text string | source string |
|---|---|
easy meal, vegetarians, and those who enjoy a balance of carbohydrates and vegetables in their diet. QA1: What is the target group of people for the pasta and broccoli dish? Vegetarians QA2: What type of meal is the pasta and broccoli dish? Low-calorie Generated Answer: Vegetarians Q:This dish is suitable for which gro... | https://arxiv.org/abs/2505.19354v1 |
There are two bottles of dark liquid, possibly coffee or tea, with a white cup of dark liquid in the image. QA1: What are placed on the table? Bottles QA2: What could be the contents of the bottles? Beer Generated Answer: Beer Q:what is in the bottles? GT: alcohol/liqueur/baileys Cap1: a sandwich on a plate with a glas... | https://arxiv.org/abs/2505.19354v1 |
arXiv:2505.19355v1 [cs.CL] 25 May 2025Preprint. Under review. Estimating Online Influence Needs Causal Modeling! Counterfactual Analysis of Social Media Engagement Lin Tian & Marian-Andrei Rizoiu University of Technology Sydney {Lin.Tian-3,Marian-Andrei.Rizoiu }@uts.edu.au Abstract Understanding true influence in socia... | https://arxiv.org/abs/2505.19355v1 |
public discourse on highly polarizing topics, and who does truly drive misinformation spread. True influence is unobserved, and notoriously difficult to estimate (Ram & Rizoiu, 2024). We therefore train and test our models on observed exogenous attention signals—like search trends, news coverage cycles, or influencer a... | https://arxiv.org/abs/2505.19355v1 |
{ˆe(t0+τobs+k∆t)}K k=1, where ∆tis a fixed time step (e.g., one day), and K=⌊T/∆t⌋is the number of prediction points over a horizon T(e.g., one month). 3 Methodology Our framework (as shown in Fig. 1) models external signals as continuous-intensity treat- ments through modeling that capture both immediate responses and... | https://arxiv.org/abs/2505.19355v1 |
baseline func- tion, g∗ a(t)models dependence on past treatments, g∗ o(t)captures dependence on past engagement outcomes, and g∗ g(t)incorporates Google Trends intensity. Forg∗ g(t), we use discrete-time window sampling at 10-minute intervals ( ∆g=10 minutes): g∗ g(t) =w−1 ∑ k=0αk·g(t−k∆g)·1[t−k∆g,t−(k−1)∆g)(t), Here, ... | https://arxiv.org/abs/2505.19355v1 |
Figure 2: Engagement trajectory and counterfactual scenarios. Blue dashed lines represent observed social media engagement ( λobs). Red lines indicate Google Trends signals ( [πB]). ˜πBas the partial observed Google Trends signals. The vertical dashed line at day 9 marks the start of prediction period, with the gray sh... | https://arxiv.org/abs/2505.19355v1 |
scenario, we estimate the expected engagement outcomes under the transformed signal and calculate the causal effect as the difference between counterfactual and actual outcomes: ∆C=E[Y|GC]−E[Y|G]. 3.4 Model Training and Optimization We train our joint model using a combined loss function that explicitly optimizes both ... | https://arxiv.org/abs/2505.19355v1 |
each post using a temporal lag window τlagto model potential causal influences on engagement. 4.2 Models and Baselines We evaluate eight architectural variants, four Transformer-based and four Mamba-based, each modified to incorporate external signals, with details in Appendix A.3. Transformer-based Variants. We extend... | https://arxiv.org/abs/2505.19355v1 |
– – – – – – – Mamba (Gu & Dao, 2024) 0.189* – – – – – – – – – MBPP (Rizoiu et al., 2022) 0.193 0.326 0.295 0.281 0.225 0.225 0.226 0.329 0.315 0.284 Transformer + Token 0.128 0.214 0.220 0.225 0.210 0.203 0.197 0.238 0.245 0.250 Transformer + Attention 0.122 0.198 0.191 0.183 0.194 0.188 0.182 0.224 0.198 0.194 Transfo... | https://arxiv.org/abs/2505.19355v1 |
predicting engagement metrics (quotes, replies, retweets, favorites). 8 Preprint. Under review. Table 3: Average Treatment Effect (ATE), computed via G-computation (Robins, 1986), quan- tifies causal impacts of interventions over a 7-day horizon with 95% bootstrap confidence intervals ( ±), normalized across four engag... | https://arxiv.org/abs/2505.19355v1 |
6 6 3 2 5 4 9 8 5 2 3 6 5 5 4 2 6 2 6 5 4 7 8 5 3 3 2 9 5 7 8 5 2 7 1 3 5 3 3 4 8 3 8 2 5 8 3 2 5 5 5 5 6 6 10 2 2 2 4 2 5 5 6 11 7 5 1 2 1 5 2 3 4 6 12 14Effect Effect Scores vs. Follower Counts 2468101214 Figure 3: Decile Heatmaps (Spearman ρcorrelation coefficient, Kendall’s Wrank agreement, Concordance Correlation ... | https://arxiv.org/abs/2505.19355v1 |
Xie (2017b) to misinformation contexts, confirming that content responsiveness to external promotion depends on baseline popular- ity and temporal dynamics across three climate change narratives: “Climate change isn’t real” (85.2th percentile), “Climate change is a UN hoax” (55.6th percentile), and “Changes in earth’s ... | https://arxiv.org/abs/2505.19355v1 |
sources of misinformation and improving platform governance, they could potentially be misused to optimize manipulation campaigns or target influential users for spreading harmful content. We acknowledge these dual-use concerns and propose safeguards, such as integrating with existing content moderation frameworks to p... | https://arxiv.org/abs/2505.19355v1 |
27th ACM International Conference on Multimedia , pp. 2682–2686, 2019. Dean Eckles, Ren ´e F Kizilcec, and Eytan Bakshy. Estimating peer effects in networks with peer encouragement designs. Proceedings of the National Academy of Sciences , 113(27): 7316–7322, 2016. Song Gao, Jinmeng Rao, Yuhao Kang, Yunlei Liang, Jake ... | https://arxiv.org/abs/2505.19355v1 |
Ma, Yuxiao Dong, Kuansan Wang, and Jie Tang. Deepinf: Social influence prediction with deep learning. In Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining , pp. 2110–2119, 2018. Rohit Ram and Marian-Andrei Rizoiu. Empirically measuring online social influence. EPJ Data Sci... | https://arxiv.org/abs/2505.19355v1 |
Qingyuan Zhao, Murat A Erdogdu, Hera Y He, Anand Rajaraman, and Jure Leskovec. Seismic: A self-exciting point process model for predicting tweet popularity. In Proceedings of the 21th ACM SIGKDD international conference on knowledge discovery and data mining , pp. 1513–1522, 2015. Fan Zhou, Xovee Xu, Goce Trajcevski, a... | https://arxiv.org/abs/2505.19355v1 |
Sequence modeling has recently been dominated by transformer architectures (Vaswani et al., 2017) and state space models. Time series adaptations include Informer (Zhou et al., 2021b), using ProbSparse self-attention, and Autoformer (Wu et al., 2021), leveraging auto-correlation for period- based dependencies. State Sp... | https://arxiv.org/abs/2505.19355v1 |
in Section 4.3, which represent the mean Root Mean Squared Error (RMSE) and Binary Cross Entropy (BCE) across these datasets. Here, we report individual dataset results to offer a granular view of model behavior under different counterfactual scenarios. 16 Preprint. Under review. Model Base Scenario 1: Exposure Scenari... | https://arxiv.org/abs/2505.19355v1 |
0.478 0.433 0.408 0.538 0.478 0.463 Transformer + Attention 0.508 0.423 0.378 0.448 0.393 0.353 0.528 0.463 0.443 Transformer + Layer 0.513 0.433 0.383 0.458 0.403 0.373 0.508 0.428 0.413 Transformer + Adapter 0.493 0.403 0.353 0.428 0.373 0.338 0.543 0.483 0.463 Mamba + Token 0.528 0.453 0.393 0.468 0.418 0.398 0.533 ... | https://arxiv.org/abs/2505.19355v1 |
0.193 0.183 0.193 0.188 0.183 0.238 0.218 0.213 Mamba + Adapter 0.110 0.183 0.178 0.168 0.178 0.173 0.168 0.228 0.213 0.208 Table 10: Root Mean Squared Error (RMSE) of predicted social media engagement metrics over a 7-day horizon under counterfactual scenarios for the covid dataset. Model Scenario 1: Exposure Scenario... | https://arxiv.org/abs/2505.19355v1 |
Optimized Text Embedding Models and Benchmarks for Amharic Passage Retrieval Kidist Amde Mekonnen* University of Amsterdam k.a.mekonnen@uva.nlYosef Worku Alemneh* Independent Researcher yosefwalemneh@gmail.comMaarten de Rijke University of Amsterdam m.derijke@uva.nl Abstract Neural retrieval methods using transformer- ... | https://arxiv.org/abs/2505.19356v1 |
which address this limitation by jointly encoding query-document pairs, capturing richer contextual interactions, with a computational overhead thatarXiv:2505.19356v1 [cs.IR] 25 May 2025 restricts their use to re-ranking candidate docu- ments (Humeau et al., 2020). As an alternative, late-interaction models (e.g., ColB... | https://arxiv.org/abs/2505.19356v1 |
Amharic occupational terms to male forms even when the context is gender-neutral. Such errors reflect broader research gaps in NLP, where systems disproportionately prioritize high- resource languages, thereby exacerbating inequities faced by underrepresented linguistic communi- ties (Shen et al., 2024). Amharic, the w... | https://arxiv.org/abs/2505.19356v1 |
Nonetheless, their ef- fectiveness in morphologically complex languages like Amharic remains unexplored, as current evalu- ations do not account for challenges arising from root-based and templatic morphologies. Beyond data scarcity, retrieval performance is further constrained by morphological complexity and tokenizat... | https://arxiv.org/abs/2505.19356v1 |
a fixed-length vector representation via a transformer encoder Enc (·): qenc=EncQ(q), p enc=EncP(p) (1) The relevance score between a query qand a pas- sagepis computed using a similarity function f(q, p) =sim(qenc, penc), where sim (·,·)typically denotes the dot product or cosine similarity. ColBERT: Late interaction ... | https://arxiv.org/abs/2505.19356v1 |
and the corresponding article bodies as passages. As the dataset lacks explicit relevance judgments, we adopt a heuristic super- vision approach: each headline is assumed to be relevant to its associated article. To validate this as- sumption, we manually examined a random subset of query-passage pairs and confirmed hi... | https://arxiv.org/abs/2505.19356v1 |
baseline are marked with†, based on a paired t-test. RQ2 How do different retrieval paradigms com- pare in effectiveness, establishing a bench- mark for Amharic passage retrieval? (Sec- tion 6.2) RQ3 How does tokenization quality, particularly subword segmentation, impact retrieval ef- fectiveness in morphologically ri... | https://arxiv.org/abs/2505.19356v1 |
subword segmentation, impacts re- trieval effectiveness in morphologically rich, low- resource languages, using Amharic as a case study. We focus on subword fertility, defined as the av- erage number of subword tokens per word (Pietra et al., 1997), as a key indicator of tokenization quality. Figure 1 presents fertilit... | https://arxiv.org/abs/2505.19356v1 |
baseline are marked with†, based on a paired t-test. ColBERT-BERT- Med-AmharicColBERT-RoBERT a- Med-AmharicColBERT-RoBERT a- Base-Amharic Model0.8000.8250.8500.8750.9000.9250.9500.9751.000Retrieval Metric Score 40M 42M 110M Metrics MRR@10 NDCG@10 Recall@10 Recall@50 Recall@100 Figure 3: Effect of base model size on Col... | https://arxiv.org/abs/2505.19356v1 |
(iv) Qualitative observations: Manual inspection of top-ranked outputs shows that Amharic-optimized dense models generally retrieve more contextually appropriate content. However, even the best mod- els struggle with negation, temporal shifts, and nuanced entailment. For instance, given the query “Was the planned prote... | https://arxiv.org/abs/2505.19356v1 |
pre-trained on a relatively modest corpus of 300 million tokens from web, news, and social me- dia sources. This is substantially smaller than the corpora used for high-resource language, e.g., En- glish BERT (3.3B) and RoBERTa (30B). Such data limitations may affect model generalization and downstream retrieval perfor... | https://arxiv.org/abs/2505.19356v1 |
We encourage the community to use our models and datasets responsibly, and to continue advanc- ing equitable IR systems that serve linguistically diverse users. References Ife Adebara, AbdelRahim Elmadany, Muhammad Abdul-Mageed, and Alcides Alcoba Inciarte. 2023. SERENGETI: Massively multilingual language mod- els for ... | https://arxiv.org/abs/2505.19356v1 |
of deep bidirectional transformers for language under- standing. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Tech- nologies, Volume 1 (Long and Short Papers) , pages 4171–4186, Minneapolis, Minnesota. Association for Computational ... | https://arxiv.org/abs/2505.19356v1 |
problem! Exploring the viability of pretrained multilingual language models for low- resourced languages. In Proceedings of the 1st Work- shop on Multilingual Representation Learning , pages 116–126, Punta Cana, Dominican Republic. Associa- tion for Computational Linguistics. Stephen Della Pietra, Mark Epstein, Salim R... | https://arxiv.org/abs/2505.19356v1 |
arXiv preprint arXiv:1803.05355 . Cagri Toraman, Eyup Halit Yilmaz, Furkan ¸ Sahinüc, and Oguzhan Ozcelik. 2023. Impact of tokenization on language models: An analysis for Turkish. ACM Trans. Asian Low-Resour. Lang. Inf. Process. , 22(4). Ahmet Üstün, Gosse Bouma, and Gertjan van Noord. 2019. Cross-lingual word embeddi... | https://arxiv.org/abs/2505.19356v1 |
evaluation proto- cols to ensure fair, consistent, and progress-driving comparisons in future research. A.1 Generalization to 2AIRTC: Amharic-Specific vs. Multilingual Models To assess the generalization capacity of retrieval models trained on the Amharic Passage Retrieval Dataset, we evaluate their zero-shot performan... | https://arxiv.org/abs/2505.19356v1 |
results as indicative for completeness. A.4 Toward Robust Benchmarks for Amharic Information Retrieval Although this study provides strong baselines for Amharic dense retrieval, the limitations of 2AIRTC, particularly its small query pool (240 topics) and sparse, sometimes inconsistent relevance annota- tions, signific... | https://arxiv.org/abs/2505.19356v1 |
weakening the learning signal during contrastive training. To address these gaps, future work should: •Incorporate curated or user-derived queries (e.g., search logs or community Q&A), •Employ better hard negative mining strategies, and •Collect human-annotated relevance labels for ro- bust evaluation. Model MRR NDCG R... | https://arxiv.org/abs/2505.19356v1 |
to learning rate and training duration. Figure 4: Negation failure case: The model retrieves the same top passage for both a positive (Query 1) and a negated (Query 2) version of the query, with comparable similarity scores. This reflects a lack of semantic sensitivity to negation. 2e-5 5e-5 Learning Rate64 128 256Batc... | https://arxiv.org/abs/2505.19356v1 |
arXiv:2505.19360v1 [cs.CL] 25 May 2025ChartLens: Fine-grained Visual Attribution in Charts Manan Suri , Puneet Mathur *, Nedim Lipka , Franck Dernoncourt ,Ryan A. Rossi ,Dinesh Manocha University of Maryland, College Park Adobe Research manans@umd.edu ,puneetm@adobe.com Abstract The growing capabilities of multimodal l... | https://arxiv.org/abs/2505.19360v1 |
2023b; Li et al., 2023b), and post-hoc attribution (Huo et al., 2023; Chen et al., 2023) aim to mitigate hallucination by enabling users to trace responses back to their origins. For resolving visual hallucinations specifically, post-generation validation approaches like (Zhou et al., 2023; Lee et al., 2023b; Yin et al... | https://arxiv.org/abs/2505.19360v1 |
financial documents and policy datasets, such as SEC Filings, the World Bank Open Data, Open Government Data, and the Global Terrorism Database. The benchmark fea- tures diverse chart styles and includes retrieval, reasoning, and computation-based questions, all paired with fine-grained visual attribution annota- tions... | https://arxiv.org/abs/2505.19360v1 |
incorporated OCR sub-networks . Pre-trained models like STL-CQA (Singh and Shekhar, 2020) and VisionTaPas (Masry et al., 2022) further improved performance by leverag- ing transformer-based architectures . Generation-based approaches dominate tasks like chart captioning and chart-to-table conver- sion. Models such as D... | https://arxiv.org/abs/2505.19360v1 |
as Ac, representing all potential regions within the chart. 4.1 Data Sources The ChartV A-Eval Benchmark is constructed from a diverse set of data sources to ensure compre- hensive evaluation of post hoc attribution in chart- based visual question answering (VQA). By incor- porating both synthetic and real-world charts... | https://arxiv.org/abs/2505.19360v1 |
assessed the annotations based on two criteria: (1) Relevance — ensuring the annotated elements directly support the answer, and (2) Completeness — verifying that all necessary chart elements were included. This process ensured high-quality and precise attribution annotations for both datasets. Further details on attri... | https://arxiv.org/abs/2505.19360v1 |
noisy and low-quality images more robustly. It produces pre- cise masks that closely align with the boundaries of chart elements, even in complex cases. Addi- tionally, SAM naturally suppresses background fea- 5 tures like grid lines by generating weaker masks (low IoU) for these elements, as they lack the spa- tial co... | https://arxiv.org/abs/2505.19360v1 |
erencing these elements, the model’s response be- comes more transparent and easier to verify. 6 Experiments 6.1 Baselines Zero-shot GPT-4o Bounding Box Prompting: As a baseline, we prompt GPT-4o (OpenAI, 2024) to predict normalized bounding box coordinates for chart components (e.g., lines, bars, pie sectors) based on... | https://arxiv.org/abs/2505.19360v1 |
charts in- volves referring to singular points. Since grounding models generate bounding boxes or regions, it is challenging to precisely match these regions to in- dividual ground truth points without ambiguity. To address this, two metrics are defined for evaluation: 1.Detection Rate: Measures the proportion of groun... | https://arxiv.org/abs/2505.19360v1 |
KOSMOS2 achieve high detection rate, this can largely be explained by the high Chart% area covered by their attributions; covering large areas of the chart makes capturing specific points non- trivial but reduces the specificity of attributions, making them less effective at fine-grained localiza- tion. In contrast, Ch... | https://arxiv.org/abs/2505.19360v1 |
it can be improved or replaced with more advanced methods in future iterations. Second, our approach primarily focuses on vi- sual chart elements, such as bars, points, or sectors, and does not account for textual components like captions, labels, or titles. Addressing this limita- tion and integrating text-based reaso... | https://arxiv.org/abs/2505.19360v1 |
ure dataset for visual reasoning. arXiv preprint arXiv:1710.07300 . 9 Shankar Kantharaj, Rixie Tiffany Ko Leong, Xiang Lin, Ahmed Masry, Megh Thakkar, Enamul Hoque, and Shafiq Joty. 2022. Chart-to-text: A large-scale benchmark for chart summarization. arXiv preprint arXiv:2203.06486 . Yannis Katsis, Saneem Chemmengath,... | https://arxiv.org/abs/2505.19360v1 |
in Natural Language Processing: System Demonstrations , pages 250–258. Fanqing Meng, Wenqi Shao, Quanfeng Lu, Peng Gao, Kaipeng Zhang, Yu Qiao, and Ping Luo. 2024. Char- tassisstant: A universal chart multimodal language model via chart-to-table pre-training and multitask instruction tuning. arXiv preprint arXiv:2401.0... | https://arxiv.org/abs/2505.19360v1 |
Shi, Yu Qiao, and Junchi Yan. 2023. Structchart: Perception, structur- ing, reasoning for visual chart understanding. arXiv preprint arXiv:2309.11268 . Jianwei Yang, Hao Zhang, Feng Li, Xueyan Zou, Chun- yuan Li, and Jianfeng Gao. 2023. Set-of-mark prompting unleashes extraordinary visual grounding in gpt-4v. arXiv pre... | https://arxiv.org/abs/2505.19360v1 |
different years, countries, and regions. The dataset comprises 841 unique vari- ables and 160 entities, with data spanning from 1960 to 2016. These statistics are represented in three main plot types: bar plots, line plots, and scat- ter plots. The plots vary in their visual elements, including legend positions, fonts,... | https://arxiv.org/abs/2505.19360v1 |
with reviewing bounding box annotations to assess Relevance, ensuring that the annotated chart elements directly supported the provided an- swers, and Completeness, verifying that all nec- essary chart elements were included. The annota- tion process was conducted in an interactive setting where annotators could inspec... | https://arxiv.org/abs/2505.19360v1 |
Label Pie Chart Sectors 1:procedure DETECT PIECHART SEC- TORS (image_path, predictor) 2: Input: Image path image _path 3: Output: Processed image with labeled pie chart sectors 4: Step 1: Preprocess Image 5: Load the image and convert it to grayscale 6: Apply binary thresholding with Otsu’s method 7: Detect external co... | https://arxiv.org/abs/2505.19360v1 |
arXiv:2505.19376v1 [cs.CL] 26 May 2025Belief Attribution as Mental Explanation: The Role of Accuracy, Informativity, and Causality Lance Ying∗1,2, Almog Hillel∗1, Ryan Truong∗2 Vikash K. Mansinghka1, Joshua B. Tenenbaum1, Tan Zhi-Xuan1 1Massachusetts Institute of Technology, Cambridge, MA, USA 2Harvard University, Camb... | https://arxiv.org/abs/2505.19376v1 |
on the left shows a player in a trea- sure game, who is trying to find a blue key to unlock the blue door to retrieve the gold chest. On the right, participants are asked to rank three statements about the player’s beliefs based on their likelihood of attributing each statement. belief attribution and its relationship ... | https://arxiv.org/abs/2505.19376v1 |
we study the selective attribution of beliefs, which we model as explanatory choice. To do so, we use the fact that a generative theory-of-mind is also a causal model, allowing us to evaluate the effect of hypothetical inter- ventions on agents’ minds (Ho et al., 2022; Chen et al., 2024; S. Wu et al., 2024). We use thi... | https://arxiv.org/abs/2505.19376v1 |
the agent’s action Atand a partial observation Ot(i.e. ev- erything in Stexcept the contents of unopened boxes). We also assume a uniform prior over initial environment states P(S0)(corresponding to all ways up to 2 keys can be placed in boxes while ensuring the goal is reachable), and an initial belief prior P(B0)that... | https://arxiv.org/abs/2505.19376v1 |
a listener who can also observe the environ- ment but has not drawn inferences from the agent’s actions:1 Info(ϕ,t):= KL[P(St,Bt|ϕt=T,Ot=ot)||P(St,Bt|Ot=ot)] 1Since our experiment restricts the statements that participants rank to those about the player’s beliefs, it is natural to model the lis- tener as knowing about ... | https://arxiv.org/abs/2505.19376v1 |
seeing that it is empty), even if they are necessary to explain certain actions. Probabilistic Attribution of Belief Statements Each of the factors above can combined into an overall score measuring the explanatory strength of a belief statement ϕ. For a set of factors F, we do this via linear combination of their loga... | https://arxiv.org/abs/2505.19376v1 |
the fitted parameters αf.Info* is an alternative informativity measure for listeners with no visibility of the environment. statement to the agent. Specifically, we provide participants with the following prompt, which avoids framing the task in terms of explanation or communication to a specific listener: “Please rank... | https://arxiv.org/abs/2505.19376v1 |
highly ranked statement that humans did in mostscenarios (see clusters at bottom left of each plot). How- ever, ranking the other two statements appears to be more challenging, and both models and humans show significant uncertainty. We also found a high correlation of r=0.81 be- tween the Acc+Infomodel and the Causal ... | https://arxiv.org/abs/2505.19376v1 |
if the player assigned a high enough probability to the red key being in either box 2 or 3, even if they were uncertain as to which), a rational player would first go to check box 2, since box 2 is closer to both the player and the goal. This results in low likelihood of the observed actions under the intervention that... | https://arxiv.org/abs/2505.19376v1 |
model’s fit to human judgments. Fourth, our results shows the potential value of unifying the causal and communicative aspects of explanation selection. Recent work by Kirfel et al. (2024) provides one promising direction, suggesting that people select explanations that are causally relevant to the listeners. Extending... | https://arxiv.org/abs/2505.19376v1 |
959– 971. Icard, T. F., & Knobe, J. (2016). Causality, normality, and sampling propensity. In Proceedings of the annual meeting of the cognitive science society (V ol. 38). Icard, T. F., Kominsky, J. F., & Knobe, J. (2017). Normality and actual causal strength. Cognition ,161, 80–93.Kirfel, L., Harding, J., Shin, J. Y ... | https://arxiv.org/abs/2505.19376v1 |
(NIPE): Modeling probabilistic social inferences from linguistic inputs. arXiv preprint arXiv:2306.14325 . Ying, L., Collins, K. M., Wong, L., Sucholutsky, I., Liu, R., Weller, A., . . . Tenenbaum, J. B. (2025). On benchmark- ing human-like intelligence in machines. arXiv preprint arXiv:2502.20502 . Ying, L., Zhi-Xuan,... | https://arxiv.org/abs/2505.19376v1 |
arXiv:2505.19384v1 [cs.CL] 26 May 2025GSA-TTS: TOWARD ZERO-SHOT SPEECH SYNTHESIS BASED ON GRADUAL STYLE ADAPTOR Seokgi Lee*, Jungjun Kim* ABSTRACT We present the gradual style adaptor TTS (GSA-TTS) with a novel style encoder that gradually encodes speaking styles from an acoustic reference for zero-shot speech synthesi... | https://arxiv.org/abs/2505.19384v1 |
high-fidelity and generalizable zero-shot style transfer, we propose a gradual style adaptor-based text-to- speech model (GSA-TTS), including local style encoder and global style encoder. We introduce a style segmen- tation strategy for the first time, employing a pre-trained automatic speech recognition (ASR) model. A... | https://arxiv.org/abs/2505.19384v1 |
2.2. Local Style Encoder The local-style encoder (LSE) receives the queued style seg- ments as input, aiming to extract a vector containing the word-level style information. Following [21], LSE consists of spectral processing, temporal processing and multihead- attention with temporal average pooling. The spectral pro-... | https://arxiv.org/abs/2505.19384v1 |
w ) =γ(w)·x−µ σ+β(w), γ(w) =Eγ∗w, β (w) =Eβ∗w,(1) where xdenotes the input text features. µandσare mean and standard deviation along the feature dimension.3. EXPERIMENTAL SETUP We train GSA-TTS using two multi-speaker datasets: LibriTTS- R [27] (2456 speakers) and the VCTK dataset [28] (109 speakers). For testing, we e... | https://arxiv.org/abs/2505.19384v1 |
a plug-and-play manner. To evaluate the robustness of GSA- TTS, we synthesize the audio via non-parallel style transfer that uses a random audio from a target speaker as acoustic reference. As shown in Table 1, we confirm statistically sig- nificant differences between GSA-TTS and comparative mod- els in CSMOS with the... | https://arxiv.org/abs/2505.19384v1 |
POS tagging. Table 4 illustrates the voiced frame ratio of the style segment based on verbs, adjectives, nouns, and others, result- ing in percentages of 73.45%, 65.73%, 59.72%, and 54.68%, respectively. Combining the above results, we speculate that the style segment tagged with noun, adjective, and verb is more infor... | https://arxiv.org/abs/2505.19384v1 |
pp. 1–5. [7] Kaisheng Yao and Geoffrey Zweig, “Sequence-to- sequence neural net models for grapheme-to-phoneme conversion,” arXiv preprint arXiv:1506.00196 , 2015. [8] Tero Karras, Samuli Laine, and Timo Aila, “A style- based generator architecture for generative adversarial networks,” in Proceedings of the IEEE/CVF co... | https://arxiv.org/abs/2505.19384v1 |
Auli, and David Grangier, “Language modeling with gated convolu- tional networks,” in International conference on ma- chine learning . PMLR, 2017, pp. 933–941. [23] Chulhee Yun, Srinadh Bhojanapalli, Ankit Singh Rawat, Sashank J Reddi, and Sanjiv Kumar, “Are transformers universal approximators of sequence-to-sequence ... | https://arxiv.org/abs/2505.19384v1 |
arXiv:2505.19388v1 [cs.CL] 26 May 2025GEC -METRICS : A Unified Library for Grammatical Error Correction Evaluation Takumi Goto, Yusuke Sakai, Taro Watanabe Nara Institute of Science and Technology (NAIST) {goto.takumi.gv7, sakai.yusuke.sr9, taro}@is.naist.jp Abstract We introduce GEC-METRICS , a library for us- ing and... | https://arxiv.org/abs/2505.19388v1 |
has similarly con- solidated evaluation metrics into a unified library, which has further accelerated and simplified model development. In the same way, a unified framework for the GEC evaluation metric is highly desired. We introduce GEC-METRICS , a unified frame- 1 He go to the school .Source:Human Annotation He goes... | https://arxiv.org/abs/2505.19388v1 |
=P e∈IweP e∈Heditwe,Recall =P e∈IweP e∈Reditwe (2) ERRANT (Felice et al., 2016; Bryant et al., 2017) sets we= 1 .0for all of edits, and PT-ERRANT (Gong et al., 2022) computes a weight by BERTScore (Zhang et al., 2020) or BARTScore (Yuan et al., 2021). GoToScorer (Go- tou et al., 2020) uses the error correction difficul... | https://arxiv.org/abs/2505.19388v1 |
2014). Nonetheless, the number of available meta-evaluation datasets re- mains limited. One contributing factor is the lack of a unified framework for GEC evaluation metrics, which hinders consistent and comprehensive vali- dation and increases the cost of implementing base- lines when constructing meta-evaluation data... | https://arxiv.org/abs/2505.19388v1 |
PT-ERRANT , and GoToScorer as edit-level met- rics, GLEU andGREEN asn-gram level met- rics. For reference-free metrics, it supports SOME , Scribendi ,IMPARA ,LLM-S , and LLM-E6as sentence-level metrics. We carefully designed the library for extensibility and ease of changing hyper- parameters and base models, supportin... | https://arxiv.org/abs/2505.19388v1 |
in L20 and L23 take a metric 4 System-level Sentence-level Metric GJG15 SEEDA-S SEEDA-E GJG15 SEEDA-S SEEDA-E Base +Fluency Base +Fluency Base +Fluency Base +Fluency r ρ r ρ r ρ r ρ r ρ Acc. τAcc. τAcc. τAcc. τAcc. τ ERRANT .647 .687 .539 .343 -.592 -.156 .682 .643 -.508 .033 .654 .307 .594 .189 .544 .087 .608 .217 .55... | https://arxiv.org/abs/2505.19388v1 |
2024b) enables discussions on evaluation performance by focusing on competitive systems in human evaluation, and theedit-level attribution shows which edit opera- tion a metric focuses on in the evaluation (Goto et al., 2024). GEC-METRICS provides tools for such analyses and result visualization. 7We provide the docume... | https://arxiv.org/abs/2505.19388v1 |
by (Kobayashi et al., 2024a), indicating the need for further discussion on the validity of the approach. Figure 4 shows the window-analysis results for IMPARA. We used hu- man TrueSkill rankings of SEEDA-S and used 4 as the window size. An observation is that the correla- tions suddenly drops at x= 7, which is consist... | https://arxiv.org/abs/2505.19388v1 |
metric im- plementations, GEC-METRICS aims to support and strengthen these efforts. Impacts for the community. GEC-METRICS serves as a powerful tool for researchers to eas- ily develop evaluation methods. It also accelerates their application in the GEC field, including bias investigations, integration with learning an... | https://arxiv.org/abs/2505.19388v1 |
Language Tech- nologies, Volume 1 (Long and Short Papers) , pages 4171–4186, Minneapolis, Minnesota. Association for Computational Linguistics. Mariano Felice, Christopher Bryant, and Ted Briscoe. 2016. Automatic extraction of learner errors in ESL sentences using linguistically enhanced alignments. InProceedings of CO... | https://arxiv.org/abs/2505.19388v1 |
Conference , pages 303–313, Tokyo, Japan. Associa- tion for Computational Linguistics. Mengsay Loem, Masahiro Kaneko, Sho Takase, and Naoaki Okazaki. 2023. Exploring effectiveness of GPT-3 in grammatical error correction: A study on performance and controllability in prompt-based methods. In Proceedings of the 18th Wor... | https://arxiv.org/abs/2505.19388v1 |
Jiaxi Yang, Jingren Zhou, and 25 oth- ers. 2025. Qwen2.5 technical report. Preprint , arXiv:2412.15115. Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, and 1 others. 2019. Language models are unsupervised multitask learn- ers.OpenAI blog , 1(8):9. Vyas Raina and Mark Gales. 2023. Minimu... | https://arxiv.org/abs/2505.19388v1 |
Learning Representations . Yike Zhao, Xiaoman Wang, Yunshi Lan, and Weining Qian. 2025. UnifiedGEC: Integrating grammatical 9 error correction approaches for multi-languages with a unified framework. In Proceedings of the 31st Inter- national Conference on Computational Linguistics: System Demonstrations , pages 37–45,... | https://arxiv.org/abs/2505.19388v1 |
not contain the training scripts, but we make them public in a separate repository13. bert-base-cased is used for computing the sim- ilarity score with the threshold 0.9. LLM-S and LLM-E. For GPT-4-S, we use beta.chat.completions.parse API for the OpenAI models and use OUTLINES li- brary (Willard and Louf, 2023)14for t... | https://arxiv.org/abs/2505.19388v1 |
sources fileDrag and drop file hereLimit 200MB per file • TXT Browse filesEnter hypothesesEnter hypotheses (one per line)He goes to a school .Or, upload hypotheses fileDrag and drop file hereLimit 200MB per file • TXT Browse filesEnter references0Enter references0 (one per line)He goes to school .Or, upload references0... | https://arxiv.org/abs/2505.19388v1 |
arXiv:2505.19392v1 [cs.CL] 26 May 2025Simple and Effective Baselines for Code Summarisation Evaluation Jade Robinson and Jonathan K. Kummerfeld University of Sydney jonathan.kummerfeld@sydney.edu.au Abstract Code documentation is useful, but writing it is time-consuming. Different techniques for gen- erating code summa... | https://arxiv.org/abs/2505.19392v1 |
in QA, our method does not favour longer (or shorter) summaries. These differences highlight the distinctiveness of code summarisation and therefore, the value of research in this space. Our work provides novel baselines that are simple and effective, forming a solid foundation for further exploration. 2 Related Work C... | https://arxiv.org/abs/2505.19392v1 |
but for code summarisation they have the additional benefit that they could be used to flag low quality summaries within an existing code base. 3 Task Code summarisation is the task of generating a summary of a code snippet. We are proposing new metrics for this task. The aim of the metric is to output a score that cap... | https://arxiv.org/abs/2505.19392v1 |
al., 2021; Haque et al., 2022; Mastropaolo et al., 2024), we aim to max- imise correlation with human evaluation scores. We follow Haque et al. (2022)’s methodology: (1) when there are multiple human scores for a sample, we compare with the mean to reduce the impact of noise from disagreement, and (2) we use Spear- man... | https://arxiv.org/abs/2505.19392v1 |
embedding meth- ods is likely to continue to provide improvements here. One key difference between these approaches is cost, which will be discussed below. ask-LLM-no-ref is just as effective. The per- formance of the Ask-LLM-Directly style metrics is stable regardless of whether the reference sum- mary is provided, wi... | https://arxiv.org/abs/2505.19392v1 |
mercial tools and compute costs open source model OLMo-2 (we used an A100). These results show that these approaches are clearly much cheaper than running human evaluations, but still more ex- pensive than metrics which can be run locally, e.g. gte-base-en, Sentence-BERT and n-gram methods. 7 Conclusion We introduce a ... | https://arxiv.org/abs/2505.19392v1 |
Automatic semantic augmentation of language model prompts (for code summarization). In Proceedings of the IEEE/ACM 46th International Conference on Software Engineer- ing, ICSE ’24, New York, NY , USA. Association for Computing Machinery. Uri Alon, Shaked Brody, Omer Levy, and Eran Ya- hav. 2019. code2seq: Generating s... | https://arxiv.org/abs/2505.19392v1 |
Softw. Eng. Methodol. , 32(1):Article 23. Sakib Haque, Zachary Eberhart, Aakash Bansal, and Collin McMillan. 2022. Semantic similarity met- rics for evaluating source code summarization. In Proceedings of the 30th IEEE/ACM International Conference on Program Comprehension , page 36–47. Association for Computing Machine... | https://arxiv.org/abs/2505.19392v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.