text string | source string |
|---|---|
Others WikiTables EI 0.05 4.33 0.00 0.17 0.00 0.00 0.13 0.03 1.170.00 0.02 0.00 0.00 0.13 0.02 2.33 0.00 0.11 0.00 0.00 0.07 MI 0.01 0.80 0.00 0.03 0.00 0.00 0.00 0.01 0.93 0.00 0.00 0.00 0.00 0.01 0.00 0.73 0.00 0.00 0.00 0.00 0.00 Partial 0.22 25.00 0.00 0.35 0.00 0.01 0.09 0.3220.34 0.00 0.55 0.00 0.02 0.10 0.30 22.... | https://arxiv.org/abs/2505.22176v1 |
perfor- mance, it is not without its limitations. Its reliance on large-scale language models comes at the cost of increased computational overhead, which may impactscalability. Moreover,themethodfaceschal- lengeswhendealingwithhierarchicaltables,where nested headers and multi-level groupings make alignment significant... | https://arxiv.org/abs/2505.22176v1 |
method for automatic evalu- ation of machine translation. In Proceedings of the 40thAnnualMeetingoftheAssociationforCompu- tational Linguistics , pages 311–318, Philadelphia, Pennsylvania,USA.AssociationforComputational Linguistics. Panupong Pasupat and Percy Liang. 2015. Composi- tional semantic parsing on semi-struct... | https://arxiv.org/abs/2505.22176v1 |
0.9and compute a normalized difference: GT−Ref Ref = 0.4. Thus, γp= 0.9×0.4 = 0 .36. Ourproposedrubricisdomainagnostic,thebeautyofourworkisthatourweightingschemeisflexible which makes the evaluation metric domain specific based on domain knowledge, for instance one can set higher weights to cell values with numeric typ... | https://arxiv.org/abs/2505.22176v1 |
| Year.T1/Y ear.T2 | Position.T1/Position.T2 | ... | 201 1/201 1 | 2nd/2nd | ... | ... | ... | ... Output (shortened): | Year.T1/Y ear.T2 | Competition.T1/- | V enue.T1/Place.T2 | ... | | ... | ... | ... | ... | The final alignment shows all matched and unmatched columns/rows, with slashes and dashes where appropriate.... | https://arxiv.org/abs/2505.22176v1 |
differences (MI, EI, Partial) ... Format: | Category | Numerical | String | Bool | Date | List | Time | Others | |----------|-----------|--------|------|------|------|------|--------| | MI | ... | ... | ... | ... | ... | ... | ... | | EI | ... | ... | ... | ... | ... | ... | ... | | Partial | ... | ... | ... | ... | ..... | https://arxiv.org/abs/2505.22176v1 |
arXiv:2505.22222v1 [cs.CV] 28 May 2025Look & Mark: Leveraging Radiologist Eye Fixations and Bounding boxes in Multimodal Large Language Models for Chest X-ray Report Generation Yunsoo Kim1Jinge Wu1Su-Hwan Kim2 Pardeep Vasudev1Jiashu Shen3Honghan Wu1,4 1UCL2Technical University of Munich3University of Oxford4University ... | https://arxiv.org/abs/2505.22222v1 |
diagnostic process by tracking where doctors look and how long they spend examining different areas of a chest X-ray. Bounding box annotations help localize the language output in well-defined spatial coordi- nates, reducing the risk of free-form hallucinations and improving multimodal large language model’s “grounded”... | https://arxiv.org/abs/2505.22222v1 |
MAIRA2 focuses solely on bounding box ground- ing and does not incorporate the dynamic reasoning patterns captured through radiologists’ eye fixa- tions. 2.2 Chest X-ray Diagnosis with Eye Fixation Kim et al.(Kim et al., 2024b) explored the role of radiologists’ fixation data in guiding multi- modal LLMs for CXR analys... | https://arxiv.org/abs/2505.22222v1 |
xi2, yi2, li), (1)where (xi1, yi1)and (xi2, yi2)are the top-left and bottom-right coordinates, respectively, and liis the associated abnormality label (e.g., “Car- diomegaly”). For each bounding box, an abnormal- ity caption is assigned, as shown in Figure 1. 3.2 Eye Fixation Integration Fixation data G={g1, g2, . . . ... | https://arxiv.org/abs/2505.22222v1 |
report. To address these limitations, we chose to use the dictated reports from the REFLACX dataset rather than the standard MIMIC_CXR findings and impressions. REFLACX provides multiple dic- tated reports per image, which better capture the variability in radiologist interpretation and report terminology. This approac... | https://arxiv.org/abs/2505.22222v1 |
contrast, LLaV A-Med was trained using an instruction tuning dataset derived from figures and legends in PubMed papers. This model, with 8 billion parameters, was designed for medical visual question answering (VQA) tasks, demonstrating strong reasoning capabilities but lacking specific training on MIMIC-CXR data. To p... | https://arxiv.org/abs/2505.22222v1 |
widely used in past work, several recent studies have shown that it correlates poorly with expert radi- ologist evaluations, particularly for complex, multi- sentence report generation tasks. Instead, we adopt more recent and clinically grounded evaluation metrics such as RadGraph-XL (Delbrouck et al., 2024) and RaTESc... | https://arxiv.org/abs/2505.22222v1 |
of clinical writing samples and grounding cues of L&M. ForQwen2.5VL , I&L&M yields significant gains, especially in RadGraph-XL (0.0812 vs. 0.0534) and C.A VG (74.83% vs. 61.27%). Sim- ilarly, Llama3.2V sees marked improvements in Model Method RG-L BERT RadG RaTE C.A VG (%) A.A VG (%) CXR-LLaV A - 0.1653 0.8586 0.1107 ... | https://arxiv.org/abs/2505.22222v1 |
prompt. Still, L&M performed bet- ter than L in all models, although very small (0.8%) for MAIRA2. This shows that grounding can be more effective than fixation for report generation. 5.2.2 General-Purpose Models with In-Context Learning (Figure 3) For general-purpose models, in-context learning combined with Look & Ma... | https://arxiv.org/abs/2505.22222v1 |
CXR-LLaV A (-) . Figure 4 provides qualitative analysis of model outputs, comparing generated reports from differ- ent methods against ground truth reports. Three ex- amples are shown, with clinically significant errors marked in red, as identified by radiologists. The examples include cases of pneumothorax, pleural ef... | https://arxiv.org/abs/2505.22222v1 |
lexical and clinical evaluation metrics, with the largest gains observed in clinical metrics such as RaTEScore and RadGraph-XL. For instance, CXR-LLaV A achieved a 1.2% improvement in overall metrics (A.A VG) compared to baseline prompting, while LLaV A-Med demonstrated a remarkable 9.2% boost. General-purpose models a... | https://arxiv.org/abs/2505.22222v1 |
clude the propagation of biases present in training datasets and the possibility of over-reliance on AI- generated reports, particularly in high-stakes clini- cal environments. To mitigate these risks, L&M is explicitly designed as an assistive tool to support, rather than replace, radiologist decision-making. Addition... | https://arxiv.org/abs/2505.22222v1 |
Wu. 2024a. Enhancing human-computer in- teraction in chest x-ray analysis using vision and lan- guage model with eye gaze patterns. arXiv preprint arXiv:2404.02370 . Yunsoo Kim, Jinge Wu, Yusuf Abdulle, Yue Gao, and Honghan Wu. 2024b. Human-in-the-loop chest x- ray diagnosis: Enhancing large multimodal models with eye ... | https://arxiv.org/abs/2505.22222v1 |
in medical visual question an- swering. In The Second Tiny Papers Track at ICLR 2024 . Hanguang Xiao, Feizhong Zhou, Xingyue Liu, Tianqi Liu, Zhipeng Li, Xin Liu, and Xiaoxuan Huang. 2024. A comprehensive survey of large language models and multimodal large language models in medicine. arXiv preprint arXiv:2405.08603 .... | https://arxiv.org/abs/2505.22222v1 |
Look & Mark en- hances the adaptability of general-purpose models for clinical tasks, particularly when evaluated withclinical relevance metrics. This heatmap provides a clear comparative anal- ysis of model performance under different prompt- ing strategies, emphasizing the contributions of Look & Mark in reducing hal... | https://arxiv.org/abs/2505.22222v1 |
arXiv:2505.22231v1 [cs.SD] 28 May 2025Advancing Hearing Assessment: An ASR-Based Frequency-Specific Speech Test for Diagnosing Presbycusis Stefan Bleeck Institute of Sound and Vibration Research, University of Southampton May 29, 2025 Abstract Traditional audiometry often fails to fully characterize the functional impa... | https://arxiv.org/abs/2505.22231v1 |
This capability allows for a more fine-grained, frequency-specific diagnostic insight into speech percep- tion deficits. This paper details the development of a novel ASR-based frequency-specific speech test designed to provide such insights, particularly for conditions like presbycu- sis. Our approach leverages ASR to... | https://arxiv.org/abs/2505.22231v1 |
custom Levenshtein distance algorithm for phoneme sequences, quantifying substitutions, insertions, and deletions. A phoneme- level confusion matrix was generated, counting the frequency of original-to-transcribed phoneme mappings. This analysis involved a backtracing mechanism to identify the spe- cific sequence of op... | https://arxiv.org/abs/2505.22231v1 |
refine the speech test for optimal diagnostic capability, an ASR-based diagnostic value assessment was performed. This involved comparing ASR performance under two distinct conditions: Normal Hearing (NH) Simulation (ASR processing speech stimuli with noise but with- out simulated hearing loss) and Hearing Loss (HL) Si... | https://arxiv.org/abs/2505.22231v1 |
200 W→B 186 T→D 172 AE1→EH1 163 N→L 155 IY1→IH1 147 T→K 142 R→B 140 IH1→IY1 136 M→L 134 S→K 134 N→T 131 R→T 129 OW1→AA1 123 R→K 122 AA1→AH1 121 AO1→AA1 114 Figure 2 presents the top N (e.g., N=20) most frequently selected specific phoneme con- fusion types from the curated test item set. This highlights the individual ... | https://arxiv.org/abs/2505.22231v1 |
place of articulation confusions within the cu- rated test item set for substitution errors, confirming that curated items predominantly reflect confusions among phonemes with high-frequency acoustic cues. 9 Figure 3: Place of Articulation Confusion Matrix (Curated Items). This heatmap de- picts the distribution of pla... | https://arxiv.org/abs/2505.22231v1 |
girls girl 88 4 84 challenged challenge 80 4 76 repainting recanting 98 34 64 around ’round 96 36 60 musical musica 84 34 50 boys boyce 98 54 44 even given 100 60 40 effects effect 58 18 40 few feel 80 42 38 lost ast 98 62 36 shall chalk 82 46 36 wash wat 92 58 34 even aven 98 64 34 keep kip 94 62 32 Continued on next ... | https://arxiv.org/abs/2505.22231v1 |
the overall discriminative power of the test. 16 Figure 9: Receiver Operating Characteristic (ROC) Curve. This figure presents the ROC curve, illustrating the trade-off between Sensitivity and (1 - Specificity) across various potential test failure thresholds. The dashed black line represents the performance of a rando... | https://arxiv.org/abs/2505.22231v1 |
audible cues [5]. Regard- ing the third research question, the simulation demonstrated that a test battery curated from these ASR-derived confusions can effectively differentiate between normal-hearing and hearing-impaired listeners in a simulated environment. The diagnostic difference cal- culated for curated word pai... | https://arxiv.org/abs/2505.22231v1 |
with the *audibility*-centric view of mild-to-moderate loss [1,2], future work should consider incorporating models of supra-threshold deficits [3,4] for broader applicability, especially in noise. The TIMIT corpus, while phonetically rich, consists of read speech and may not fully represent the variability of natural ... | https://arxiv.org/abs/2505.22231v1 |
A Linguistically Motivated Analysis of Intonational Phrasing in Text-to-Speech Systems: Revealing Gaps in Syntactic Sensitivity Charlotte Pouw1, Afra Alishahi2, Willem Zuidema1 1Institute for Logic, Language and Computation, University of Amsterdam 2Cognitive Science and Artificial Intelligence, Tilburg University {c.m... | https://arxiv.org/abs/2505.22236v1 |
when it reliably signals the need for an intonational boundary (i.e., obvious clause bound- aries in simple sentence structures), although the duration of intonational boundaries is also mod- ulated by lexical cues. In more complex cases such as garden path sentences and attachment am- biguities, systems need explicit ... | https://arxiv.org/abs/2505.22236v1 |
flower . The attachment site of the PP with the flower was ambiguous, as the room contained a frog toy with a flower on its head, as well as a frog and a flower separately. When speakers were aware of the ambiguity, they produced a bound- ary after frogto signal the VP-attachment structure (i.e., when they wanted the a... | https://arxiv.org/abs/2505.22236v1 |
the structure of sentences with (temporary) syntactic ambiguity, and place intonational boundaries in the correct positions accordingly. Using controlled stimuli, we analyze which cues are used by the systems to disambiguate these sentences. 4.1 Syntactic Disambiguation Garden path sentences contain temporary syntactic... | https://arxiv.org/abs/2505.22236v1 |
duration of each Adv. NP1 VP1 <A> NP2 <B> NP3 VP20.00.51.0Duration (s) 1* 2*1* 2*1* 2*1* 2* When Roger lef t ,the house was darkWhen Roger lef t the house , it was darkComma Early Closure Late Closure Adv. NP1 VP1 <A> NP2 <B> NP3 VP2 1 21 2 When Roger lef t the house was darkWhen Roger lef t the house it was darkNo Com... | https://arxiv.org/abs/2505.22236v1 |
From Simple Wikipedia2, we select sentences that contain exactly one comma, marking a syntactic boundary.3We select boundaries that signal major structural breaks, which typically lead to an audible intonational boundary in spoken language. We use dependency parsing to detect such structural breaks (examples are listed... | https://arxiv.org/abs/2505.22236v1 |
similar pattern: the strongest intonational boundaries are produced in the Syntactic + Comma cue condition. None of the models produce an intonational boundary in the No cue condition. The Syntactic cue and Un- natural comma cue conditions are inbetween, with the comma cue leading to a slightly stronger in- tonational ... | https://arxiv.org/abs/2505.22236v1 |
20 percent. We use R2(explained variance) as our evaluation metric to gauge how well the predicted regression lines fit the data.Category Predictor Punctuation Comma Presence (1 or 0) Lexical Preceding POS tag (one-hot) Following POS tag (one-hot) Constituency Is Clause Boundary (1 or 0) Num. Closing Brackets Max. Tree... | https://arxiv.org/abs/2505.22236v1 |
obvious clause boundaries, garden path sentences are likely under- represented in their training data. Sentences with attachment ambiguity may occur more frequently. However, even for such sentences, the intonation patterns we aim to capture (where high attachment introduces an intonational boundary and low attach- men... | https://arxiv.org/abs/2505.22236v1 |
semantic bias towards high attachment, and 2500 sentences with a bias towards low attachment (re- sulting in ~6 hours of speech). We synthesized these sentences using Tacotron2, inserting com- mas at positions that would correspond to intended pauses (e.g., before the preposition with in high attachment cases). We agai... | https://arxiv.org/abs/2505.22236v1 |
of phenomena to better understand the types of lin- guistic associations captured by TTS systems. One potential direction would be to develop a resource similar to BLiMP (Warstadt et al., 2020) for TTS, which could serve as a more comprehensive frame- work for evaluating their syntactic sensitivity. Ad- ditionally, str... | https://arxiv.org/abs/2505.22236v1 |
for Computational Lin- guistics, ACL 2024 , pages 14727–14742. Association for Computational Linguistics (ACL). Ambika Kirkland, Shivam Mehta, Harm Lameris, Gus- tav Eje Henter, Eva Székely, and Joakim Gustafson. 2023. Stuck in the MOS pit: A critical analysis of MOS test methodology in TTS evaluation. In 12th Speech S... | https://arxiv.org/abs/2505.22236v1 |
Shen, Afra Alishahi, Arianna Bisazza, and Grze- gorz Chrupała. 2023. Wave to Syntax: Probing spo- ken language models for syntax. In Proc. INTER- SPEECH 2023 , pages 1259–1263. Jonathan Shen, Ruoming Pang, Ron J Weiss, Mike Schuster, Navdeep Jaitly, Zongheng Yang, Zhifeng Chen, Yu Zhang, Yuxuan Wang, Rj Skerrv-Ryan, et... | https://arxiv.org/abs/2505.22236v1 |
open question how varying these voice characteristics might influence the resulting intonation patterns. 0.00.51.0Duration (s) 1* 2*1* 2*1* 2*1* 2*Comma (Tacotron2) Early Closure Late Closure 1 21 2No Comma (Tacotron2) Adv. NP1 VP1 <A> NP2 <B> NP3 VP20.00.51.0Duration (s) 1* 2*1* 2*1* 2*1* 2* When Roger lef t ,the hous... | https://arxiv.org/abs/2505.22236v1 |
climbed trees ( leaves ) were falling. As John hunted the frightened deer ( it) escaped through the woods. When Anne visited the British relatives ( they) were moving to London. When Rita washed her favorite sweater ( it) was torn to shreds. When Joan left her old boyfriend ( he) stalked her for two months. While the a... | https://arxiv.org/abs/2505.22236v1 |
arXiv:2505.22240v1 [cs.CL] 28 May 2025BioHopR: A Benchmark for Multi-Hop, Multi-Answer Reasoning in Biomedicine Yunsoo Kim1Yusuf Abdulle1,2Honghan Wu1,3 1UCL2King’s College London3University of Glasgow {yunsoo.kim.23, honghan.wu}@ucl.ac.uk Abstract Biomedical reasoning often requires travers- ing interconnected relatio... | https://arxiv.org/abs/2505.22240v1 |
comprehensive, multi-answer re- sponses. Existing benchmarks for multi-hop reasoning, such as Hetionet (Himmelstein et al., 2017) and other biomedical QA datasets (Rao et al., 2022), have laid the groundwork for evaluating multi- hop capabilities in the biomedical domain. How- ever, these benchmarks primarily focus on ... | https://arxiv.org/abs/2505.22240v1 |
reasoning capabili- ties of LLMs (Kim et al., 2024). While this is a step forward, it remains constrained to single-hop reasoning and does not address the need for multi- hop reasoning or the generation of multiple valid answers—a common requirement in biomedical inquiries. Knowledge Graph Question Answering. Knowledge... | https://arxiv.org/abs/2505.22240v1 |
multi-answer rea- soning, ensuring that the benchmark captures the intricate relational structures and knowledge depen- dencies present in biomedical science. The differ- ences between our dataset and relevant datasets are summarized in Table 1. 3 BioHopR: Multi-hop Reasoning in Biomedicine One to many relationship (Bi... | https://arxiv.org/abs/2505.22240v1 |
− − − − − → Bridge (Drug) treats− − − → Target (Diseases) .(2) Here, the bridge node (e.g., drug), used to query for 1-hop questions, serves as the intermediate entity linking the query and target. Answer Definition. The target nodes are the final answers to the query. For 1-hop reasoning, this corresponds to all nodes... | https://arxiv.org/abs/2505.22240v1 |
type is summarized in Tables 2 and 3. The restriction to one-to-many-to-many relation- ships ensures that the dataset mirrors real-world biomedical reasoning scenarios, where single en- tities often relate to multiple downstream entities. This design makes the dataset uniquely suited for evaluating large language model... | https://arxiv.org/abs/2505.22240v1 |
truth answers. The cosine similarity for a prediction pand an answer aiis defined as: cos(p, ai) =p·ai ∥p∥∥ai∥. (8) If the maximum cosine similarity across all ground truth answers satisfies: max i∈{1,...,n}cos(p, ai)> τ, (9) 1https://platform.openai.com/docs/modelsthen the prediction is considered a true positive. The... | https://arxiv.org/abs/2505.22240v1 |
performance ( BOTH_WR: 52.14% ). However, even the best-performing models show substantial error rates in both hops, indicating that multi-step inference remains a bottleneck. 5.3 Multi-Hop Reasoning Remains a Bottleneck Across all models, performance declines sharply from 1-hop to 2-hop tasks. For example, GPT4O’s pre... | https://arxiv.org/abs/2505.22240v1 |
were not explicitly part of the predefined an- swer set. This behavior suggests that proprietary models may apply broader medical reasoning com- pared to open-source models. Proprietary models generally outperform open-source models in both task adherence and reasoning accuracy. 5.5 Ablation Study: Prompting Strategy P... | https://arxiv.org/abs/2505.22240v1 |
listing multiple answers rather than specifying a specific answer. Based on these findings, we restricted our evalu- ation of all other models to Single-Answer Prompt- ing. This decision is motivated by higher robust- ness and computational overhead of multi-answer prompting. Also in many real-world scenarios, users ty... | https://arxiv.org/abs/2505.22240v1 |
Kexin Huang, and Marinka Zitnik. 2023. Building a knowledge graph to enable pre- cision medicine. Scientific Data , 10(1):67. Junying Chen, Zhenyang Cai, Ke Ji, Xidong Wang, Wanlong Liu, Rongsheng Wang, Jianye Hou, and Benyou Wang. 2024. Huatuogpt-o1, towards med- ical complex reasoning with llms. arXiv preprint arXiv:... | https://arxiv.org/abs/2505.22240v1 |
answering using knowledge graph embeddings and language models. arXiv preprint arXiv:2211.05351 . François Remy, Kris Demuynck, and Thomas De- meester. 2023. Biolord-2023: Semantic textual rep- resentations fusing llm and clinical knowledge graph insights. arXiv preprint arXiv:2311.16075 . Julian Schnitzler, Xanh Ho, J... | https://arxiv.org/abs/2505.22240v1 |
Whereas, for complex conditions such as Lung Cancer and Alzheimer’s Disease, we can evaluate the models ability to reason through more intricate, multi-factorial diseases. Our qualitative analysis showed several key find- ings regarding the models’ reasoning capabilities across different diseases. Interestingly, none o... | https://arxiv.org/abs/2505.22240v1 |
1 predictions, failing to generate the correct responses. Similarly, LLaMA Ultra Medi- cal did not fully adhere to the prompt’s instructions — when asked to provide a single answer for Hop 2, it instead generated a list of multiple possible an- swers. While the listed responses were correct, this deviation indicates a ... | https://arxiv.org/abs/2505.22240v1 |
arXiv:2505.22251v1 [eess.AS] 28 May 2025Evaluation of LLMs in Speech is Often Flawed: Test Set Contamination in Large Language Models for Speech Recognition Yuan Tseng* Titouan Parcollet Rogier van Dalen Shucong Zhang Sourav Bhattacharya AI Center Cambridge, Samsung, United Kingdom {t.parcollet, r.vandalen, s1.zhang, s... | https://arxiv.org/abs/2505.22251v1 |
with nearly two-thirds of each of the four LibriSpeech evaluation sets and one-third of the Common V oice English evaluation sets (Section II). Second, we train billion-parameter LLMs from scratch, and intentionally contaminate some of them by interspersing LibriSpeech books throughout pretraining. Results comparing co... | https://arxiv.org/abs/2505.22251v1 |
LibriSpeech dev and test sets. B. Common Voice Common V oice, as of version 20.0, is a multilingual corpus containing more than 22,000 hours of read speech from 133 languages. Recordings are gathered through crowdsourcing, where volunteers read sentences from Wikipedia or other user-submitted sources. However, many LLM... | https://arxiv.org/abs/2505.22251v1 |
are trained on the Pile [14], a dataset also used for training various other LLMs [30]– [33]. Note that Pythia models were trained on document segments without start-of-sentence tokens, hence the likelihoods reported do not consider the first token in each sentence. We restore casing and punctuation to both leaked and ... | https://arxiv.org/abs/2505.22251v1 |
varying amounts of data. An increase in training data does not “dilute” the effects of contamination; leaked sentences still remain more likely to be generated by contaminated LLMs. –0.05 –0.04 –0.03 –0.02 –0.01 0 0.01 0 Contamination Level /check /check/check (c) 1.4B LLMs that are contaminated once and twice compared... | https://arxiv.org/abs/2505.22251v1 |
2, this resulted in a more pronounced likelihood improvement for leaked sentences (–0.038), and minor improvements for non- leaked sentences (–0.006). We also consider an alternative explanation of the improved likelihoods of leaked sentences: Contaminated LLMs are more likely to generate leaked sentences compared to u... | https://arxiv.org/abs/2505.22251v1 |
[35], [36]. As shown in Figure 4, the LLM pre- dicts transcriptions of input speech in an autoregressive fashion, given sequences of speech embeddings obtained from the encoder and an optional text prompt embedding sequence. Following [17], we freeze both the speech encoder and the LLM, and only fine-tune parameters of... | https://arxiv.org/abs/2505.22251v1 |
performance on leaked data, but also affect how well the model generalises to unseen data. However, this increased probability of generating leaked sentences is not necessarily reflected in ASR error rates, as the error rates of contaminated systems can worsen even when the likelihoods of the same system improve. Manua... | https://arxiv.org/abs/2505.22251v1 |
Automatic Speech Recognition and Understanding Workshop , 2023. [6] R. Ma, M. Qian, P. Manakul, M. Gales, and K. Knill, “Can generative large language models perform asr error correction?” arXiv preprint arXiv:2307.04172 , 2023. [7] G. Song, Z. Wu, G. Pundak, A. Chandorkar, K. Joshi, X. Velez, D. Caseiro, B. Haynor, W.... | https://arxiv.org/abs/2505.22251v1 |
recognition,” arXiv preprint arXiv:2410.03752 , 2024. [20] G. Yang, Z. Ma, Z. Gao, S. Zhang, and X. Chen, “Ctc-assisted llm-based contextual asr,” in Proceedings of Spoken Language Technology Workshop , 2024, pp. 126–131. [21] H. Laurenc ¸on, L. Saulnier, T. Wang, C. Akiki, A. V . del Moral, T. L. Scao, L. V . Werra, C... | https://arxiv.org/abs/2505.22251v1 |
large language models across training and scaling,” in Proceedings of International Conference on Machine Learning , vol. 202, 2023, pp. 2397– 2430. [31] S. Black, S. Biderman, E. Hallahan, Q. Anthony, L. Gao, L. Golding, H. He, C. Leahy, K. McDonell, J. Phang, M. Pieler, U. S. Prashanth, S. Purohit, L. Reynolds, J. To... | https://arxiv.org/abs/2505.22251v1 |
arXiv:2505.22273v1 [cs.CL] 28 May 2025Comprehensive Evaluation on Lexical Normalization: Boundary-Aware Approaches for Unsegmented Languages Shohei Higashiyama and Masao Utiyama National Institute of Information and Communications Technology, Kyoto, Japan {shohei.higashiyama,mutiyama}@nict.go.jp Abstract Lexical normal... | https://arxiv.org/abs/2505.22273v1 |
superior recall even with fewer in- stances. Third, domains rich with unknown infor- mal words exhibit low performance, especially the typo-correction domain.1 1Our dataset and code will be available at {URL} . 2 Related Work Text normalization has been studied for some rep- resentative purposes: mapping dialectal vari... | https://arxiv.org/abs/2505.22273v1 |
tion instances, i.e., non-standard and standard form pairs (details in Appendix A.1). The large data size and domain diversity of our dataset are ad- vantages over existing Japanese datasets, enabling multi-perspective evaluations, as shown in §6. Basic Designs We followed Higashiyama et al. (2021b)’s annotation criter... | https://arxiv.org/abs/2505.22273v1 |
to predict the set of non-standard word spans and their standard forms S={(b, e, s )}. Here, (b, e)(0≤b≤e≤n) indicates a span of an non-standard word xb:ewith length e−bin the source sentence, and sindicates its standard form. Each standard form sis a string with length ≥0, 3cwj-3.1.0 (https://clrd.ninjal.ac.jp/unidic/... | https://arxiv.org/abs/2505.22273v1 |
the chunk has no need for normalization. Notably, we estimate a length value per character token, so an non-standard word chunk of mcharacters yields mredundant length values (e.g., yl 0:4= [5,5,5,5]). We determine a single length value by taking a majority vote within the chunk. Thus, the combina- tion of two sequence... | https://arxiv.org/abs/2505.22273v1 |
instruction tuning. Here, {inst} ,{src} , and {tgt} are placeholders, and EOSrepresents the end of text token. We use En- glish instruction texts explaining the corresponding content for STRUCT andSPAN as described above; the exact wording is provided in Appendix C.4. 6 Experiments We set the following experimental que... | https://arxiv.org/abs/2505.22273v1 |
Results and Analysis 7.1 Normalization Accuracy for Japanese We evaluated LN methods with three type of ar- chitectures on the JMLN test-C set. Table 3 shows the performance of encoder-only models, and encoder-decoder and decoder-only models with both S TRUCT and S PAN approaches.10 The observed results are as follows.... | https://arxiv.org/abs/2505.22273v1 |
models (E: encoder, S: seq2seq, D: decoder). “ ,→” indicates the continual pre-trained model derived from the base model listed in the previous row. The best score within each size group is shown in bold . For each backbone model, the better of the STRUCT andSPAN approaches is underlined . Scores where the SPAN approac... | https://arxiv.org/abs/2505.22273v1 |
critical drawback, but they run much faster on the H200. Thus, when high-spec GPUs are available, Sarashina models are viable options in accuracy-critical scenarios. 7.4 Investigation of Training Data Size We generated size- Ntraining sets by sampling random N∈{500,1k,2k,4k,8k,12k}sentences from the entire JMLN trainin... | https://arxiv.org/abs/2505.22273v1 |
(Norm-In-Lex rate) and (ii) the proportion of origi- 7 Data Register Surf-Out-Train Avg-3M DeBERTa-L T5-L Sarashina2.2-3B F0.5 P R P R P R 01BJ-OC Q&A site 0.59 0.595 0.654 0.416 0.572 0.369 0.717 0.578 02BJ-OY Blog 0.48 0.713 0.720 0.543 0.741 0.551 0.802 0.662 03RC-BLG Blog 0.37 0.775 0.801 0.691 0.793 0.664 0.803 0.... | https://arxiv.org/abs/2505.22273v1 |
while cutting- edge decoder-only models delivered high precision, notably high recall, and reasonable throughput on a high-spec GPU. (2) Normalization accuracy consis- tently increased with training data size, yet even 4k– 8k training sentences yielded reasonable precision around 0.7. Sarashina-series models, in partic... | https://arxiv.org/abs/2505.22273v1 |
annotators and the pay- ment amount to each annotator were determined by the company. The annotation work was per- formed by four annotators, including an annotation manager, all of whom are native Japanese speakers. Under the contract for the annotation work, it was agreed that the intellectual property rights to the ... | https://arxiv.org/abs/2505.22273v1 |
Short Papers) , pages 4171–4186, Minneapolis, Minnesota. Association for Computational Linguistics. Kazuki Fujii, Taishi Nakamura, Mengsay Loem, Hi- roki Iida, Masanari Ohi, Kakeru Hattori, Hirai Shota, Sakae Mizuki, Rio Yokota, and Naoaki Okazaki. 2024. Continual pre-training for cross-lingual LLM adaptation: Enhancin... | https://arxiv.org/abs/2505.22273v1 |
Soumya Batra, Spencer Whitman, Sten Sootla, Stephane Collot, Suchin Gururangan, Syd- ney Borodinsky, Tamar Herman, Tara Fowler, Tarek Sheasha, Thomas Georgiou, Thomas Scialom, Tobias Speckbacher, Todor Mihaylov, Tong Xiao, Ujjwal 10 Karn, Vedanuj Goswami, Vibhor Gupta, Vignesh Ramanathan, Viktor Kerkez, Vincent Gonguet... | https://arxiv.org/abs/2505.22273v1 |
Bansal, Nandhini Santhanam, Natascha Parks, Natasha White, Navyata Bawa, Nayan Singhal, Nick Egebo, Nicolas Usunier, Nikhil Mehta, Nikolay Pavlovich Laptev, Ning Dong, Norman Cheng, Oleg Chernoguz, Olivia Hart, Omkar Salpekar, Ozlem Kalinli, Parkin Kent, Parth Parekh, Paul Saab, Pavan Balaji, Pe- dro Rittner, Philip Bo... | https://arxiv.org/abs/2505.22273v1 |
lexical normalization. In Proceedings of the 2021 Conference of the North American Chap- ter of the Association for Computational Linguistics: Human Language Technologies , pages 5532–5541, Online. Association for Computational Linguistics. Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen- Zhu, Yuanzhi Li, Shean ... | https://arxiv.org/abs/2505.22273v1 |
Anne Göhring, Tanja Samardži ´c, and Elisabeth Stark. 2018. Encoder- decoder methods for text normalization. In Proceed- ings of the Fifth Workshop on NLP for Similar Lan- guages, Varieties and Dialects (VarDial 2018) , pages 18–28, Santa Fe, New Mexico, USA. Association for Computational Linguistics. Kikuo Maekawa, Ma... | https://arxiv.org/abs/2505.22273v1 |
Saito, Kyosuke Nishida, Kugatsu Sadamitsu, Ku- niko Saito, and Junji Tomita. 2017. Automatically extracting variant-normalization pairs for Japanese text normalization. In Proceedings of the Eighth In- ternational Joint Conference on Natural Language Processing (Volume 1: Long Papers) , pages 937– 946, Taipei, Taiwan. ... | https://arxiv.org/abs/2505.22273v1 |
Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Ł ukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in Neural Information Pro- cessing Systems , volume 30. Curran Associates, Inc. Aobo Wang, Min-Yen Kan, Daniel Andrade, Takashi Onishi, and Kai Ishikawa. 2013. Chinese informal word normalizat... | https://arxiv.org/abs/2505.22273v1 |
286 Total 18,862 316,216 8,375 Table 8: Statistics of the JMLN Cur-sets. ID Name #Sent #Word #Nrom 01 BJ-OC 200 3,981 33 02 BJ-OY 201 3,821 56 03 RC-BLG 200 2,903 57 04 RC-REV 200 3,872 27 05 RK-ICB 200 2,942 16 06 RK-TRV 200 3,139 18 07 RK-RCP 200 2,763 25 08 AM 200 2,932 12 09 NC-VID 150 2,418 23 10 NC-PED 182 2,672 ... | https://arxiv.org/abs/2505.22273v1 |
Social media posts 2020–2022 12 JW Japanese Wikipedia Typo Dataset Encyclopedia edit history –2021 13 NU Nagoya University Conversation Corpus Conv. transcriptions 2001–2003 14 SK Skype Conversation Corpus Conv. transcriptions 2012 Table 10: Data sources of JMLN. The Year column indicates the years of original text pub... | https://arxiv.org/abs/2505.22273v1 |
level alignments remains challenging. A.6 Inter-Annotator Agreement Table 12 shows the detailed statistics on inter- annotator agreement. Sentences in domains 01–07 were annotated by Annotators B and C, and those in domains 08–10 were by B and D. Notably, after we fixed annotation disagreements for these sentences thro... | https://arxiv.org/abs/2505.22273v1 |
rate 3e-5 Learning rate scheduler linear Warmup ratio 0.1 Gradient norm clipping threshold 1.0 Optimizer AdamW Training epochs ja: 30; th: 15 Batch size ja: 32; th: {4, 16} Learning rate 2e-4 Learning rate scheduler constant Warmup ratio 0.1 Gradient norm clipping threshold 1.0 Optimizer AdaFactor Beam width for infere... | https://arxiv.org/abs/2505.22273v1 |
a single warm-up inference followed by three inference passes over all 3,786 JMLN test sentences (a total of 11.1k characters), which were sorted by increasing token count according to its tokenizer. These evalua- tions were run at multiple batch sizes; we selected the batch size that yielded the highest throughput (sh... | https://arxiv.org/abs/2505.22273v1 |
F 0.5scores in Table3 for each model series—T5, mT5, Llama-3.2, Qwen2.5, and Sarashina2.2. The scores for the STRUCT andSPAN approaches are shown with solid and dotted lines, re- spectively. results. The encoder-only models (3 backbone models ×2 model size) with the FULL-SEG approach obtained +0.01–0.05 F 0.5gains from... | https://arxiv.org/abs/2505.22273v1 |
F-S EG-POS 0.978 0.963 0.636 0.976 0.649 0.532 0.621 0.508 0.970 96.56 RoBERTa-BP-S EG – – 0.630 – 0.662 0.515 0.626 0.488 0.976 97.04 F-S EG 0.982 – 0.669 0.980 0.726 0.537 0.678 0.502 0.981 97.22 F-S EG-POS 0.982 0.971 0.677 0.981 0.713 0.566 0.678 0.525 0.973 97.27 DeBERTa-BP-S EG – – 0.637 – 0.678 0.516 0.638 0.489... | https://arxiv.org/abs/2505.22273v1 |
(husband) FP&FN △ (e)BJ-OY :(···)ウォルトンでカンツリ Neg:Unkと利根川水系でバスをやんべ。 (I’m going kan-tsuri (= fishing at a managed sport) at Walton and bass fishing in Tone River system.) Gold – DeBERTa-L – – – T5-L – – – Sarashina2.2-3B カンツリ →キャッチ＆リリース OOL (catch & release) FP ✗ (f)NU:(···)最高気温25度とかっしょう Pos:Unk? (The high is like 25°C or... | https://arxiv.org/abs/2505.22273v1 |
arXiv:2505.22293v1 [cs.CL] 28 May 2025Compensating for Data with Reasoning: Low-Resource Machine Translation with LLMs Samuel Frontull andThomas Ströhle Department of Computer Science / University of Innsbruck {samuel.frontull,thomas.stroehle}@uibk.ac.at Abstract Large Language Models (LLMs) have demon- strated strong ... | https://arxiv.org/abs/2505.22293v1 |
LLMs have opened up new possibilities, especially in low-resource scenarios. This work aims to evaluate the effectiveness of different prompting techniques for machine transla- tion to and from the low-resource language Ladin using LLMs, in the case of Italian and the two stan- dard variants of Ladin: Val Badia and Ghe... | https://arxiv.org/abs/2505.22293v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.