text string | source string |
|---|---|
model achieved 67% accuracy (weighted F1: 0.65). For Issue Clarity (84 high-score vs. 329 low-score examples), it achieved 79% accuracy (weighted F1: 0.76). A more detailed prediction quality analysis, including precision and recall per class, can be found in Appendix F, Table 4. The LLM-generated labels for Issue Clar... | https://arxiv.org/abs/2505.20411v1 |
for independent verification and can potentially lead to inconsistencies or misleading reporting practices such as reporting pass@N as pass@1 or implicitly using information derived from final tests. The reliance on closed-source frameworks for many submissions further reduces the transparency and reproducibility of th... | https://arxiv.org/abs/2505.20411v1 |
our ReAct scaffolding. 3.3 Result analysis We leverage the decontaminated nature of SWE-rebench to analyze performance trends over time and identify potential signs of contamination effects in prior benchmarks. Specifically, we evaluate models on two distinct temporal subsets of tasks: those created in January 2025 and... | https://arxiv.org/abs/2505.20411v1 |
explicit reasoning mechanisms •DeepSeek models demonstrate the strongest performance among open-source models across both SWE-rebench subsets and the SWE-bench Verified benchmark. Notably, both the December and March releases of DeepSeek-V3 consistently outperform other open models in resolution rate and pass@5, highli... | https://arxiv.org/abs/2505.20411v1 |
static benchmarks and thereby facilitate more rapid and open progress in the field of AI for software engineering. Our future work will concentrate on several key areas: •Expanding data coverage and volume: We aim to significantly increase the dataset volume by extending our collection methodology from tasks strictly t... | https://arxiv.org/abs/2505.20411v1 |
https://arxiv.org/abs/2503.23803 . Yuxiang Wei, Olivier Duchenne, Jade Copet, Quentin Carbonneaux, Lingming Zhang, Daniel Fried, Gabriel Synnaeve, Rishabh Singh, and Sida I. Wang. Swe-rl: Advancing llm reasoning via reinforcement learning on open software evolution, 2025. URL https://arxiv.org/abs/2502. 18449 . Alexand... | https://arxiv.org/abs/2505.20411v1 |
https://tracto.ai/ . Ilya Grigorik. Gh archive, 2011. URL https://www.gharchive.org/ . Accessed 15 May 2025. Qwen, :, An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, Huan Lin, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jingr... | https://arxiv.org/abs/2505.20411v1 |
with evol-instruct, 2023. URL https://arxiv.org/abs/2306.08568 . Jiawei Liu, Thanh Nguyen, Mingyue Shang, Hantian Ding, Xiaopeng Li, Yu Yu, Varun Kumar, and Zijian Wang. Learning code preference via synthetic evolution, 2024. URL https://arxiv. org/abs/2410.03837 . DeepSeek-AI. Deepseek-r1: Incentivizing reasoning capa... | https://arxiv.org/abs/2505.20411v1 |
Yu, Hongli Yu, Song Yu, Menghan Yu, Huan Yu, Siyu Yuan, Jun Yuan, Yutao Zeng, Tianyang Zhan, Zheng Zhang, Yun Zhang, Mofan Zhang, Wang Zhang, Ru Zhang, Zhi Zhang, Tianqi Zhang, Xinyi Zhang, Zhexi Zhang, Sijun Zhang, Wenqiang Zhang, Xiangxiang Zhang, Yongtao Zhang, Yuyu Zhang, Ge Zhang, He Zhang, Yue Zhang, Renjie Zheng... | https://arxiv.org/abs/2505.20411v1 |
However, leveraging RL effectively in this setting demands gathering a large collection of real-world, interactive SWE tasks which is challenging, as each task requires the construction of a suitable and reliable execution environment. Some works attempt to gather such tasks by artificially injecting bugs into code [Ya... | https://arxiv.org/abs/2505.20411v1 |
* Testing setup Think step by step: * Identify README files, as they often contain installation instructions. * Look for setup.py, pyproject.toml, requirements.txt, environment.yml. * Consider files in directories like \texttt{docs} that might contain installation guides. * Look for test configuration files that might ... | https://arxiv.org/abs/2505.20411v1 |
placeholders like {test_name}. * If a Makefile runs tests, extract the actual test command (e.g., pytest) . You must ensure the final JSON includes required fields (, install, test_cmd), and optionally packages, pre_install, reqs_path, env_yml_path, pip_packages if relevant. Base your reasoning on all provided files an... | https://arxiv.org/abs/2505.20411v1 |
root causes of the errors in the logs. * Modify the installation configuration to address these issues. * You might need to: - Add missing dependencies - Fix command syntax errors - Change installation order - Add environment variables - Modify test commands First, provide brief reasoning (<100 words) about what 's cau... | https://arxiv.org/abs/2505.20411v1 |
':'tarohi24__typedflow-68 ', 'repo ':'tarohi24/typedflow ', 'base_commit ':'2127e74314d2b97d596cfc12ed8fb257bb688d6f ', 'version ':'1.0', 'created_at ':'2019-12-10 15:26:34 ', 'problem_statement ': "The new syntax doesn 't work\nIt doesn 't accept args in the correct way. For instance, life of cache tables are never in... | https://arxiv.org/abs/2505.20411v1 |
macro avg 0.66 0.71 0.64 0.75 0.78 0.76 413 weighted avg 0.78 0.68 0.70 0.83 0.81 0.82 413 Issue low-score 0.83 0.94 0.88 0.82 0.94 0.88 329 Clarity high-score 0.51 0.26 0.35 0.47 0.20 0.28 84 accuracy 0.80 0.79 413 macro avg 0.67 0.60 0.61 0.65 0.57 0.58 413 weighted avg 0.77 0.80 0.77 0.75 0.79 0.76 413 •Patch genera... | https://arxiv.org/abs/2505.20411v1 |
number of words in the solution patch must not exceed 500. •Problem statement length: The problem statement (github issue description) must contain between 16 and 1000 words (inclusive). •Problem statement language: The problem statement must be in English. •Recency: The github issue the task is based on must have been... | https://arxiv.org/abs/2505.20411v1 |
last open file will be reopened. If line_number is provided, the current window will move to show that line. Usage: replace [OPTIONS] SEARCH REPLACE Replaces a given string with another string in the currently open file. Options: --replace-all Replace all occurrences of the SEARCH text. Usage: scroll_down [OPTIONS] Scr... | https://arxiv.org/abs/2505.20411v1 |
put it into your reasoning. 25 * If the issue description reports some error, create a script to reproduce the error and run it to confirm the error. THIS IS USUALLY A VERY GOOD FIRST STEP! * Edit the source code of the repo to resolve the issue * Rerun your reproduce script and confirm that the error is fixed! THIS IS... | https://arxiv.org/abs/2505.20411v1 |
/project, current file: none) bash-$ </environment> <agent> Let me analyze the issue and fix the prime number generator script step by step. First, let 's create a Python script to reproduce the reported issue and verify the current behavior. ```command create prime_numbers.py ``` </agent> <environment> [File: prime_nu... | https://arxiv.org/abs/2505.20411v1 |
17, 19, 23, 29] (Current directory: /project, current file: none) bash-$ </environment> <agent> The output now correctly shows the first 10 prime numbers. Let 's go through the checklist: 1. Reproduced the error: Yes, we saw the script incorrectly included non- prime numbers in its output. 2. Confirmed the fix resolves... | https://arxiv.org/abs/2505.20411v1 |
repository so that the requirements specified in the <pr_description> are met? I 've already taken care of all changes to any of the test files described in the <pr_description>. This means you DON'T have to modify the testing logic or any of the tests in any way! Your task is to make the minimal changes to non-tests f... | https://arxiv.org/abs/2505.20411v1 |
Current file: /scikit- rf/skrf/network.py) bash-$ Agent Output There is a skrf directory which is likely where the main package code lives, but there is no obvious tests directory here. Let 's look inside the skrf directory to see if the tests are located there. ls -l skrf Environment Observation total 812 ... (Current... | https://arxiv.org/abs/2505.20411v1 |
again: edit 183:183 << 'EOF' def get_adapted_type(self, coltype: TypeEngine) -> TypeEngine: if isinstance(coltype, TypeDecorator) and not hasattr(coltype.__class__, 'impl '): log.warning(f 'TypeDecorator implementation { coltype.__class__.__name__} lacks a class-level variable \ 'impl\ '. Setting to default type TEXT. ... | https://arxiv.org/abs/2505.20411v1 |
the issue. repo str The repository owner/name identifier from GitHub. base_commit str The commit hash representing the repository’s HEAD before the solution PR was applied. hints_text str Comments made on the issue before the creation of the solution PR’s first commit. created_at str The creation timestamp of the pull ... | https://arxiv.org/abs/2505.20411v1 |
91 173 412.25 Files Edited 3.46 2 4 10 Lines Edited 142.17 37 112 500 Fail-to-Pass Tests 14.56 2 5 37 Pass-to-Pass Tests 85.81 22 64 351 Total Tests 105.43 31 82.25 428 Difficulty Score 1.13 1 2 2 Issue Text Score 1.04 1 2 3 Test Score 1.38 2 2 3 On average, task instances in SWE-rebench feature issue descriptions of a... | https://arxiv.org/abs/2505.20411v1 |
arXiv:2505.20415v1 [cs.CL] 26 May 2025Enhancing Logical Reasoning in Language Models via Symbolically-Guided Monte Carlo Process Supervision Xingwei Tan1, Marco Valentino1,3, Mahmud Akhter2, Maria Liakata2,4, Nikolaos Aletras1 1University of Sheffield, UK 2Queen Mary University of London, UK 3Idiap Research Institute, ... | https://arxiv.org/abs/2505.20415v1 |
found that, while appearing plausible at the surface, LLM rea- soning still possesses several limitations. When multiple valid deductive steps are available, for in- stance, LLMs are incapable of planning to system- atically explore different possibilities (Saparov and He, 2023). To address this limitation, prompting s... | https://arxiv.org/abs/2505.20415v1 |
2022) and DPO (Rafailov et al., 2023a) without additional human supervision to derive models that can perform logical reasoning by leveraging explicit symbolic representations. Extensive experiments on logical reasoning benchmarks such as FOLIO (Han et al., 2024) and LogicAsker (Wan et al., 2024) demonstrate the effect... | https://arxiv.org/abs/2505.20415v1 |
Symbolic Reasoning Trajectories (2.2) Filter Symbolic Reasoning Trajectories(2.3) Fine-TuningSymbolically Guided LLMSymbolic React Prompt ... Action: Apply the law of quantifier negation on the third formula Observation: ∀x ¬(Jokes(x) → Unaware(x)) ...Figure 2: Overall pipeline for training a symbolically-guided LLM wi... | https://arxiv.org/abs/2505.20415v1 |
solver which can have a higher flexibility than a symbolic solver. This work offers improvements in both standard logical reasoning benchmarks and in real-world problems. 3 Symbolically-Guided Monte Carlo Process Supervision Since LLMs are mostly trained on natural language data, without explicit guidance, they tend to... | https://arxiv.org/abs/2505.20415v1 |
logical schemes across different problems. To this end, we define a process reward model to provide detailed step-wise feedback over the generated reasoning trajectories. A process reward model is usually trained using the following loss function: LPRM =−NX i=1ysilogrsi+ (1−ysi)log(1−rsi),(1) where ysiis the golden lab... | https://arxiv.org/abs/2505.20415v1 |
and the development set of FOLIOv2 as the development set for our experiment.LogicAsker encompasses a comprehensive set of atomic logical rules, integrating them to con- struct queries that require long inference chains. It requires the model to determine whether the target statement is true or false given the context.... | https://arxiv.org/abs/2505.20415v1 |
consist of partially formalized ex- pressions which only appear in classroom instead of the real-world. Therefore, we include more re- alistic datasets to investigate whether our methods offer out-of-domain generalisability. We use claim verification datasets, where a model needs to deter- mine whether a statement/clai... | https://arxiv.org/abs/2505.20415v1 |
.00 process-DPO (Jiao et al., 2024) + ReAct 53.92 55 .67 Sym. Trajectory-FT (Ours) 63.24 63.67 Sym. Trajectory-FT-DPO (Ours) 61.27 60 .67 Sym. Trajectory-FT + ReAct (Ours) 62.75 60 .67 Sym. Trajectory-FT-DPO + ReAct (Ours) 60.29 61 .33Qwen2.5-7B-InstructReAct (Yao et al., 2023b) 57.84 69.56 SymbCoT (Xu et al., 2024) 53... | https://arxiv.org/abs/2505.20415v1 |
2 presents the out-of-domain results of all methods. We directly employ the model check- points selected for FOLIO and LogicAsker testing, reporting the higher accuracy of the two. Models fine-tuned using our proposed approach demon- strate increased accuracy compared to process- DPO in most instances, with the excepti... | https://arxiv.org/abs/2505.20415v1 |
fewer generated tokens. Overall, these examples demonstrate how symbolic reasoning tra- jectories can support higher levels of abstraction, leading empirically to the observed improvement in generalization. Reasoning steps: 1. Let’s analyze the premises one by one: Premise 1: For all x20, if x20makes tea, then x20is pl... | https://arxiv.org/abs/2505.20415v1 |
obtained by the research teams. Acknowledgments We acknowledge IT Services at The University of Sheffield, Oxford Advanced Research Computing, and Bristol Centre for Supercomputing for the pro- vision of services for High Performance Comput- ing. References DeepSeek-AI. 2025. Deepseek-v3 technical report. Preprint , ar... | https://arxiv.org/abs/2505.20415v1 |
Learning Representations . Liangchen Luo, Yinxiao Liu, Rosanne Liu, Samrat Phatale, Meiqi Guo, Harsh Lara, Yunxuan Li, Lei Shu, Yun Zhu, Lei Meng, Jiao Sun, and Abhinav Rastogi. 2024. Improve mathematical reasoning in language models by automated process supervision. Preprint , arXiv:2406.06592. Qing Lyu, Shreya Havald... | https://arxiv.org/abs/2505.20415v1 |
cot or not to cot? chain-of- thought helps mainly on math and symbolic reason- ing. In The Thirteenth International Conference on Learning Representations . Aarohi Srivastava et al. 2023. Beyond the imitation game: Quantifying and extrapolating the capabili- ties of language models. Transactions on Machine Learning Res... | https://arxiv.org/abs/2505.20415v1 |
Hsu. 2024. Faithful logical rea- soning via symbolic chain-of-thought. In Proceed- ings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Pa- pers) , pages 13326–13365, Bangkok, Thailand. As- sociation for Computational Linguistics. An Yang, Baosong Yang, Beichen Zhang, Binyuan... | https://arxiv.org/abs/2505.20415v1 |
5 shows the comparison between vanilla ReAct and symbolic ReAct with larger models. C Prompt Examples In this section, we show an example of the prompt we used for symbolic ReAct. 2https://ollama.com/Model FOLIO LogicAsker llama3.1:8b 53.431.83 52.330.82 qwen2.5:7b 57.841.74 69.560.57 llama3.3:70b 70.261.29 77.330.54 q... | https://arxiv.org/abs/2505.20415v1 |
(no evidence of scien- tific malpractice), then it can be inferred that C (actions are normal and do not threaten the integrity of peer review). Therefore, the statement is: True. Answer: A) True Input to the LLMs Solve a question answering task by having a Thought, then Finish with your answer. Thought can reason abou... | https://arxiv.org/abs/2505.20415v1 |
statements in the context to logic premises Observation: ∀x (Drinks(x) →Dependent(x)) ::: All people who regularly drink coffee are dependent on caffeine. ∀x (Drinks(x) ∨Jokes(x)) ::: People either regularly drink coffee or joke about being addicted to caffeine. ¬∃x (Jokes(x) →Unaware(x)) ::: No one who jokes about bei... | https://arxiv.org/abs/2505.20415v1 |
arXiv:2505.20416v1 [cs.CL] 26 May 2025GraphGen: Enhancing Supervised Fine-Tuning for LLMs with Knowledge-Driven Synthetic Data Generation Zihong Chen1Wanli Jiang1Jinzhe Li1 Zhonghang Yuan1Huanjun Kong1Wanli Ouyang1,3Nanqing Dong1,2* 1Shanghai Artificial Intelligence Laboratory2Shanghai Innovation Institute 3The Chinese... | https://arxiv.org/abs/2505.20416v1 |
diversity, with repetitive sentence templates and similar difficulty levels. This lack of variation can lead to overfitting, reducing the generalization abil- ity of fine-tuned models and, in extreme cases, causing catastrophic forgetting or model col- lapse (Shumailov et al., 2024). Recent efforts have attempted to im... | https://arxiv.org/abs/2505.20416v1 |
large language models to train smaller mod- els (West et al., 2022). Unlike KG-driven methods, LLMs can generate diverse, human-like text with- out reliance on predefined templates (Liang et al., 2023). However, they often suffer from limited con- trollability and hallucination (Ji et al., 2023), lead- ing to factual i... | https://arxiv.org/abs/2505.20416v1 |
scenar- ios: atomic QA, aggregated QA, and multi-hop QA. From a knowledge organization perspective, these scenarios exemplify the most representative knowledge-intensive tasks in the context of closed- book QA. The framework comprises a four-step workflow involving two interdependent LLMs: the Synthesizer Model ( Msynt... | https://arxiv.org/abs/2505.20416v1 |
comprehension assessment. Through binary yes/no questions, we capture precise semantic information for confidence modeling. which measures the gap between the LLM’s current understanding and complete mastery of the knowl- edge point. By assessing the comprehension loss of Mtrain, we can systematically evaluate whether ... | https://arxiv.org/abs/2505.20416v1 |
Metrics We employ a set of natural language metrics (Cao et al., 2024) to evaluate the quality of generated text. Details are provided in Appendix F.3. Since most of these metrics are better suited for evaluating complete sentences than brief responses, we compared the aggregated QAs produced by GraphGen with those fro... | https://arxiv.org/abs/2505.20416v1 |
QA English 73,642 7,405 135 25 1,985 82 Table 1: Description of datasets employed for experiments. For calculating the token count, the tokenizer used is from Qwen2.5 series (Yang et al., 2025). The corpus is employed for graph construction and data synthesis, while the test set is utilized to evaluate the performance ... | https://arxiv.org/abs/2505.20416v1 |
EntiGraph 532,971 47.4 30.1 92.6 93.3 93.1 56.3 28.8 55.2 SELF-QA 561,798 83.4 34.7 91.3 92.8 92.3 59.5 39.3 58.8 GraphGen 54,287 657.9 75.8 87.8 95.7 90.4 85.0 31.8 75.2 Table 2: Comparison results with other data synthesis methods on data quality evaluation metrics. The results indicate that the quality of data gener... | https://arxiv.org/abs/2505.20416v1 |
values in- dicate knowledge blind spots within the model. These high-loss instances often involve long-tail or rare knowledge that the model may struggle with. Figure 6: Comprehension loss of the Trainee Model. The reduction in loss after training highlights the effec- tiveness of data synthesis and the enhanced compre... | https://arxiv.org/abs/2505.20416v1 |
long as a correlation exists, variations in understanding lev- els within the subgraphs do not significantly impact the final outcomes. However, the underlying pat- terns merit further exploration in future research. 6 Conclusion In this paper, we propose GraphGen, an effective KG-based approach to synthetic data gener... | https://arxiv.org/abs/2505.20416v1 |
Proceedings of the 55th Annual Meeting of the Association for Computational Lin- guistics (Volume 1: Long Papers) , pages 1342–1352. Zichu Fei, Xin Zhou, Tao Gui, Qi Zhang, and Xuan- Jing Huang. 2022. Lfkqg: A controlled generation framework with local fine-tuning for question gen- eration over knowledge bases. In Proc... | https://arxiv.org/abs/2505.20416v1 |
Processing , pages 2348–2370. Yuanyuan Liang, Jianing Wang, Hanlun Zhu, Lei Wang, Weining Qian, and Yunshi Lan. 2023. Prompting large language models with chain-of-thought for few- shot knowledge base question generation. In Pro- ceedings of the 2023 Conference on Empirical Meth- ods in Natural Language Processing , pa... | https://arxiv.org/abs/2505.20416v1 |
Zhang, Yoshua Ben- gio, William W. Cohen, Ruslan Salakhutdinov, and Christopher D. Manning. 2018. Hotpotqa: A dataset for diverse, explainable multi-hop question answer- ing. arXiv preprint , arXiv:1809.09600. Zitong Yang, Neil Band, Shuangping Li, Emmanuel Candès, and Tatsunori Hashimoto. 2024. Syn- thetic continued p... | https://arxiv.org/abs/2505.20416v1 |
This is especially useful for those who frequently adjust settings, enhancing efficiency. Figure 7: User interface of GraphGen. “Input Configuration” is utilized to specify the data sources and the target format. “Traverse Strategy” determines the method of graph organization. “Model Configuration” is employed to set p... | https://arxiv.org/abs/2505.20416v1 |
evaluation settings. Table 5 provides the hyperparameters employed during training, while Table 4 outlines the parameters used in our evaluation pipeline. The time required for processing varies with the size of the dataset and changes in the graph organization strategy. On average, generating a batch of approximately ... | https://arxiv.org/abs/2505.20416v1 |
pairs generated for each text segment. SELF-QA involves two critical steps: first, it generates ten questions based on the original text, and then it answers these questions contextually, yielding a total of ten QA pairs. EntiGraph begins by extracting entities from the text, then combines these entities in pairs or tr... | https://arxiv.org/abs/2505.20416v1 |
Meta-Llama-3.1-8B-Instruct and MiniCPM3-4B, as trainee models. The results, illustrated in Figure 18, are consistent with our primary findings. Specifically, our method continues to deliver stable and significant performance improvements across three knowledge-intensive datasets. These additional experiments reinforce ... | https://arxiv.org/abs/2505.20416v1 |
The reward model inference score of QA pairs. Reward models in this pa- per include BAAI/IndustryCorpus2_DataRater1andOpenAssistant/reward- model-deberta-v3-large-v22. 1https://huggingface.co/BAAI/IndustryCorpus2_DataRater 2https://huggingface.co/OpenAssistant/reward-model-deberta-v3-large-v2 Table 7: Key metrics for e... | https://arxiv.org/abs/2505.20416v1 |
scores. Metric GraphGen EntiGraph Genie LongForm SELF-QA WRAP General Benchmarks GPQA 33.84 27.78 29.80 30.81 34.85 33.84 CMLU 77.61 77.56 78.58 77.42 77.54 77.89 GSM8K 80.89 79.45 80.44 80.74 80.21 81.20 BBH 68.51 67.72 67.80 67.57 66.92 66.57 MATH 52.45 52.14 53.73 52.91 52.99 54.71 Lukaemon 73.56 71.11 73.61 72.77 7... | https://arxiv.org/abs/2505.20416v1 |
arXiv:2505.20422v1 [cs.CL] 26 May 2025SEMMA : A Semantic Aware Knowledge Graph Foundation Model Arvindh Arun◦1, Sumit Kumar◦2, Mojtaba Nayyeri1, Bo Xiong3 Ponnurangam Kumaraguru2, Antonio Vergari4, Steffen Staab15 1Institute for AI, University of Stuttgart,2IIIT Hyderabad, 3Stanford University,4University of Edinburgh,... | https://arxiv.org/abs/2505.20422v1 |
symbols and their formal meaning. Though related work has used word embeddings of knowledge graph symbols and other attributed texts (e.g. comments) when embedding individ- ual knowledge graphs (Nayyeri et al., 2023; Yuan et al., 2025; Yao et al., 2019; Wei et al., 2023; Xu et al., 2021), KGFMs have so far neglected th... | https://arxiv.org/abs/2505.20422v1 |
head or tail entity masked, denoted as (h, r,?)and (?, r, t). We assume that for every relation, de- noted by r, there is an inverse relation, denoted byr−1. Let GTRAIN = (ETRAIN,RTRAIN,FTRAIN) correspond to the training graph and GTEST = (ETEST,RTEST,FTEST)denote the graph during in- ference, where GTRAIN∪G TEST=GO. T... | https://arxiv.org/abs/2505.20422v1 |
Kingdom” . All of the 57 datasets from the benchmark for this task (see Section 5.1) attribute textual information. While some real- world TAKGs, like Wikidata, exhibit even more complex graph structures (e.g., involving time) and text attributes (e.g., multi-lingual), our definitions allow for a sufficiently expressiv... | https://arxiv.org/abs/2505.20422v1 |
the original graph GO to perform the final link prediction. By learning relative structural patterns via GSTR Rand condition- ing representations on the query, ULTRA avoids learning fixed relation embeddings, enabling its zero-shot transfer capabilities. TRIX(Zhang et al., 2025b) and MOTIF (Huang et al., 2025) are foll... | https://arxiv.org/abs/2505.20422v1 |
2), which explicitly separates structural processing (as inULTRA ) from a module dedicated to processing relation’s textual semantics derived from LLMs in parallel. This dual-module approach provides mod- ularity, allowing the text processing module to be deactivated when the textual attributions are noisy or unhelpful... | https://arxiv.org/abs/2505.20422v1 |
we require a method to derive their semantic embeddings. In- tuitively, if the original relation embedding signi- fies a transformation in a certain conceptual direc- tion in the embedding space, its inverse should logically point the other way (Sun et al., 2019). So, given a relation’s embedding τr, we generate the in... | https://arxiv.org/abs/2505.20422v1 |
NBFNet for final prediction. We use the same training objective as ULTRA for the whole pipeline. 5 Experiments Through our experiments, we address the follow- ing questions: RQ1: Where does textual semantics matter? RQ2: Does adding textual semantics in- crease average performance? RQ3: Can textual semantics help gener... | https://arxiv.org/abs/2505.20422v1 |
this insight, we introduce the SEMMA HYBRID variant, where we switch off the text processing modulebased on the validation set’s performance. 5.3 Does adding textual semantics increase average performance? From Table 1, we can see that SEMMA clearly out- performs ULTRA averaged across all 54 datasets. MRR has an increa... | https://arxiv.org/abs/2505.20422v1 |
all such queries, whereas GTEXT R will be able to distinguish them based on the relation identifier. For a working example, refer to Fig. 8. As observed in Table 4, ULTRA ’s performance drops substantially in the new setting where S EMMA is 2x better. Takeaway 2. The limitation of purely struc- tural approaches like UL... | https://arxiv.org/abs/2505.20422v1 |
versus tex- tual signals, potentially improving upon the current SEMMA HYBRID ’s validation-dependent switch, especially when relation text quality varies signifi- cantly. Finally, SEMMA only explores ontological concepts that can be modeled with textual semantic similarity; there is scope for future work to broaden th... | https://arxiv.org/abs/2505.20422v1 |
2016. Are names meaningful? quantifying social meaning on the se- mantic web. In The Semantic Web–ISWC 2016: 15th International Semantic Web Conference, Kobe, Japan, October 17–21, 2016, Proceedings, Part I 15 , pages 184–199. Springer. 9 DeepSeek-AI, Aixin Liu, Bei Feng, et al. 2025. Deepseek-v3 technical report. Prep... | https://arxiv.org/abs/2505.20422v1 |
Gao, and Heyan Huang. 2024a. Fundamen- tal capabilities of large language models and their applications in domain scenarios: A survey. In Pro- ceedings of the 62nd Annual Meeting of the Associa- tion for Computational Linguistics (Volume 1: Long Papers) , pages 11116–11141, Bangkok, Thailand. Association for Computatio... | https://arxiv.org/abs/2505.20422v1 |
Jay Revolinsky, and Jiliang Tang. 2024. Towards better benchmark datasets for in- ductive knowledge graph completion. Preprint , arXiv:2406.11898. Saba Sturua, Isabelle Mohr, Mohammad Kalim Akram, Michael Günther, Bo Wang, Markus Krimmel, Feng Wang, Georgios Mastrapas, Andreas Koukounas, Nan Wang, and Han Xiao. 2024. j... | https://arxiv.org/abs/2505.20422v1 |
using graph neural networks for multi-node representation learning. In Proceedings of the 35th International Conference on Neural Information Processing Sys- tems, NIPS ’21, Red Hook, NY , USA. Curran Asso- ciates Inc. Yucheng Zhang, Beatrice Bevilacqua, Mikhail Galkin, and Bruno Ribeiro. 2025b. Trix: A more expressive... | https://arxiv.org/abs/2505.20422v1 |
leakage arising from 1https://openrouter.ai/ 12 Sample Output "cleaned_relations": { "Causes": "Causes", "/organization/organization/headquarters./location/mailing_address/citytown": "Headquarters City", "GpMF": "Gene participates Molecular Function" ... } "relation_descriptions": { "Causes": ["leads to effect", "effec... | https://arxiv.org/abs/2505.20422v1 |
Koutra (2020) 2034 42 32888 1827 1828 h/t WDsinger Lv et al. (2020) 10282 135 16142 2163 2203 h/t FB15k237_10 Lv et al. (2020) 11512 237 27211 15624 18150 tails FB15k237_20 Lv et al. (2020) 13166 237 54423 16963 19776 tails FB15k237_50 Lv et al. (2020) 14149 237 136057 17449 20324 tails FB15k237 Toutanova and Chen (201... | https://arxiv.org/abs/2505.20422v1 |
to effectively identify important semantic parallels, thereby bridging linguistic vari- ations and compensating for less formal relational definitions. The usefulness of textual semantic similarity is also dependent upon the threshold used. A high threshold predominantly captures high-certainty similarities (e.g., exac... | https://arxiv.org/abs/2505.20422v1 |
the relation vocabulary present in the test graph. Formally, given GTESTand queries Q=FT\ F O, we define a new evaluation setting: RQ∩ RGTEST=∅ where RQdenotes the set of relations used in queries, and RGTESTdenotes the set of relations present in the test graph. This scenario is moti- vated by the practical case of te... | https://arxiv.org/abs/2505.20422v1 |
7 14371 10000 7 14371 1596 10000 7 14110 1566 MT2 org 10000 10 23233 10000 10 23233 2581 10000 11 21976 2441 MT2 sci 10000 16 16471 10000 16 16471 1830 10000 16 14852 1650 MT3 art 10000 45 27262 10000 45 27262 3026 10000 45 28023 3113 MT3 infra 10000 24 21990 10000 24 21990 2443 10000 27 21646 2405 MT4 sci 10000 42 125... | https://arxiv.org/abs/2505.20422v1 |
GNN ENT# layers 6 hidden dim 64 message DistMult aggregation sum g(·) 2-layer MLP Learningoptimizer AdamW learning rate 0.0005 training steps 200,000 adv temperature 1 # negatives 128 batch size 64 Training graph mixture FB15k237, WN18RR, CoDExMedium Table 9: SEMMA hyperparameters for pre-training. GNN STRcorresponds t... | https://arxiv.org/abs/2505.20422v1 |
words as possible while clearly conveying the core meaning. Avoid filler words, unnecessary adjectives, and overly generic language. Emphasize Key Semantics: Focus on the distinctive action or relationship the relation name implies. Ensure that the description highlights the unique aspects that differentiate it from si... | https://arxiv.org/abs/2505.20422v1 |
specific dataset. 21 DatasetULTRA SEMMA MRR H@10 MRR H@10 Inductive (e, r) datasets FBIngram:25 0.383 0.636 0.400 0.642 FBIngram:50 0.332 0.536 0.344 0.546 FBIngram:75 0.395 0.598 0.404 0.600 FBIngram:100 0.436 0.634 0.445 0.635 WKIngram:25 0.288 0.481 0.303 0.509 WKIngram:50 0.152 0.304 0.174 0.318 WKIngram:75 0.372 0... | https://arxiv.org/abs/2505.20422v1 |
arXiv:2505.20428v1 [cs.CL] 26 May 2025The UD-N EWS CRAWL Treebank: Reflections and Challenges from a Large-scale Tagalog Syntactic Annotation Project Angelina A. Aquino2,4*Lester James V . Miranda1*Elsie Marie T. Or3* 1Allen Institute for AI2Charles Darwin University 3Department of Linguistics, University of the Philip... | https://arxiv.org/abs/2505.20428v1 |
from the Leipzig Tagalog NewsCrawl corpus.2The text was annotated by na- tive speakers with domain knowledge in linguistics, and the annotations were verified through manual inspection and semi-automated quality control. Fur- thermore, we introduce transformer-based parsers onUD-N EWSCRAWL which serve as baselines for ... | https://arxiv.org/abs/2505.20428v1 |
2005) which we discuss in the next section. Tagalog has a non-configurational phrase struc- ture (Kroeger, 1993) wherein sentences are canoni- cally predicate-initial and sentence arguments fol- lowing the predicate have flexible order and can also be fronted using the ayinversion marker. Tagalog lacks a copula, and th... | https://arxiv.org/abs/2505.20428v1 |
argument to be the subject, unlike in Indo-European languages like English (Foley, 2008; Riesberg et al., 2019). In Figure 1 therefore, while the verbs in each sentence exhibit different voice markings and the subject carry different thematic roles, the exam- ples could all be translated as the equivalent of the Englis... | https://arxiv.org/abs/2505.20428v1 |
these can be in- flected with verb affixes, thus they can alternatively be glossed as ’to give’ and ’to surpass or defeat’, respectively. The POS designation of Tagalog roots in UD- NEWSCRAWL was determined by the morphologi- cal structure of the word. Thus, if a root is affixed with a nominal affix, such as pag- in th... | https://arxiv.org/abs/2505.20428v1 |
the sentence tokenization of texts in the corpus and selected a subset of 15,000 sentences. These were then divided into 150 files of 100 sentences each in order of increasing sentence length, with an ap- proximately even distribution of sentence lengths across files. This data preparation process dis- tributed the ann... | https://arxiv.org/abs/2505.20428v1 |
the treebank, we employed both semi-automated and manual post-annotation work- flows. First, we trained a silver-standard parsing model on our existing annotations using spaCy (Honnibal et al., 2020) and identified instances where the model’s morphological annotations and dependency parsing relations disagree with huma... | https://arxiv.org/abs/2505.20428v1 |
in Appendix D. Finally, we evaluated each component in its corresponding linguistic task. We report the accuracy for both lemmatization and POS tagging tasks and the macro F1-score for the morpholog- ical annotation task. In addition, we also report the unlabeled and labeled attachment scores (UAS / LAS) for the depend... | https://arxiv.org/abs/2505.20428v1 |
UD-NewsCrawl36.69% 25.64% 17.51% 13.32% 3.19% 2.83% 0.82%Figure 3: Topic distribution of UD-N EWSCRAWL using categories from SIB-200 (Adelani et al., 2024). previously reported benchmarks on both treebanks. The cross-treebank generalization indicates that de- spite potential differences in domain, annotation style, or ... | https://arxiv.org/abs/2505.20428v1 |
borne and Gerdes, 2019; Kanayama and Iwamoto, 2020). For example, in UD-N EWSCRAWL , the nsubj relation is defined in terms of subject NP semantic roles typical of Indo-European languages. Although UD provides language-specific documen- tation, English (and Euro-western linguistic frame- works) remains the default refe... | https://arxiv.org/abs/2505.20428v1 |
annotation of texts by native Tagalog speakers, who were compensated for their work. We ensured that the annotators were treated fairly and that their contributions were acknowledged. The data used in the corpus was sourced from pub- licly available news articles, and we adhered to ethical guidelines regarding data usa... | https://arxiv.org/abs/2505.20428v1 |
190– 202, Manila, Philippines. Association for Computa- tional Linguistics. Mikel Artetxe, Sebastian Ruder, and Dani Yo- gatama. 2019. On the cross-lingual transferabil- ity of monolingual representations. arXiv preprint arXiv:1910.11856 . Maria Bardají, Elsie Or, Angelina Aquino, and Nikolaus Himmelmann. 2024. The cha... | https://arxiv.org/abs/2505.20428v1 |
Matthew S. Dryer and Martin Haspelmath. 2013. WALS Online (v2020.4). Data set. Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783 . David M. Eberhard,... | https://arxiv.org/abs/2505.20428v1 |
nominalism and its consequences: A tagalog case study. Theoretical Linguistics , 35(1):1–49. Dan Kondratyuk and Milan Straka. 2019. 75 languages, 1 model: Parsing Universal Dependencies univer- sally. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joi... | https://arxiv.org/abs/2505.20428v1 |
tagalog case study. Linguistic Inquiry , 36(4):565–599. Ria P. Rafael. 2016. Sinong pasimuno? paggamit ng subject at topic sa pag-aaral ng wika’ (who’s the initiator? the use of subject and topic in language studies). Daluyan: Journal ng Wikang Filipino , 22(1/2):165–180. Lawrence Reid. 2005. Tagalog and philippine lan... | https://arxiv.org/abs/2505.20428v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.