. This focus aligns with the researcher’s pattern of emphasizing methodologies that tackle specific challenges, as indicated by the high in... | https://arxiv.org/abs/2505.21082v2 |
Methodology) Personalized ReasoningTo generate a title for the scholarly paper, I begin by analyzing the abstract, which emphasizes the challenges of person re-identification due to variations in appearance. The researcher has a strong preference for methodologies, as indicated by the high influence rate (96.5%) in the... | https://arxiv.org/abs/2505.21082v2 |
paper, I begin by analyzing the abstract, which emphasizes the challenges of person re-identification due to variations in appearance. The researcher has a strong preference for methodologies, as indicated by the high influence rate (96.5%) in the Methodology category. This suggests that the researcher values innovativ... | https://arxiv.org/abs/2505.21082v2 |
meaningful categories that group similar features - Aim for general factors that apply across different contexts - Prioritize FUNDAMENTAL factors that cannot be further reduced - Ensure factors are ORTHOGONAL to each other (minimal conceptual overlap) - Create factors with high EXPLANATORY POWER across multiple domains... | https://arxiv.org/abs/2505.21082v2 |
input ] }} Title : {title} Features : {features} 31 Table 18: Personalized reasoning construction prompt for LaMP-5. Prompt: Personalized Reasoning Construction You are an expert in scholarly paper analysis. Your task is to generate a logical personalized reasoning path that explains how a researcher would arrive at a ... | https://arxiv.org/abs/2505.21082v2 |
Will It Still Be True Tomorrow? Multilingual Evergreen Question Classification to Improve Trustworthy QA Sergey Pletenev*,2,1, Maria Marina*, 2,1, Nikolay Ivanov1, Daria Galimzianova4, Nikita Krayko4,Mikhail Salnikov2,1,Vasily Konovalov2,5, Alexander Panchenko1,2, Viktor Moskvoretskii1,3 1Skoltech,2AIRI,3HSE University... | https://arxiv.org/abs/2505.21115v1 |
trained to identify evergreen questions. We demonstrate the usefulness of EG-E5 inarXiv:2505.21115v1 [cs.CL] 27 May 2025 EverGreenQA TimeQA MuLan FreshQA TAQA (our work) (Chen et al., 2021) (Fierro et al., 2024) (Vu et al., 2024) (Zhao et al., 2024) Both EG and mutable questions Train-Test split Human-Evaluated Multili... | https://arxiv.org/abs/2505.21115v1 |
address tempo- ral knowledge gaps externally, another direction is to update the internal knowledge of LLMs. Up- dating internal knowledge in LLMs is computa- tionally expensive, as retraining or editing mod- els often requires substantial resources and cannot be performed daily or hourly in practice. Tech- niques like... | https://arxiv.org/abs/2505.21115v1 |
ity, we recruited human evaluators for each target language, all of whom are either native speakers or possess advanced proficiency (B2–C1 level). We randomly sampled 100 questions from the test set (50 mutable, 50 evergreen) for evaluation. No er-rors were found in the translations for English, He- brew, German, or Ar... | https://arxiv.org/abs/2505.21115v1 |
wat position t, andVis the vocabulary. Results. Table 3 shows that most models ex- hibit only mild correlations between uncertainty and evergreen-ness, with Mistral 7B and Qwen 2.5 32B achieving the strongest signals.Model Perplexity Mean Token Entropy Gemma 2-9B-it 0.23 0.27 Gemma 2-27B-it 0.26 0.29 LLaMA 3.1-8B-it 0.... | https://arxiv.org/abs/2505.21115v1 |
0.54 0.46 0.62 0.43 0.68 0.70 0.21 0.78 0.67 0.11 0.87 Evergreen EG 0.50 0.72 0.52 0.52 0.20 0.79 0.47 0.65 0.62 0.49 0.31 0.65 0.51 0.28 0.68 0.50 0.10 0.87 Table 4: Self-knowledge identification performance. We report classification quality using AUROC and AUPRC, and calibration efficiency using PRR. EG stand for Eve... | https://arxiv.org/abs/2505.21115v1 |
city is the next winter olympics in Beijing Milan MuSiQue 2022 Who is the mayor presiding now where Merrill Elam was born? Lance Bottoms Andre Dickens SQuAD 2020 How many teams are in the Greek Super League? 18 14 HotpotQA 2018 Yau Ma Tei North is a district of a city with how many citizens? 7.2 million 7.4 million MuS... | https://arxiv.org/abs/2505.21115v1 |
need for manual annotation and facilitating the scalable construction of large QA corpora. 6.1 Popular QA Datasets Analysis Mutable questions pose a serious challenge for fair QA evaluation: outdated gold answers can make correct responses from modern LLMs ap- pear wrong, especially when models are evaluated at differe... | https://arxiv.org/abs/2505.21115v1 |
32B-it 0.36 Qwen 2.5 72B-it 0.35 EG-E5 0.66 EverGreen 0.77 Table 7: Correlation of ChatGPT with UC and EG. All results are significant (p-value < 0.05). EverGreen denotes ground true labels in the selected dataset part. 7 Explaining GPT-4o Retrieval GPT-4o autonomously decides when to invoke its retrieval system using ... | https://arxiv.org/abs/2505.21115v1 |
distinguishes its knowl- edge. For instance, learning to differentiate be- tween truly stable physical facts (such as the area of Liechtenstein) and more variable ones (like the brightest star in the sky), or between completedhistorical events (e.g., the French Revolution) and ongoing developments (such as upcoming pre... | https://arxiv.org/abs/2505.21115v1 |
was collected, stored, or used. All examples are factual in nature and were manuallyreviewed to ensure compliance with privacy and ethical standards. Dataset labels and translations were created by trained linguists and multilingual annotators. Anno- tators were compensated fairly according to local labor regulations. ... | https://arxiv.org/abs/2505.21115v1 |
the North American Chapter of the Association for Computational Linguistics: Human Language Technologies: Short Papers, NAACL 2024, Mexico City, Mexico, June 16-21, 2024 , pages 762–771. Association for Computational Linguistics. Marina Fomicheva, Shuo Sun, Lisa Yankovskaya, Frédéric Blain, Francisco Guzmán, Mark Fishe... | https://arxiv.org/abs/2505.21115v1 |
2024. Realtime qa: What’s the answer right now? Preprint , arXiv:2207.13332. Tom Kwiatkowski, Jennimaria Palomaki, Olivia Red- field, Michael Collins, Ankur P. Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Jacob Devlin, Ken- ton Lee, Kristina Toutanova, Llion Jones, Matthew Kelcey, Ming-Wei Chang, Andrew M... | https://arxiv.org/abs/2505.21115v1 |
Petrakov, Rui Xing, Abdelrahman Sadallah, Kirill Grishchenkov, Alexander Panchenko, Timothy Baldwin, Preslav Nakov, Maxim Panov, and Artem Shelmanov. 2024. Benchmarking uncertainty quan- tification methods for large language models with lm-polygraph. Transactions of the Association for Computational Linguistics , 13:22... | https://arxiv.org/abs/2505.21115v1 |
pages 15015– 15040. Association for Computational Linguistics. A Evergreen Testing Details LLM Verbal Parameters. Each example comes with 5-shot for mutable and 5-shot for immutable examples. For llama 3.1 sampling parameters are following: TEMPERATURE =0.7, TOP_P=0.9 . For Qwen 2.5: TEMPERATURE =0.6, TOP_P=0.95, TOP_K... | https://arxiv.org/abs/2505.21115v1 |
[3, 5, 7, 9, 11], max_features: [0.2, 0.4, sqrt, log2, None], bootstrap: [True, False], criterion: [gini, entropy], class_weight: [balanced, 0: 1, 1: 1, None] E Predictive Analysis of Uncertainty for Temporality Table 9 reports McFadden’s pseudo- R2values from logistic regression models trained to predict evergreen-nes... | https://arxiv.org/abs/2505.21115v1 |
and complete. Example of an incomplete or partially useful answer that is not suitable as a golden answer: Question: Do spiders have teeth? Answer: Yes, spiders have teeth. Comment: A fact-check in open sources reveals that this answer is not accu- rate enough to be considered a golden answer. The correct response woul... | https://arxiv.org/abs/2505.21115v1 |
enhanced the variety and coverage of our training set. Synthetic Instruction Can you generate different question-answer pair: slow-changing questions, in which the answer typically changes over the course of several years (up to 10); fast-changing ques- tion, in which the answer typically changes within a year or less;... | https://arxiv.org/abs/2505.21115v1 |
0.842 0.845 0.841 0.832 0.825 0.831 0.836 E5 Small (Wang et al., 2024) 0.821 0.822 0.819 0.815 0.804 0.807 0.817 0.815 E5 Large (Wang et al., 2024) 0.910 0.913 0.909 0.910 0.904 0.900 0.897 0.906 Table 10: Comparison of different models on a training dataset. All models are multilingual variants. The best scores are sh... | https://arxiv.org/abs/2505.21115v1 |
arXiv:2505.21137v1 [cs.CL] 27 May 2025Scaling and Prompting for Improved End-to-End Spoken Grammatical Error Correction Mengjie Qian, Rao Ma, Stefano Bann `o, Kate M. Knill, Mark J.F . Gales 1ALTA Institute, Department of Engineering, University of Cambridge, UK {mq227,rm2114,sb2549,kmk1001,mjfg100 }@cam.ac.uk Abstract... | https://arxiv.org/abs/2505.21137v1 |
rather than simply presenting a ‘ready-made’ correction. Feedback is a crucial component in CALL applications, offering learners actionable insights into where and how they have made errors. Effective feedback must be easy to understand, informative, and support- ive of language learning. Therefore, in contrast to SGEC... | https://arxiv.org/abs/2505.21137v1 |
speech trans- lation across various language pairs [20, 21], and other spoken language understanding tasks beyond ASR [22, 23]. In this work, we extend Whisper for E2E spoken grammati- cal error correction by fine-tuning it on grammatically corrected transcriptions (Whisper gec). The model directly generates gram- mati... | https://arxiv.org/abs/2505.21137v1 |
focus on language structure and im- proving its ability to generate accurate GEC transcriptions and provide useful feedback. 3. Experimental Setup 3.1. Datasets This paper uses Linguaskill [24] labelled and unlabelled train- ing sets to build systems, Linguaskill dev set to select hyper- parameters, and Linguaskill tes... | https://arxiv.org/abs/2505.21137v1 |
studies [10, 7] have demonstrated that both Translation Edit Rate (TER) and Word Error Rate (WER) are relevant metrics for spoken GEC. Both metrics report similar trends in evaluating spoken GEC, making it unnecessary to use both. In this work, we adopt WER as the primary metric for its simplicity and clarity. To asses... | https://arxiv.org/abs/2505.21137v1 |
and large-v2 models. Model FT (cont.)small.en large-v2 LNG lbl S&I LNG lbl S&I Whisper flt+ GEC 13.24 16.91 11.81 13.99 Whisper gecLNG lbl 13.48 17.76 11.10 13.21 LNG unl 14.16 18.11 12.93 15.92 + LNG lbl12.72 16.84 11.10 13.93 4.2. Prompting with Additional Information In this experiment, we investigate whether prompt... | https://arxiv.org/abs/2505.21137v1 |
performance using model prompting, larger model size and pseudo-labelled data, we assess their impact on feedback per- formance. Here, we focus on the large-v2 models as they con- sistently outperform the small.en model. Table 4 presents the SGEC feedback performance for various GEC models based on large-v2, evaluated ... | https://arxiv.org/abs/2505.21137v1 |
2023. [Online]. Available: https://doi.org/10.1162/coli a00478 [2] H. T. Ng, S. M. Wu, T. Briscoe, C. Hadiwinoto, R. H. Susanto, and C. Bryant, “The CoNLL-2014 shared task on grammatical er- ror correction,” in Proceedings of the 18th conference on compu- tational natural language learning: shared task , 2014, pp. 1–14... | https://arxiv.org/abs/2505.21137v1 |
inFindings of the Association for Computational Linguistics: NAACL 2024 . Mexico City, Mexico: Association for Computa- tional Linguistics, Jun. 2024, pp. 754–781. [Online]. Available: https://aclanthology.org/2024.findings-naacl.49 [16] M. Qian, K. Knill, S. Banno, S. Tang, P. Karanasou, M. J. Gales, and D. Nicholls, ... | https://arxiv.org/abs/2505.21137v1 |
T. Wolf et al. , “Transformers: State-of-the-art natural language processing,” in Proc. Conf. on Empirical Methods in Natural Lan- guage Processing: System Demonstrations , 2020, pp. 38–45. [31] D. Dahlmeier and H. T. Ng, “Better evaluation for grammatical error correction,” in Proceedings of the 2012 Conference of the... | https://arxiv.org/abs/2505.21137v1 |
arXiv:2505.21138v1 [cs.CL] 27 May 2025Leveraging LLM and Self-Supervised Training Models for Speech Recognition in Chinese Dialects: A Comparative Analysis Tianyi Xu1,2, Hongjie Chen1, Wang Qing1, Lv Hang1, Jian Kang1, Li Jie1, Zhennan Lin2, Yongxiang Li1, Xie Lei2 1Institute of Artificial Intelligence (TeleAI), China ... | https://arxiv.org/abs/2505.21138v1 |
subsequently fed into an LLM for decoding. This method aims to integrate acous- tic features more effectively with linguistic context to enhance ASR accuracy. Specifically, SALMONN [20] employs Whis- per [1] to extract semantic content while utilizing BEATs [23] to enable comprehensive perception of speech, music, and ... | https://arxiv.org/abs/2505.21138v1 |
are first extracted using the speech encoder, producing the encoder output Hs, represented as: Hs=Encoder (S). (3) The encoder output Hsis then passed through a projector, fol- lowed by a linear layer, to generate a feature sequence Eswith the same dimensionality as the input to the LLM, denoted as: Es=Linear (Projecto... | https://arxiv.org/abs/2505.21138v1 |
we trained Data2vec2 on a larger dataset consisting of 300,000 hours of unlabeled multilingual and accented speech. For the current study, we use the TeleSpeechPT variant of Data2vec2 with a model size of 24 layers and 301 million parameters. The model was trained at a sampling rate of 25Hz, which is lower than the com... | https://arxiv.org/abs/2505.21138v1 |
set to 5. This ensures that gradients over 5 or below -5 are clipped to 5 or -5, respectively. Additionally, we use a gra- dient accumulation of 20 to increase the effective batch size to 100. Regarding the training approach, when training the LLM, we freeze the LLM body and only update the LLM using LoRA fine-tuning [... | https://arxiv.org/abs/2505.21138v1 |
6.25Hz 18.95 18.30 10.98 23.31 6.48 7.54 6.77 Table 4: The improvement in CER (%) of the model across mul- tiple dialectal datasets in 4-stage fine-tuning. Stage Projector He Nan Shang Hai Hu Nan Cantonese kespeech Test net Test meeting 1Linear 58.59 84.44 115.68 25.32 25.74 17.68 20.82 Conv1d 61.13 58.16 169.03 21.35 ... | https://arxiv.org/abs/2505.21138v1 |
the 4-stage fine-tuning on the 40,000- hour dataset. The results are summarized in Table 3. For Qwen2 0.5B, we compare LoRA fine-tuning versus full fine-tuning in LLM. We observe a significant performance increase, which suggests that the LoRA parameter is not sufficient for acous- tic representation training. In parti... | https://arxiv.org/abs/2505.21138v1 |
Niu, “WeNet 2.0: More Productive End- to-End Speech Recognition Toolkit,” in Interspeech , 2022. [6] Z. Yao, L. Guo, X. Yang, W. Kang, F. Kuang, Y . Yang, Z. Jin, L. Lin, and D. Povey, “Zipformer: A faster and better encoder for automatic speech recognition,” in The Twelfth International Conference on Learning Represen... | https://arxiv.org/abs/2505.21138v1 |
in Thirty-fifth Conference on Neural Information Processing Systems Datasets and Bench- marks Track (Round 2) , 2021. [20] C. Tang, W. Yu, G. Sun, X. Chen, T. Tan, W. Li, L. Lu, Z. Ma, and C. Zhang, “SALMONN: Towards Generic Hearing Abilities for Large Language Models,” CoRR , 2023. [21] Z. Ma, G. Yang, Y . Yang, Z. Ga... | https://arxiv.org/abs/2505.21138v1 |
arXiv:2505.21148v1 [cs.CL] 27 May 2025Assessment of L2 Oral Proficiency using Speech Large Language Models Rao Ma, Mengjie Qian, Siyuan Tang, Stefano Bann `o, Kate M. Knill, Mark J.F . Gales 1ALTA Institute, Department of Engineering, University of Cambridge, UK {rm2114,mq227,st941,sb2549,kmk1001,mjfg100 }@cam.ac.uk Ab... | https://arxiv.org/abs/2505.21148v1 |
their ability to assess pronunciation and fluency aspects directly To address these lim- itations, [17, 18] proposed building an E2E grader from the wav2vec 2.0 model. The speech foundations models are ca- pable of leveraging rich acoustic features, however, they do not explicitly account for aspects related to content... | https://arxiv.org/abs/2505.21148v1 |
Half-point scores (e.g., 3.5) are allowed for intermediate ratings. Our goal is to develop auto graders that produce scores closely aligned with the references. As LLMs are trained with the next-word prediction target, they are suitable for tasks with text-based outputs. Thus, applying LLMs to holistic scoring, which r... | https://arxiv.org/abs/2505.21148v1 |
score (arg max cyi,c) (4) This approach is referred to as the “hard” decoding strategy. Nevertheless, it only takes into account the highest prediction and ignores the model predictions over other class labels. Al- ternatively, we propose a “soft” decoding strategy where we aggregate the model outputs using the fair av... | https://arxiv.org/abs/2505.21148v1 |
the following, we refer to this model as Qwen2Audio . In the ex- periments, we evaluate the zero-shot classification performance of Qwen2Audio on L2 oral assessment. When adapted, LoRA adaptors with a rank of 16 are inserted into both encoder and de- coder layers, adding 10M parameters to the pre-trained model. In all ... | https://arxiv.org/abs/2505.21148v1 |
of mod- els adapted with LoRA tuning. By only introducing 10M new parameters, the model performance on all metrics largely im- prove. After the training, Qwen2Audio learns to more faithfully grade the learner’s performance, with the PCC value increas- ing from 0.371 to 0.892 on LinGen. For models trained with different... | https://arxiv.org/abs/2505.21148v1 |
perform well on tasks that they are not specifically trained on. In this section, we aim to assess the gen- eralisation capability of the grader on out-of-domain test sets. For Linguaskill, although all five parts of the test aim to mea- sure the candidate’s spoken language proficiency, they vary in format and focus. F... | https://arxiv.org/abs/2505.21148v1 |
the proposed L2 graders. Table 5: Within and cross task evaluation on test parts 1,3,4,5 using models trained on Linguaskill and S&I. Train ModelTest (PCC) LinGen LinBus S&I LNGBERT 0.941 0.931 0.796 Qwen2Audio 0.951 0.938 0.824 S&IBERT 0.901 0.901 0.753 Qwen2Audio 0.929 0.914 0.833 4. Conclusions In this paper, we exa... | https://arxiv.org/abs/2505.21148v1 |
1 (Long and Short Papers) , 2019, pp. 4171–4186. [9] V . Raina, M. J. F. Gales, and K. M. Knill, “Universal adversarial attacks on spoken language assessment systems,” in Proc. Inter- speech , 2020, pp. 3855–3859. [10] X. Wang, K. Evanini, Y . Qian, and M. Mulholland, “Automated scoring of spontaneous speech from young... | https://arxiv.org/abs/2505.21148v1 |
no. 2, p. 100050, 2023. [24] K. P. Yancey, G. Laflair, A. Verardi, and J. Burstein, “Rating short L2 essays on the CEFR scale with GPT-4,” in Proceedings of the 18th Workshop on Innovative Use of NLP for Building Educa- tional Applications (BEA 2023) . Toronto, Canada: Association for Computational Linguistics, Jul. 20... | https://arxiv.org/abs/2505.21148v1 |
Leveraging GANs for citation intent classification and its impact on citation network analysis Davi A. Bezerra1, Filipi N. Silva2and Diego R. Amancio1 1Institute of Mathematics and Computer Science, University of S˜ ao Paulo, S˜ ao Carlos, Brazil 2Observatory on Social Media, Indiana University, Bloomington, IN, USA Ab... | https://arxiv.org/abs/2505.21162v1 |
and Result. Before the advent of machine learning models, citation intention classification was al- ready done manually, focusing on just one article to understand how citations occurred [45, 50]. With the advancement of automatic techniques for intention classification, anno- tated databases, and computational power, ... | https://arxiv.org/abs/2505.21162v1 |
[1]. Ensemble methods and nearest-neighbor algorithms also featured prominently in early computational approaches [17, 47]. A major step forward came from the work of Jurgens et al. [26], who proposed a robust annotation schema classifying citations into six distinct categories: background, motivation, uses, extension,... | https://arxiv.org/abs/2505.21162v1 |
AND METHODS This section provides a detailed overview of the methodology developed and refined throughout this study. The main objective is to investigate citation intent classification and its significance in enhancing the understanding of each citation’s relevance within a citation network. To achieve this, we adopt ... | https://arxiv.org/abs/2505.21162v1 |
six nuanced classes, whereas SciCite simplifies classification into three broader categories, emphasizing general applicability. The average context lengths are comparable (33-36 words), but SciCite stands out with a maximum length of 510 words, offering richer contextual information compared to ACL-ARC (178 words) and... | https://arxiv.org/abs/2505.21162v1 |
problem, concept, approach, topic, or the significance of the problem in the field.Recent evidence suggests that co- occurring alexithymia may explain deficits [12]. Locally high-temperature melting regions can act as permanent termination sites [6–9]. One line of work is focused on changing the objective function (Mao... | https://arxiv.org/abs/2505.21162v1 |
integrating the contex- tual representation capabilities of BERT with the adversarial training framework of GAN, GAN-BERT enhances performance in sentence classification tasks. This hybrid model is particularly advantageous in scenarios where labeled data is limited, as it effectively utilizes both supervised and unsup... | https://arxiv.org/abs/2505.21162v1 |
it identifies synthetic examples produced by the generator as belonging to a separate, distinct class ( k+ 1). Through this dual functionality, the discriminator leverages both labeled and unlabeled data structures effectively, enhancing 12 noise G Synthetic SciBertlabelc Labeled UnlabeledDk classes k + 1 fake/real XLN... | https://arxiv.org/abs/2505.21162v1 |
that only relevant and high-impact citations are included. The network analysis consists of three main steps: 1.Intent-Based Filtering : citations are categorized using a citation intent classifier trained with the SciCite dataset, employing GAN-BERT enhanced by SciBERT em- beddings. The SciCite dataset was selected du... | https://arxiv.org/abs/2505.21162v1 |
semi-supervised SS-cGAN + SciBERT model achieved an F 1score of 81.75%, significantly outperforming baseline methods such as CitePrompt (68.39%) and SciBERT Finetune (70.98%). To further illustrate its strong generalization ca- pability, Figure 3 shows the confusion matrix of the SS-cGAN + SciBERT predictions, where th... | https://arxiv.org/abs/2505.21162v1 |
visualizations in Figures 4 and 5 depict the SciBERT model embeddings of the SciCite dataset, showing the predicted and target values for the test samples, respectively. Each point in the image represents a citation context, with each color corresponding to a citation intent. In Figure 4, the classes appear relatively ... | https://arxiv.org/abs/2505.21162v1 |
asresult . We constructed a citation network comprising 76,640 nodes and 171,403 edges. Our anal- ysis focused on the largest component, extracted from the weakly connected subgraph, which consists of 71,939 nodes and 165,615 edges. Figure 7 displays the resulting graph, where a 20 background method result Predicted la... | https://arxiv.org/abs/2505.21162v1 |
betweenness – before and after filtering. These changes were visualized using bump charts, highlighting the top 20 papers identified by each cen- trality metric. Figure 8 illustrates that, after removing background citations, the in-degree centrality analysis exhibits significant shifts in the rankings of the most cite... | https://arxiv.org/abs/2505.21162v1 |
and after background citation filtering. Paper 1412.6980 rises from rank 2 to 1, indicating increased centrality independent of background citations, while paper 1411.4038 experiences a sharp decline from rank 6 to 24, suggesting strong dependence on background-oriented connections. V. CONCLUSION Citations have traditi... | https://arxiv.org/abs/2505.21162v1 |
proportion of method or background citations may indicate continuity rather than disrup- tion. Also, future work could investigate whether certain linguistic cues in citation contexts signal that a paper is perceived as disruptive when cited. Additionally, intent-based analysis may support bias and ethics monitoring by... | https://arxiv.org/abs/2505.21162v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.