Natural Language Processing Journal最新文献

{"title":"Position-aware end-to-end cross-document event coreference resolution for Dutch","authors":"Loic De Langhe,&nbsp;Orphée De Clercq,&nbsp;Veronique Hoste","doi":"10.1016/j.nlp.2025.100184","DOIUrl":"10.1016/j.nlp.2025.100184","url":null,"abstract":"<div><div>Natural language understanding entails the ability to comprehend the relations between various people, objects or events throughout one, or multiple, text(s). Event coreference resolution (ECR) is a discourse-based natural language processing (NLP) task which aims to link those textual events, be they real or fictional, that refer to the same conceptual event. In this paper, we introduce a novel end-to-end approach for cross-document ECR which combines expert-level positional knowledge and graph-based representations in order to create a memory-efficient and accurate system meant for the detection and resolution of events in large document collections. We make three fundamental architectural changes to a current state-of-the-art cross-document ECR system and show that our approach outperforms this earlier model (+ 4% CONLL F1) on a large Dutch ECR dataset. Moreover, we show through in-depth qualitative and quantitative analysis that our proposed approach consistently detects more relevant events and suffers notably less from the typical issues models exhibit when predicting coreference chains.</div></div>","PeriodicalId":100944,"journal":{"name":"Natural Language Processing Journal","volume":"13 ","pages":"Article 100184"},"PeriodicalIF":0.0,"publicationDate":"2025-12-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145098437","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Whose morality do they speak? Unraveling cultural bias in multilingual language models","authors":"Meltem Aksoy","doi":"10.1016/j.nlp.2025.100172","DOIUrl":"10.1016/j.nlp.2025.100172","url":null,"abstract":"<div><div>Large language models (LLMs) have become integral tools in diverse domains, yet their moral reasoning capabilities across cultural and linguistic contexts remain underexplored. This study investigates whether multilingual LLMs, such as GPT-3.5-Turbo, GPT-4o-mini, Llama 3.1, and MistralNeMo, reflect culturally specific moral values or impose dominant moral norms, particularly those rooted in English. Using the updated Moral Foundations Questionnaire (MFQ-2) in eight languages, Arabic, Farsi, English, Spanish, Japanese, Chinese, French, and Russian, the study analyzes the models’ adherence to six core moral foundations: care, equality, proportionality, loyalty, authority, and purity. The results reveal significant cultural and linguistic variability, challenging the assumption of universal moral consistency in LLMs. Although some models demonstrate adaptability to diverse contexts, others exhibit biases influenced by the composition of the training data. These findings underscore the need for culturally inclusive model development to improve fairness and trust in multi-lingual AI systems.</div></div>","PeriodicalId":100944,"journal":{"name":"Natural Language Processing Journal","volume":"12 ","pages":"Article 100172"},"PeriodicalIF":0.0,"publicationDate":"2025-09-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144534372","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Analyzing social media discourse of avian influenza outbreaks","authors":"Marzieh Soltani ,&nbsp;Shayan Sharif ,&nbsp;Rozita Dara","doi":"10.1016/j.nlp.2025.100176","DOIUrl":"10.1016/j.nlp.2025.100176","url":null,"abstract":"<div><div>The ongoing avian influenza outbreaks have had significant implications for the global poultry industry in addition to a wide range of wild birds and mammals. To enhance our understanding of public perceptions and reactions during such outbreaks, the present study examined social media discourse surrounding avian influenza on X (formerly known as Twitter). By employing advanced large language models, including DistilBERT for post filtering (average 89.5% accuracy via 5-fold cross-validation) along with Mixtral-8x7B, BERTopic, and RoBERTa for sentiment and topic/user analysis, this research categorizes the discussions and sentiments expressed by users over time. Our analysis focused on three aspects: main topics, sentiment, and temporal patterns of user engagement surrounding avian influenza outbreaks. Sentiment analysis revealed that a majority of posts related to economic impact (81.2%), wildlife (71.7%), and human cases (67.9%) expressed negative sentiment. Through topic modeling, prevalent topics of concern were identified in discussions, including concerns about transmission to humans and mammals, as well as issues related to food security and food prices. Additionally, the analysis of user engagement patterns showed distinct categories of users and highlighted the contributions of top users in shaping the discourse. Emotion analysis showed that over 80% of posts on major topics conveyed emotions such as anger, sadness, and fear, especially during periods of high case reports. The present study underscores the potential of social media analysis to understand public reactions to avian influenza outbreaks and to facilitate effective responses to public concerns and needs.</div></div>","PeriodicalId":100944,"journal":{"name":"Natural Language Processing Journal","volume":"12 ","pages":"Article 100176"},"PeriodicalIF":0.0,"publicationDate":"2025-09-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144886249","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"On English-Chinese Neural Machine Translation leveraging Transformer model","authors":"Subrota Kumar Mondal ,&nbsp;Yijun Chen ,&nbsp;Yuning Cheng ,&nbsp;Hong-Ning Dai ,&nbsp;Syed B. Alam ,&nbsp;H.M. Dipu Kabir","doi":"10.1016/j.nlp.2025.100166","DOIUrl":"10.1016/j.nlp.2025.100166","url":null,"abstract":"<div><div>In today’s era of globalization, people’s cross-cultural communication has become increasingly frequent, and photo translation (photo, image, or scene text translation) technology has become an important tool. By using this translation technology, people can easily recognize and translate text from other languages without the need for manual input or translation. This has important practical value for people in fields such as tourism, business, education, and research. Therefore, photo translation technology has become an indispensable tool, providing more convenience to people’s lives and work. To this, this paper aims to achieve high accuracy English to Chinese photo translation, which can be divided into three stages: <span>text detection</span>, <span>text recognition</span>, and <span>text translation (i.e., machine translation)</span>. We observe that in text detection and recognition, we have challenges with occluded text, hand-written text, scene text, text with complex layout, distorted text, and many others. However, in this paper, we limit our analysis to Translation phase. For detection and recognition phase, we make use of current state-of-the-art methodologies, such as <span>DBNet</span> (Liao et al., 2020) model for detection and the <span>ABINet</span> (Fang et al., 2021) model for recognition. In the translation part, we use Transformer model with modifications towards improving the translation accuracy. The modifications are mainly reflected in two aspects: <span>data preprocessing</span> and <span>optimizer</span>. In the data preprocessing part, we use the <span>BPE</span> (Byte Pair Encoding) algorithm instead of basic word-centered tokenization algorithms. In the context, <span>BPE</span> algorithm can divide words into smaller subwords, which can solve the problem of rare words to some extent and provide better word vectors for language model training. In the optimizer part, we use the <span>Lion</span> model proposed by Google instead of the widely used <span>Adam</span> optimizer that helps reduce the loss more quickly than using Adam optimizer for small size batch — with batch size 256 achieves the lowest test loss 0.392842 (−1.072171) and the highest BLEU4 score 0.381281 (+0.24063). This adds value in reducing the consumption of training resources and the sustainability of deep learning.</div></div>","PeriodicalId":100944,"journal":{"name":"Natural Language Processing Journal","volume":"12 ","pages":"Article 100166"},"PeriodicalIF":0.0,"publicationDate":"2025-09-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144472472","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Intermediate-task transfer learning for Indonesian NLP tasks","authors":"Adrianus Saga Ekakristi,&nbsp;Alfan Farizki Wicaksono,&nbsp;Rahmad Mahendra","doi":"10.1016/j.nlp.2025.100161","DOIUrl":"10.1016/j.nlp.2025.100161","url":null,"abstract":"<div><div>Transfer learning, a common technique in recent Natural Language Processing (NLP) research, involves pre-training a model on a large, unlabeled dataset using self-supervised methods and then fine-tuning it on a smaller, labeled dataset for a specific task. Recent studies have demonstrated that introducing an additional training step between pre-training and fine-tuning can further enhance model performance. This method is called intermediate-task transfer learning (ITTL). Although this approach can potentially improve performance in the target task, choosing an intermediate task that leads to the highest performance increase remains challenging. Furthermore, despite the extensive research on intermediate training methods in English NLP, the application of these techniques to Indonesian language processing is still relatively understudied. In this study, we apply the ITTL method to nine Indonesian NLP datasets, using each as both intermediate and target tasks, to investigate its behavior. Furthermore, we show that linear regression analysis can effectively identify factors that maximize performance improvement in target tasks when using ITTL. Our experiments reveal that ITTL enhances F1 score performance in the majority of cases, provided suitable intermediate tasks are selected. Specifically, our best performing model achieved performance gains in 9 out of 10 target task sets, with improvements reaching up to 18.6%. Our detailed analysis indicates that factors such as task type matching, task complexity, vocabulary size, and dataset size significantly influence the effectiveness of ITTL on target task performance. This research shows that ITTL, coupled with our proposed guidelines for intermediate task selection, offers a promising training paradigm for Indonesian NLP.</div></div>","PeriodicalId":100944,"journal":{"name":"Natural Language Processing Journal","volume":"12 ","pages":"Article 100161"},"PeriodicalIF":0.0,"publicationDate":"2025-09-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144291182","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"A survey and evaluation of text-to-speech systems for the Tamil language","authors":"Ahrane Mahaganapathy,&nbsp;Kengatharaiyer Sarveswaran","doi":"10.1016/j.nlp.2025.100171","DOIUrl":"10.1016/j.nlp.2025.100171","url":null,"abstract":"<div><div>This survey provides a comprehensive review of existing Tamil Text-to-Speech (TTS) synthesis systems, synthesis approaches, evaluation approaches, and highlights state-of-the-art approaches and challenges in handling linguistic nuances. Voice-based interfaces are becoming part of life. Therefore, it is import to have an expensive TTS system which can make human experience better. Tamil, with its rich linguistic features and diagnostic nature, presents significant challenges to speech synthesis. In addition to the survey, importantly this work proposes a perceptual evaluation framework which consists of expressiveness, low listening fatigue, and overall quality, in addition to traditional intelligibility and naturalness, dimensions to evaluate better human experience. This study also uses the Comparative Mean Opinion Score (CMOS) for the subjective evaluation instead of the Mean Opinion Score. A dataset for the evaluation was also carefully prepared and six widely used Tamil TTS systems were evaluated using Word Error Rate and the subjective evaluation was done using the proposed evaluation framework with the support of 30 evaluators. The reliability of the subjective evaluation is also assessed using Krippendorff’s Alpha. The results indicate the existing systems have significant room for improvement in all perceptual dimensions. The study underscores the need for evaluation datasets and evaluation approaches that cater to subjective perceptual dimensions of speech synthesis for better human experience and lays a foundation for future research and development in Tamil and similar TTS systems.</div></div>","PeriodicalId":100944,"journal":{"name":"Natural Language Processing Journal","volume":"12 ","pages":"Article 100171"},"PeriodicalIF":0.0,"publicationDate":"2025-09-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144549312","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Can “consciousness” be observed from large language model (LLM) internal states? Dissecting LLM representations obtained from Theory of Mind test with Integrated Information Theory and Span Representation analysis","authors":"Jingkai Li","doi":"10.1016/j.nlp.2025.100163","DOIUrl":"10.1016/j.nlp.2025.100163","url":null,"abstract":"<div><div>Integrated Information Theory (IIT) provides a quantitative framework for explaining consciousness phenomenon, positing that conscious systems comprise elements integrated through causal properties. We apply IIT 3.0 and 4.0 — the latest iterations of this framework — to sequences of Large Language Model (LLM) representations, analyzing data derived from existing Theory of Mind (ToM) test results. Our study systematically investigates whether the differences of ToM test performances, when presented in the LLM representations, can be revealed by IIT estimates, i.e., <span><math><msup><mrow><mi>Φ</mi></mrow><mrow><mo>max</mo></mrow></msup></math></span> (IIT 3.0), <span><math><mi>Φ</mi></math></span> (IIT 4.0), Conceptual Information (IIT 3.0), and <span><math><mi>Φ</mi></math></span>-structure (IIT 4.0). Furthermore, we compare these metrics with the Span Representations independent of any estimate for consciousness. This additional effort aims to differentiate between potential “consciousness” phenomena and inherent separations within LLM representational space. We conduct comprehensive experiments examining variations across LLM transformer layers and linguistic spans from stimuli. Our results suggest that sequences of contemporary Transformer-based LLM representations lack statistically significant indicators of observed “consciousness” phenomena but exhibit intriguing patterns under <em>spatio</em>-permutational analyses.</div></div>","PeriodicalId":100944,"journal":{"name":"Natural Language Processing Journal","volume":"12 ","pages":"Article 100163"},"PeriodicalIF":0.0,"publicationDate":"2025-09-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144472473","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Homophobia and transphobia span identification in low-resource languages","authors":"Prasanna Kumar Kumaresan ,&nbsp;Devendra Deepak Kayande ,&nbsp;Ruba Priyadharshini ,&nbsp;Paul Buitelaar ,&nbsp;Bharathi Raja Chakravarthi","doi":"10.1016/j.nlp.2025.100169","DOIUrl":"10.1016/j.nlp.2025.100169","url":null,"abstract":"<div><div>Online platforms have become prevalent because they promote free speech and group discussions. However, they also serve as platforms for hate speech, which can negatively impact the psychological well-being of vulnerable people. This is especially true for members of the LGBTQ＋ community, who are often the targets of homophobia and transphobia in online environments. Our study makes three main contributions: (1) we developed a new dataset with span-level annotations for homophobia and transphobia in Tamil, English, and Marathi; (2) we employed advanced language models using BERT-based architectures, Conditional Random Field (CRF), and Bidirectional Long Short-Term Memory (BiLSTM) layers to enhance span-level detection of harmful content; and (3) we conducted benchmarking to evaluate the effectiveness of monolingual and multilingual models in detecting subtle forms of hate speech. The annotated dataset, which is collected from real-world social media (YouTube) content, provides diverse language contexts and enhances the representation of low-resource languages. The span-based detection approach enables models to detect subtle linguistic nuances, leading to more precise content moderation that accounts for cultural differences. The experimental results show that our models achieve effective span detection, which provides valuable information for creating inclusive moderation tools. Our research leads to the development of AI systems, and we aim to reduce the burden on moderators and improve the quality of online experiences for LGBTQ＋ vulnerable.</div></div>","PeriodicalId":100944,"journal":{"name":"Natural Language Processing Journal","volume":"12 ","pages":"Article 100169"},"PeriodicalIF":0.0,"publicationDate":"2025-09-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144489870","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Large language models for ingredient substitution in food recipes using supervised fine-tuning and direct preference optimization","authors":"Thevin Senath ,&nbsp;Kumuthu Athukorala ,&nbsp;Ransika Costa ,&nbsp;Surangika Ranathunga ,&nbsp;Rishemjit Kaur","doi":"10.1016/j.nlp.2025.100177","DOIUrl":"10.1016/j.nlp.2025.100177","url":null,"abstract":"<div><div>In this paper, we address the challenge of recipe personalization through ingredient substitution. We make use of Large Language Models (LLMs) to build an ingredient substitution system designed to predict plausible substitute ingredients within a given recipe context. Given that the use of LLMs for this task has been barely done, we carry out an extensive set of experiments to determine the best LLM, prompt, and the fine-tuning setups. We further experiment with methods such as multi-task learning, two-stage fine-tuning, and Direct Preference Optimization (DPO). The experiments are conducted using the publicly available Recipe1MSub corpus. The best results are produced by the Mistral7-Base LLM after fine-tuning and DPO. This result outperforms the strong baseline available for the same corpus with a Hit@1 score of 22.04. Although LLM results lag behind the baseline with respect to other metrics such as Hit@3 and Hit@10, we believe that this research represents a promising step towards enabling personalized and creative culinary experiences by utilizing LLM-based ingredient substitution.</div></div>","PeriodicalId":100944,"journal":{"name":"Natural Language Processing Journal","volume":"12 ","pages":"Article 100177"},"PeriodicalIF":0.0,"publicationDate":"2025-09-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144865457","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Improving multilabel text emotion detection with emotion interrelation anchors","authors":"Polydoros Giannouris ,&nbsp;Vasileios Mygdalis ,&nbsp;Ioannis Pitas","doi":"10.1016/j.nlp.2025.100170","DOIUrl":"10.1016/j.nlp.2025.100170","url":null,"abstract":"<div><div>Emotion detection studies the problem of automatic identification of emotions expressed in text. Since multiple emotions may co-occur in a single text excerpt, state-of-the-art approaches often cast this multi-label classification task to multiple, independent binary classification tasks, each specialized for one emotion class. The main disadvantage of such approaches is that, by design, each binary classifier overlooks typical emotion interrelationships, such as co-occurrence (e.g., anger and fear) or mutual exclusiveness (e.g., sadness and joy). This paper proposes a simple and lightweight approach to re-introduce emotion interrelations into each binary classification task, where each binary classifier is able to understand the presence of other emotions, without directly inferring them. This is achieved by incorporating the proposed emotion anchors (i.e. features of representative emotional phrases) into the model of each binary classifier. More specifically, the model is trained to incorporate other emotions in its representation by learning the parameters of an attention mechanism. Based on experiments on multiple datasets, our approach improves emotion classification performance in both supervised and few-shot domain adaptation settings, outperforming standard binary models in terms of accuracy and macro averaged F1-scores. The approach is generic and can be applied to other interrelated multi-label binary classification tasks.</div></div>","PeriodicalId":100944,"journal":{"name":"Natural Language Processing Journal","volume":"12 ","pages":"Article 100170"},"PeriodicalIF":0.0,"publicationDate":"2025-09-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144489871","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}