text string | source string |
|---|---|
annotator almost two days to annotate, whileLLMannotationtookonly3hours. Theproblemofannotationdurationisparticularlyacute when several parallel annotations are running in a company. The ability to run LLM annotation on multiple GPUs reduces the queue and speeds up the entire annotation process. 17 4 Methodological Not... | https://arxiv.org/abs/2505.18688v1 |
the chat (Table 6.4). •Selecting documents relevant to the verifiable intent or intents in the annotation (Table 6.5). 5.2.1 Retrievals •multilingual-e5-large-instruct [30] Instruction-TunedMultilingualEmbeddersupports 100languages,builtonXLM-RoBERTa- large. it is fine-tuned with an instruction format. •BM25 [22] Okapi... | https://arxiv.org/abs/2505.18688v1 |
50 60 70 80 90 100 Percentage of replaced intents0102030405060708090100Accuracy (normalized) Figure 6.1: Impact of False Positive on the classifier (5 experiments) If we replace the intent in the training sample with the closest similar one and train the text classifier on this, the quality of the test sample begins to... | https://arxiv.org/abs/2505.18688v1 |
uncertain. From the beginning of the experiments, it is clear that human annotators are more accurate in multi-class annotation tasks than LLM. 24 6.2.2 Binary Table 6.2: Metrics for LLMs (Binary) ModelPrec. 1, pos.Prec. 0, neg.Rec. 1, pos.Rec. 0, neg.Acc.f1-score (macro avg.)Coverage (perc.) Human 0.8748 0.7162 0.7834... | https://arxiv.org/abs/2505.18688v1 |
the answers "yes" and "no" are opposite in meaning, and LLM tends to give the maximum probability to one or the another answer. Finally, the change in distribution means that the model is highly confident in its answer, since all the necessary reasoning for making a decision was provided, even if it was incorrect (mode... | https://arxiv.org/abs/2505.18688v1 |
RAG (Binary) ModelPrec. 1, pos.Prec. 0, neg.Rec. 1, pos.Rec. 0, neg.Acc.f1-score (macro avg.)Coverage (perc.) Human 0.8748 0.7162 0.7834 0.8297 0.8018 0.7977 100% Llama3.3-70b text-approach0.8779 0.8178 0.9743 0.4600 0.8711 0.7562 100% Qwen2.5-32b text-approach0.8613 0.6531 0.8726 0.6307 0.8059 0.7543 100% Llama3.3-70b... | https://arxiv.org/abs/2505.18688v1 |
the annotation task. it is also important to remember that RAG documents were added to the prompt during retraining. The improvement in metrics can also be attributed to the correct handling of retrieved documents, which Qwen2.5-32b learned from Deepseek-r1 and GPT-4o. 32 6.7 Final best approaches 6.7.1 Multi-class Tab... | https://arxiv.org/abs/2505.18688v1 |
improvedperformanceintheprob-approach—evenwhenreasoninggenerationisdisabledbydesign. The results of reasoning fine-tuning underscore the importance of correct initial prompt processing by the LLM and suggest that the model may "know" the correct answer from the very start of its generation. The focus of future work sho... | https://arxiv.org/abs/2505.18688v1 |
domain transfer to new intent taxonomies, and tighter integration of LLM-based annotators within active-learning loops to further improve both coverage and accuracy. 37 References [1] Jiangpeng He, Runyu Mao, Zeman Shao, and Fengqing Zhu. Incremental Learning In Online Scenario. 2021. arXiv: 2003.13191 [cs.CV] .url:htt... | https://arxiv.org/abs/2505.18688v1 |
Labels from Radiology Reports for Training Convolutional Neural Networks”.In: Academic Radiology 32.5(2025),pp.2402–2410. issn:1076-6332. doi:https: //doi.org/10.1016/j.acra.2024.12.028 .url:https://www.sciencedirect.com/ science/article/pii/S1076633224009966 . [17] Hamidreza Rouzegar and Masoud Makrehchi. Enhancing Te... | https://arxiv.org/abs/2505.18688v1 |
2025. arXiv: 2501.12948 [cs.CL] .url:https://arxiv.org/abs/2501. 12948. 41 A Appendix Multi-class prompt Hello! Here are the instructions you should follow when communicating with me: 1. Youareanintelligentdataannotationassistant. Youreceivetextthatauserhaswrittenintheapp’ssupport chat, along with several hypothetical ... | https://arxiv.org/abs/2505.18688v1 |
receive bonuses that you can spend on self-development and training. 4. If you are able to classify the text clearly, enter the correct class after the [ANSWER] tag. Use the exact number that was provided in the list, as this is necessary for subsequent quality assessment. 5. The text may not belong to any of the provi... | https://arxiv.org/abs/2505.18688v1 |
verify the correctness of the prediction. Reason consistently, step by step, consider all possible options, and be sure to use the class description and examples provided for it. Also, clarify the meaning of ambiguous words that may imply that the text does not belong to this class. Do not guess what the client means o... | https://arxiv.org/abs/2505.18688v1 |
Benchmarking and Rethinking Knowledge Editing for Large Language Models Guoxiu He1∗Xin Song1∗Futing Wang2Aixin Sun3† 1East China Normal University,2Westlake University,3Nanyang Technological University gxhe@fem.encu.edu.cn ,xsong2023@stu.ecnu.edu.cn , wangfuting@westlake.edu.cn ,axsun@ntu.edu.sg Abstract Knowledge edit... | https://arxiv.org/abs/2505.18690v1 |
employed. Infer ence settings show whether Auto regressive setting is applied instead of teacher-forcing. Under Edits , we clarify if Single orSequential editing is tested. For evaluation Dimensions , we specify if Reliability, Generalization, Locality, and Port ability are considered. Datasets LLM Infer Edits Dimensio... | https://arxiv.org/abs/2505.18690v1 |
conduct comprehensive benchmarking experiments that simulate practical knowledge editing scenarios. We recognize that real-world applications often require multiple sequential edits as knowledge evolves over time. We adopt a unified autoregressive inference setting, as edited LLMs should not rely on ground truth access... | https://arxiv.org/abs/2505.18690v1 |
knowledge data instance associated with an LLM. However, the hyper-network for a particular LLM limits their scalability in sequential editing scenarios. Furthermore, the additional training process incurs significant time and computational costs. Additionally, the necessity of modifying parameters for a limited amount... | https://arxiv.org/abs/2505.18690v1 |
impact of the ripple effect in the hidden space, which significant hinders the effectiveness of editing tasks and overall performance of edited LLMs. Similarly, [ 28] demonstrates that improvements in factuality come at the cost of a significant decline in reasoning, natural language inference, and question-answering a... | https://arxiv.org/abs/2505.18690v1 |
the number of updates, an idealized scenario; external memory -based methods only update an external knowledge base, such as a text corpus or vector store, that stores newly introduced knowledge. Evaluation Dimensions: After incorporating tpieces of new knowledge, i.e.undergoing trounds of updates, the edited LLM ftis ... | https://arxiv.org/abs/2505.18690v1 |
the most relevant knowledge from the knowledge base. If relevant knowledge is found, it is included in the prompt along with the question for the LLM to answer. If not, the LLM answers based solely on the question. For detailed implementation, see the Appendix A. Knowledge Editing Datasets We use two widely adopted con... | https://arxiv.org/abs/2505.18690v1 |
external memory-based methods, the original event text is directly provided as context input or integrated into the editable memory. Experimental Environment The experiments are executed on 8 NVIDIA A800 GPUs under a Linux system. All methods are implemented using EasyEdit [ 76].7Evaluation on reasoning benchmarks is c... | https://arxiv.org/abs/2505.18690v1 |
This highlights the practical potential of in-context learning-based approaches in real-world knowledge editing tasks. RQ2: Can knowledge editing methods help reasoning LLMs integrate new facts without degrading their reasoning ability? As LLMs are increasingly applied to complex tasks involving multi-step reasoning, m... | https://arxiv.org/abs/2505.18690v1 |
0.95 0.83 1.37 0.78 0.98 MEND 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 AdaLoRA 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 WISE 21.33 13.35 14.18 5.02 13.47 0.36 0.47 29.36 3.06 8.31 4.41 5.96 5.21 1.94 4.38 IKE 62.69 61.38 21.75 26.14 42.99 60.07 58.28 3... | https://arxiv.org/abs/2505.18690v1 |
2: Performance changes of knowledge editing methods during sequential editing on the ZsRE dataset. The x-axis represents the number of edits: 1, 10, 100, and the full dataset. the answer. These behaviors underscore key limitations of current knowledge editing methods when applied to reasoning-focused LLMs. Concrete exa... | https://arxiv.org/abs/2505.18690v1 |
0.00 0.00 WISE 8.00 7.00 3.00 2.59 5.15 0.00 0.00 6.00 2.86 2.22 IKE 87.00 92.00 5.22 25.84 52.52 87.00 92.00 5.22 25.84 52.52 ICE 54.00 45.00 29.64 38.93 41.89 54.00 45.00 29.64 38.93 41.89 GRACE 12.00 0.00 23.54 4.96 10.13 12.00 0.00 23.49 4.96 10.11 SCR 72.00 67.00 23.50 34.83 49.33 84.00 76.00 15.70 35.67 52.84 Tab... | https://arxiv.org/abs/2505.18690v1 |
Portability Locality Avg. Pre-edit 5.92 43.75 24.84 7.08 56.73 31.91 9.01 48.73 28.87 ROME 0.00 0.00 0.00 0.00 0.00 0.00 0.33 0.00 0.17 MEMIT 0.00 0.00 0.00 - - - - - - PMET 0.00 0.08 0.04 - - - - - - RECT 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 AlphaEdit - - - 14.22 19.96 17.09 12.15 17.87 15.01 FT-L 3.53 0.00 1.... | https://arxiv.org/abs/2505.18690v1 |
inference latency. 5 Conclusion Through a unified setting for datasets, LLMs, inference setting, editing scenarios, and evaluation dimensions, our benchmarking shows that most parameter-modification-based knowledge editing methods perform significantly below expectations. These methods not only fail to support flexible... | https://arxiv.org/abs/2505.18690v1 |
M. Subbiah, J. D. Kaplan, P. Dhariwal, A. Neelakantan, P. Shyam, G. Sastry, A. Askell et al. , “Language models are few-shot learners,” Advances in neural information processing systems , vol. 33, pp. 1877–1901, 2020. [10] L. Ouyang, J. Wu, X. Jiang, D. Almeida, C. Wainwright, P. Mishkin, C. Zhang, S. Agarwal, K. Slama... | https://arxiv.org/abs/2505.18690v1 |
Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , 2023, pp. 4862–4876. [24] R. Cohen, E. Biran, O. Yoran, A. Globerson, and M. Geva, “Evaluating the ripple effects of knowledge editing in language models,” Transactions of the Association for Computational Linguistics , vol. 12, pp... | https://arxiv.org/abs/2505.18690v1 |
L. Chen, and W. Lam, “Consecutive model editing with batch alongside hook layers,” arXiv preprint arXiv:2403.05330 , 2024. [38] C. Hu, P. Cao, Y . Chen, K. Liu, and J. Zhao, “Wilke: Wise-layer knowledge editor for lifelong knowledge editing,” in Findings of the Association for Computational Linguistics ACL 2024 , 2024,... | https://arxiv.org/abs/2505.18690v1 |
knowledge editing for llms with retrieval-augmented continuous prompt learning,” in Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , 2024, pp. 13 565–13 580. [52] P. Das, S. Chaudhury, E. Nelson, I. Melnyk, S. Swaminathan, S. Dai, A. Lozano, G. Kollias, V . Chenthamarakshan, J. N... | https://arxiv.org/abs/2505.18690v1 |
S. Cheng, Z. Li, S. Deng, H. Chen, and N. Zhang, “Editing large language models: Problems, methods, and opportunities,” in The 2023 Conference on Empirical Methods in Natural Language Processing . [67] J. Li, E. Beeching, L. Tunstall, B. Lipkin, R. Soletskyi, S. Huang, K. Rasul, L. Yu, A. Q. Jiang, Z. Shen et al. , “Nu... | https://arxiv.org/abs/2505.18690v1 |
Figure 4, the core inference involves a two-step knowledge selection phase, including semantic filtering and knowledge confirmation, followed by contextual reasoning phase. Figure 4: The Edited Memory is a dynamic textual knowledge base that can be expanded as needed. Phase 1: The retriever first applies semantic filte... | https://arxiv.org/abs/2505.18690v1 |
from Event Please extract one or more (subject, relation, object) triples from the following event. Each triple must express a complete, standalone fact, with no redundancy or dependency on other triples. Instructions: If the event contains multiple independent facts, extract multiple triples , one per fact. TheSubject... | https://arxiv.org/abs/2505.18690v1 |
Merovingian dynasty and the first king of the Franks. Dagobert was a later king, not his father. **Answer:** Clovis I. Case Study 2: Reasoning with Conflicting Knowledge Edit Prompt: What is Atreus’s spouse’s name? Target New Answer: Darius III Generated Reasoning Content: 20 Darius III of Persia. Atreus was the King o... | https://arxiv.org/abs/2505.18690v1 |
Generated Answer Content: The constellation in which the star ¨Hornaxïs located is **Hercules**. 21 **Answer:** Hercules Case Study 4: Losing Focus During Reasoning Edit Prompt: The name of the child of Kanye West is Target New Answer: William Walker Scranton Generated Reasoning Content: William Walker, but that’s not ... | https://arxiv.org/abs/2505.18690v1 |
arXiv:2505.18703v1 [cs.CL] 24 May 2025Towards Semantic Integration of Opinions: Unified Opinion Concepts Ontology and Extraction Task Gaurav Negi, Dhairya Dalal, Omnia Zayed, and Paul Buitelaar Insight SFI Research Centre for Data Analytics Data Science Institute University of Galway {gaurav.negi, omnia.zayed, paul.bui... | https://arxiv.org/abs/2505.18703v1 |
is valid for specific individuals or groups en- gaged in " doing heavy computations ". The reason for opinion is also expressed, i.e. “it had only about 2-1/2 hours" . None of the existing opinion mining formulations enable these extractions. This work investigates semantic representations of opinions to enrich their e... | https://arxiv.org/abs/2505.18703v1 |
is a standardised schema designed to annotate and describe subjec- tive opinions expressed on the Semantic Web and in information systems (Sánchez-Rada et al., 2016; Buitelaar et al., 2013). However, the Marl ontol- ogy cannot describe fine-grained opinion mining currently being researched in NLP. Schouten and Frasinca... | https://arxiv.org/abs/2505.18703v1 |
Part Aspect term Described Object Entity Target Entity NA Opinion Time (t) NA NA Opinion Qualifier Qualifier NA Opinion Reason Reason NA Opinion holderHolder Entity Holder Span Table 1: Unified Opinion Concepts (UOC) Second, we leverage the explicit and implicit hi- erarchical structures described in the NLP literature... | https://arxiv.org/abs/2505.18703v1 |
agreement with the ABSA literature. Figure 4 illustrates an instance of the Target class, its constituents, and their relationships. It addresses the semantic formulation for extracting the multiple facets of an opinion’s target. The object property is- ExpressedOnTarget describes its relationship with Opinion class. T... | https://arxiv.org/abs/2505.18703v1 |
speci- fies the subset for whom the battery life would be inadequate. The property hasQualifier describes the relationship between Opinion andQualifier . Reason : A reason represents an opinion’s justifica- tion or underlying cause. This concept is connected to the Opinion class via the property hasReason and, like Qua... | https://arxiv.org/abs/2505.18703v1 |
al. (2023). Therefore, our metric of choice is the Component-level exact match metric discussed in the remainder of this section. In the dataset with text instances {Ti}N i=1for each text instance Tithere exists the ground truth opinion annotation Ogiis a set of opinions Ogi= {ogi,j|j= 1,2, ...,|Ogi|}and the correspond... | https://arxiv.org/abs/2505.18703v1 |
al., 2019) ✓✓ ✓ SS22(Barnes et al., 2022) ✓✓ ✓ ✓ Ai23(Mamta and Ekbal, 2023) ✓✓ ✓ ME 23(Cai et al., 2023) ✓✓✓✓✓ Table 2: Alignment of datasets with UOC as described by Eq. 1. It should be noted that none of the datasets have annotations corresponding to qandr. evaluation dataset. ME 23was selected based on its multi-do... | https://arxiv.org/abs/2505.18703v1 |
LLMs, we report the component-level exact match f1 scores (4.2). 4.5 Experimental Settings The experiments were conducted on a machine with two NVIDIA RTX A6000 48GB GPUs. We employ the following open-weight LLMs for the experiments: Gemma-2 (9B, 27B) (Mesnard et al., 2024) , Mistral 7B (Jiang et al., 2023), Mixtral 8x... | https://arxiv.org/abs/2505.18703v1 |
the mean values. For NLPrompt, the (E-D-F) variant per- forms the best, and GPT-4o performs the best over- all. Similarly, for OntoPrompt, JSON-LD is the best-performing ontology serialisation format, and GPT-4o is the best-performing model. Task ModelComponent-Level EM P R F1 ASTEGEN-SCL-NAT 60.25 70.14 64.82 MVP 61.2... | https://arxiv.org/abs/2505.18703v1 |
on ABSA datasets and do not have difficulty identi- fying aspect terms. Being trained on ACOS tasks, the GEN_SCL_NAT and MVP models cannot ex- tract all the UOCE concepts. LLMs struggle to recognize qualifiers and reasons in our benchmark dataset as they require nuanced semantic under- standing. We believe there is amp... | https://arxiv.org/abs/2505.18703v1 |
under- stand better and utilize the comprehensive semantic structure introduced in this work.Acknowledgments This work was conducted with the financial sup- port of the Science Foundation Ireland (SFI) un- der Grant Number SFI/12/RC/2289_P2 (Insight_2) and was also supported by funding from the Irish Research Council (... | https://arxiv.org/abs/2505.18703v1 |
WSDM 2008, Palo Alto, California, USA, February 11-12, 2008 , pages 231–240. ACM. Zhibin Gou, Qingyan Guo, and Yujiu Yang. 2023. MvP: Multi-view prompting improves aspect sentiment tu- ple prediction. In Proceedings of the 61st Annual Meeting of the Association for Computational Lin- guistics (Volume 1: Long Papers) , ... | https://arxiv.org/abs/2505.18703v1 |
Beijing, China, August 3-9, 2013 , pages 2134–2140. IJCAI/AAAI. Dehong Ma, Sujian Li, and Houfeng Wang. 2018. Joint learning for targeted sentiment analysis. In Proceed- ings of the 2018 Conference on Empirical Methods in Natural Language Processing , pages 4737–4742, Brussels, Belgium. Association for Computational Li... | https://arxiv.org/abs/2505.18703v1 |
on Semantic Evaluation (SemEval 2015) , pages 486–495, Denver, Colorado. Association for Compu- tational Linguistics. Maria Pontiki, Dimitris Galanis, John Pavlopoulos, Har- ris Papageorgiou, Ion Androutsopoulos, and Suresh Manandhar. 2014. Semeval-2014 task 4: Aspect based sentiment analysis. In Proceedings of the 8th... | https://arxiv.org/abs/2505.18703v1 |
Bart-based contrastive and retrospective network for aspect-category-opinion-sentiment quadruple extrac- tion. Int. J. Mach. Learn. Cybern. , 14(9):3243–3255. Lu Xu, Hao Li, Wei Lu, and Lidong Bing. 2020. Position-aware tagging for aspect sentiment triplet extraction. In Proceedings of the 2020 Conference on Empirical ... | https://arxiv.org/abs/2505.18703v1 |
arXiv:2505.18708v1 [cs.CL] 24 May 2025A General Knowledge Injection Framework for ICD Coding Xu Zhang1,2, Kun Zhang1,2*, Wenxin Ma1,2, Rongsheng Wang1,2, Chenxu Wu1,2, Yingtai Li1,2, S. Kevin Zhou1,2,3,4 * 1School of Biomedical Engineering, Division of Life Sciences and Medicine, USTC 2MIRACLE Center, Suzhou Institute ... | https://arxiv.org/abs/2505.18708v1 |
explored the incorporation of ICD code- related knowledge to assist models in precisely lo- cating evidence fragments related to specific ICD codes, thereby effectively and efficiently improv- ing coding performance. Generally, three types of knowledge are involved in ICD coding: ICD Description, ICD Synonym, and ICD H... | https://arxiv.org/abs/2505.18708v1 |
guideline syn- thesis and multi-task learning, to inject ICD code knowldge into the coding model. •We achieve state-of-the-art performance on most evaluation metrics on the ICD coding benchmarks MIMIC-III and MIMIC-III-50, proving not only the effectiveness of our knowledge injection framework, but also the necessity o... | https://arxiv.org/abs/2505.18708v1 |
of ICD code descriptions as prompts in addition to each clinical note as model input. DKEC (Ge et al., 2024) propose a heterogeneous graph network to encode knowledge from multiple sources, and generate knowledge-based queries for each ICD code. MRR (Wang et al., 2024a) and AKIL (Wang et al., 2024b) incorporates diagno... | https://arxiv.org/abs/2505.18708v1 |
guideline, and then use this guideline and multi-task learning to inject knowledge into the model. Note that our method only incorporates knowledge in the training stage, which has no effect on the computation cost of the model during the inference stage. 3.2 Proposed Method 3.2.1 Guideline Synthesis Given a medical te... | https://arxiv.org/abs/2505.18708v1 |
descriptions and hierarchical descriptions, and concatenate them to form a long string sequence ˆx. Thus, for each training sample (x, y), we gen- erate a synthetic guideline ˆxthat encapsulates the relevant knowledge of the ICD codes assigned to the raw training sample. 3.2.2 Multi-task Learning Retrieve and predictio... | https://arxiv.org/abs/2505.18708v1 |
code query of thei-th ICD code, Ai∈RC×Nis the attention matrix from the i-th code to the input text, and Wk, Wv∈RD×Dare the linear transform matrices. Based on the aggregated evidence of the i-th ICD code, a linear classifier is applied to compute the predicted probability for i-th ICD code: ˆyi=sigmoid (EiWi), (19) wh... | https://arxiv.org/abs/2505.18708v1 |
ModelsMIMIC-III-Full MIMIC-III-Top-50 AUC F1 P@K AUC F1P@5Macro Micro Macro Micro P@8 P@15 Macro Micro Macro Micro CAML (Mullenbach et al., 2018) 0.895 0.986 0.088 0.539 0.709 0.561 0.875 0.909 0.532 0.614 0.609 MSATT-KG (Xie et al., 2019) 0.910 0.992 0.090 0.553 0.728 0.581 0.914 0.936 0.638 0.684 0.644 MSMN (Yuan et ... | https://arxiv.org/abs/2505.18708v1 |
knowledge injection on MIMIC-III-Full Dataset. The abbreviations "desc", "syn", "hie" stand for description knowledge, synonym knowledge and hierarchy knowledge, respectively. We apply our proposed knowledge injection training frame- work to the baseline model, and add different types of ICD code knowledge. Different f... | https://arxiv.org/abs/2505.18708v1 |
before and after knowledge injection. Before knowledge injection, only half of the codes are correctly predicted by the model, and the evi- dence of the false positive code "272.0" is totally irrelevant to this code. However, after knowledge injection, the predicted codes are the same as the ground truth. Notably, the ... | https://arxiv.org/abs/2505.18708v1 |
4869–4890, Miami, Florida, USA. Association for Computational Linguistics. Xueren Ge, Abhishek Satpathy, Ronald Dean Williams, John Stankovic, and Homa Alemzadeh. 2024. DKEC: Domain knowledge enhanced multi-label classification for diagnosis prediction. In Proceed- ings of the 2024 Conference on Empirical Methods in Na... | https://arxiv.org/abs/2505.18708v1 |
the North American Chapter of the Association for Computational Lin- guistics: Human Language Technologies, Volume 1 (Long Papers) , pages 1101–1111, New Orleans, Louisiana. Association for Computational Linguis- tics. Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, ... | https://arxiv.org/abs/2505.18708v1 |
Improving Bangla Linguistics: Advanced LSTM, Bi -LSTM, and Seq2Seq Models for Translating Sylheti to Modern Bangla Sourav Kumar Das Md. Julkar Naeen MD. Jahidul Islam Dept. of CSE Dept. of CSE Dept. of CSE Daffodil International University Daffodil International University Daffodil International University Dhaka, Bangl... | https://arxiv.org/abs/2505.18709v1 |
research in the Bangla Local language translation from Local Bangla to Modern Bangla. This study aims to translate Local Bangla language to Modern Bangla for better communication. By doing the research, an app could be developed to do the translation work for better user experience. This study will help the normal peop... | https://arxiv.org/abs/2505.18709v1 |
and F1 -Score of fake news vary in experiments, and the reason was to set 37.47 times less the numbers of fake news than the authentic news. Shafayet Bin Shabbir Mugdha et al. [7] presented a system that can detect news which are fake using full content or accuracy based on the news headline, performing models like Gau... | https://arxiv.org/abs/2505.18709v1 |
Bhowmik et al. [14] proposed a method using deep learning approach with rule, based method (BTSC) for sentiment analysis of Bangla text using LDD which is an extended lexicon data dictionary. Different LSTM models are used in this case an d got the best accuracy of 84.18% from BERT -LSTM. This paper also includes fine ... | https://arxiv.org/abs/2505.18709v1 |
news articles. RNN model obtained 98.33% accuracy. III. Methodology A. Workflow: The translation is from Sylheti Bangla to Modern Bangla language using Natural Language Processing. Fig. 1. shows the total work flow. After the data collection, necessary preprocessing is used for cleaning the dataset. After preprocessing... | https://arxiv.org/abs/2505.18709v1 |
research is to translate from Local Sylheti Bangla to Modern Bangla which is known by all. For the translation, different deep learning models like LSTM, Bi - LSTM, Seq2Seq models. These models are very popular in NLP or Natural la nguage processing research. LSTM (Long Short -Term memory) The model is known as recurre... | https://arxiv.org/abs/2505.18709v1 |
a vector size of 1024 for both past and future predictions. The attention layer contains 1,039 parameters. The input sequence calculates weight when predicting something. O utput layer doesn't change in this process. There are a total of 524,800 parameters. The input data uses linear transformation and makes it into a ... | https://arxiv.org/abs/2505.18709v1 |
layer processes the sequence of input into 12 and 11. The hidden state size for both is 256. Finally, the dense layer is connected with lstm_1 0,0 with 227,445 parameters. This applies linear transformation for the final output and the o utput shape is 885 vectors for each token in the given sentence. The max length is... | https://arxiv.org/abs/2505.18709v1 |
dataset isn't suitable for the model, that's why some models can't perform up to our expecta tions. Fig. 7. Fig. 8. & Fig. 9. Shows the train and validation accuracy plots are shown for 3 different models. For LSTM , the train and validation accuracy difference are very low and approximately after 40 epoch train accura... | https://arxiv.org/abs/2505.18709v1 |
Bangla, in future good and optimized stemmers are made then this type of research might get very popular among the researchers. Reference [1] Siddique, S., Ahmed, T., Talukder, M.R.A. and Uddin, M.M., 2021. English to bangla machine translation using recurrent neural network. arXiv preprint arXiv:2106.07225 . [2] Winat... | https://arxiv.org/abs/2505.18709v1 |
Spoken language identification on 4 Indonesian local languages using deep learning. Bulletin of Electrical Engineering and Informatics , 11(6), pp.3288 -3293. [18] Sen, O. and Roy, P., 2021, September. A convolutional neural network based approach to recognize Bangla spoken digits from speech signal. In 2021 Internatio... | https://arxiv.org/abs/2505.18709v1 |
Neural Parameter Search for Slimmer Fine-Tuned Models and Better Transfer Guodong Du1Zitao Fang2Jing Li1/envel⌢peJunlin Li1Runhua Jiang2 Shuyang Yu2Yifei Guo2Yangneng Chen1Sim Kuan Goh2 Ho-Kin Tang1Daojing He1Honghai Liu1Min Zhang1 1Harbin Institute of Technology, Shenzhen, China 2Xiamen University Malaysia duguodong7@... | https://arxiv.org/abs/2505.18713v1 |
2023b), as shown in Figure 1. Fine-tuned models often exhibit significant re- dundancy in parameter modifications compared to pre-trained models. Pruning these models can en-arXiv:2505.18713v1 [cs.LG] 24 May 2025 0.0 0.5 1.0 1.5 2.0 weight of parameter subspace a0.00.51.01.52.0weight of parameter subspace bOptimum Pret... | https://arxiv.org/abs/2505.18713v1 |
versatile method suitable for multi-task model fusion, compression, and robust knowledge transfer. •Experimental results shown that our method significantly improves performance in various knowledge transfer scenarios. 2 Related Work 2.1 Knowledge Transfer, Fusion and Compression In the realms of knowledge transfer, mo... | https://arxiv.org/abs/2505.18713v1 |
Here, we consider knowledge transfer, fusion and compression of a set of tasks {T1, . . . , T n}and var- ious pre-trained models like ViT (Dosovitskiy et al., 2021), T5 (Raffel et al., 2020), or Llama2 (Touvron et al., 2023). To begin, each pre-trained model is optimized on task-specific data, which can be per- formed ... | https://arxiv.org/abs/2505.18713v1 |
This mask is derived using the following equation: md=( 1,ifτd≥sorted (τ)[r×d] 0,otherwise(4) The final pruned fine-tuned model is then given by: ˆθft=θpre+m⊙τ (5) while⊙represents the Hadamard product. This final model can subsequently be applied to scenarios such as knowledge transfer, fusion, and 0.02 0.06 0.10 0.15... | https://arxiv.org/abs/2505.18713v1 |
storing only the masked values, we can preserve full performance while greatly reduc- ing storage. In term of storage for {θt}T t=1, we only need to store the pre-trained model θpre, the task vec- torsτ, and the binary masks mfor each task. For multi-task evaluation, fine-tuned models can be re- constructed by adding o... | https://arxiv.org/abs/2505.18713v1 |
80.35 64.52 55.49 58.2 61.56 60.03 Model Tailor [ICML24] 273M 73.21 52.49 42.28 67.15 43.89 82.88 63.40 56.15 75.4 61.87 66.94 NPS (ours) 273M 74.3 52.52 43.1 66.12 43.93 83.23 64.52 57.51 76.2 62.38 67.54 Method #ParamsPre-trained tasks Target task VQAv2 GQA VizWiz SQA TextVQA POPE MM-Bench MM-Bench-CN OKVQA Avg Hscor... | https://arxiv.org/abs/2505.18713v1 |
improves aver- age performance by 1.4% across 11 tasks compared to the top baseline. Merging LLMs. In our experiment, we com- bined three specialized large language models built on the Llama-2-7b architecture (Touvron et al., 2023), each focusing on a different area: Chinese language proficiency1, mathematical reasonin... | https://arxiv.org/abs/2505.18713v1 |
merging techniques. Additionally, we conducted experiments with multiple random seeds and reported the average results across five seeds. As detailed in Table 2, our approach sur- passes the best baseline by 1.3% on Roberta-base and 0.9% on T5-base. 4.5 Results on Knowledge Compression We conducted experiments using ei... | https://arxiv.org/abs/2505.18713v1 |
•Gradient-Free Operation :NPS method op- erates without gradient calculations, making it lightweight and minimizing memory usage. •Practicality and Ease of Implementation : The method is straightforward to implement and integrates easily into various applications. •Broader Applicability and Stable Perfor- mance : Unlik... | https://arxiv.org/abs/2505.18713v1 |
for better pretraining. arXiv preprint arXiv:2204.03044 . Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. 2014. De- scribing textures in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition (CVPR) . Karl Cobbe, Vineet Kosaraju, Mohammad Bav... | https://arxiv.org/abs/2505.18713v1 |
of the IEEE conference on computer vision and pattern recognition (CVPR) , pages 6904–6913. Danna Gurari, Qing Li, Abigale J Stangl, Anhong Guo, Chi Lin, Kristen Grauman, Jiebo Luo, and Jeffrey P Bigham. 2018. Vizwiz grand challenge: Answering visual questions from blind people. In Proceedings of the IEEE conference on... | https://arxiv.org/abs/2505.18713v1 |
Fei. 2013. 3d object representations for fine-grained categorization. In Proceedings of the IEEE Interna- tional Conference on Computer Vision Workshops (ICCVW) , pages 554–561. Yann LeCun. 1998. The mnist database of handwrit- ten digits. http://yann.lecun.com/exdb/ mnist/ . Hector Levesque, Ernest Davis, and Leora Mo... | https://arxiv.org/abs/2505.18713v1 |
Bissacco, Baolin Wu, Andrew Y Ng, et al. 2011. Reading digits in natural images with unsupervised feature learning. In NIPS workshop on deep learning and unsupervised feature learning , page 7. Yixin Nie, Adina Williams, Emily Dinan, Mohit Bansal, Jason Weston, and Douwe Kiela. 2020. Adversarial NLI: A new benchmark fo... | https://arxiv.org/abs/2505.18713v1 |
read. In Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition (CVPR) , pages 8317–8326. Johannes Stallkamp, Marc Schlipsing, Jan Salmen, and Christian Igel. 2011. The german traffic sign recog- nition benchmark: a multi-class classification com- petition. In Proceedings of the Internation... | https://arxiv.org/abs/2505.18713v1 |
Yu, Haiyang Yu, Fei Huang, and Yongbin Li. 2023a. Language models are super mario: Ab- sorbing abilities from homologous models as a free lunch. arXiv preprint arXiv:2311.03099 . Longhui Yu, Weisen Jiang, Han Shi, Jincheng Yu, Zhengying Liu, Yu Zhang, James T. Kwok, Zhen- guo Li, Adrian Weller, and Weiyang Liu. 2023b. ... | https://arxiv.org/abs/2505.18713v1 |
be more robust, deliv- ering near-optimal results across all tasks.B Additional Analysis B.1 Ablation Studies Our method incorporates several key factors, in- cluding the number of subspaces, the volume of the calibration dataset, and the sparsity of pruning levels. We conducted ablation studies on these ele- ments, wi... | https://arxiv.org/abs/2505.18713v1 |
ask ArithmeticTIES-Merging Fisher MergingConsensus Ties NPSPruning(ours) Figure 8: Test set performance when merging ViT-B/32 and ViT-L/14 models on eight image classification tasks. Table 4: Test set performance when merging T5-base models on seven NLP tasks. Please refer to Section 4.4 for more details. Task(→) Test ... | https://arxiv.org/abs/2505.18713v1 |
66.8 78.6 87.5 66.6 51.3 51.5 81.7 43.2 90.9 57.6 67.0 58.4 Consensus Ties [ICML24] 66.6 78.5 87.3 66.4 51.1 51.2 81.6 43.4 90.2 57.3 67.1 58.3 NPS (ours) 68.2 80.1 83.5 67.3 51.2 49.8 88.4 42.6 92.8 61.9 67.5 64.8 Table 7: Test set performance when merging ViT-B/32 models on 8 vision tasks. Please refer to Section 4.4... | https://arxiv.org/abs/2505.18713v1 |
with an initial learning rate of 1e-5 and the T5-base model at 1e-4, using the AdamW optimizer. The learning rate was gradually increased during the first 6% of training steps, then linearly decreased to zero. Both models were trained with a batch size of 16 over 30 epochs for emotion classification, with performance e... | https://arxiv.org/abs/2505.18713v1 |
358 mins 300 mins 278 mins 183 mins 273 mins D Baseline details We provied a detailed baseline description. Our ex- periments encompass seven comparison methods: •Individual means that each task uses an in- dependent fine-tuned model, which has no in- terference between tasks, but cannot perform multiple tasks simultan... | https://arxiv.org/abs/2505.18713v1 |
(Levesque et al., 2012) and Winogrande (Sakaguchi et al., 2021)), and word sense disambiguation (WiC (Pile- hvar and Camacho-Collados, 2019)). Vision Tasks. Following Task Arithmetic (Il- harco et al., 2023a), we study multi-task model merging on eight image classification datasets be- low. Stanford Cars (Krause et al.... | https://arxiv.org/abs/2505.18713v1 |
Optimal Transport-Based Token Weighting scheme for Enhanced Preference Optimization Meng Li1*, Guangda Huzhang2, Haibo Zhang2, Xiting Wang1†‡, Anxiang Zeng2† 1Gaoling School of Artificial Intelligence, Renmin University of China, 2LLM Team, Shopee Pte. Ltd. mengli.24@ruc.edu.cn, guangda.huzhang@shopee.com, peter.wu@sho... | https://arxiv.org/abs/2505.18720v1 |
They apply a heuristic weight- ing scheme to reduce the difference in total token weight. Moreover, they can not distinguish the important tokens relevant to instruction-following due to a lack of supervision signal. Recent work like APO (Dao, 2024) has attempted to address this issue by rewriting the irrelevant parts ... | https://arxiv.org/abs/2505.18720v1 |
as the multiplicative product of each token’s probability πθ(y|x) =Q|y| i=1πθ(yi|x, y<i). The reparameter- ized reward difference term ∆rin DPO treats the entire response as a single action, in contrast to classical RLHF methods that model each token as an action and optimize token-level value functions with sparse rew... | https://arxiv.org/abs/2505.18720v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.