Hasil untuk "cs.CL"

Subhadip Mitra

Evaluating large language models at scale remains a practical bottleneck for many organizations. While existing evaluation frameworks work well for thousands of examples, they struggle when datasets grow to hundreds of thousands or millions of samples. This scale is common when assessing model behavior across diverse domains or conducting comprehensive regression testing. We present Spark-LLM-Eval, a distributed evaluation framework built natively on Apache Spark. The system treats evaluation as a data-parallel problem, partitioningexamplesacrossexecutorsandaggregatingresultswithproperstatistical accounting. Beyond raw throughput, we emphasize statistical rigor: every reported metric includes bootstrap confidence intervals, and model comparisons come with appropriate significance tests (paired t-tests, McNemar's test, or Wilcoxon signed-rank, depending on the metric type). The framework also addresses the cost problem inherent in LLM evaluation through content-addressable response caching backed by Delta Lake, which allows iterating on metric definitions without re-running inference. We describe the system architecture, the statistical methodology, and report benchmark results showing linear scaling with cluster size. The framework and all evaluation code are available as open source.

Hassan Barmandah

Large language models (LLMs) for Arabic are still dominated by Modern Standard Arabic (MSA), with limited support for Saudi dialects such as Najdi and Hijazi. This underrepresentation hinders their ability to capture authentic dialectal variation. Using a privately curated Saudi Dialect Instruction dataset (Hijazi and Najdi; 5,466 synthetic instruction-response pairs; 50/50 split), we LoRA-tune ALLaM-7B-Instruct-preview, the first foundation model developed in Saudi Arabia, for Saudi dialect generation. We investigate two variants: (i) Dialect-Token training, which prepends an explicit dialect tag to the instruction, and (ii) No-Token training, which omits the tag at formatting time. Evaluation on a held-out test set combines an external dialect classifier with text fidelity metrics (chrF++ and BERTScore) and diversity measures. The Dialect-Token model achieves the best control, raising the Saudi rate from 47.97% to 84.21% and reducing MSA leakage from 32.63% to 6.21%; fidelity also improves (chrF++ +3.53, BERTScore +0.059). Both LoRA variants outperform strong generic instruction models (Falcon-7B-Instruct, Llama-3.1-8B-Instruct, Qwen-2.5-7B-Instruct, AceGPT-v2-8B-Chat, JAIS-13B-Chat) in dialect control and fidelity, while avoiding metadata-tag echoing that these baselines frequently exhibit. We do not release the dataset or any model weights/adapters; instead, we release training/evaluation/inference code and a detailed datasheet (schema and aggregate statistics) to support independent verification.

Trishala Jayesh Ahalpara

We present Tell Me, a mental well-being system that leverages advances in large language models to provide accessible, context-aware support for users and researchers. The system integrates three components: (i) a retrieval-augmented generation (RAG) assistant for personalized, knowledge-grounded dialogue; (ii) a synthetic client-therapist dialogue generator conditioned on client profiles to facilitate research on therapeutic language and data augmentation; and (iii) a Well-being AI crew, implemented with CrewAI, that produces weekly self-care plans and guided meditation audio. The system is designed as a reflective space for emotional processing rather than a substitute for professional therapy. It illustrates how conversational assistants can lower barriers to support, complement existing care, and broaden access to mental health resources. To address the shortage of confidential therapeutic data, we introduce synthetic client-therapist dialogue generation conditioned on client profiles. Finally, the planner demonstrates an innovative agentic workflow for dynamically adaptive, personalized self-care, bridging the limitations of static well-being tools. We describe the architecture, demonstrate its functionalities, and report evaluation of the RAG assistant in curated well-being scenarios using both automatic LLM-based judgments and a human-user study. This work highlights opportunities for interdisciplinary collaboration between NLP researchers and mental health professionals to advance responsible innovation in human-AI interaction for well-being.

Toshiyuki Shigemura

This paper introduces the Luca-Noise Reflex Protocol (LN-RP), a computational framework for analyzing noise-driven persona emergence in large language models. By injecting stochastic noise seeds into the initial generation state, we observe nonlinear transitions in linguistic behavior across 152 generation cycles. Our results reveal three stable persona modes with distinct entropy signatures, and demonstrate that external noise sources can reliably induce phase transitions in reflexive generation dynamics. Quantitative evaluation confirms consistent persona retention and significant differences across modes (p < 0.01). The protocol provides a reproducible method for studying reflexive generation, emergent behavior, and longrange linguistic coherence in LLMs.

Kieron Kretschmar, Walter Laurito, Sharan Maiya et al.

Prior work has introduced techniques for detecting when large language models (LLMs) lie, that is, generate statements they believe are false. However, these techniques are typically validated in narrow settings that do not capture the diverse lies LLMs can generate. We introduce LIARS' BENCH, a testbed consisting of 72,863 examples of lies and honest responses generated by four open-weight models across seven datasets. Our settings capture qualitatively different types of lies and vary along two dimensions: the model's reason for lying and the object of belief targeted by the lie. Evaluating three black- and white-box lie detection techniques on LIARS' BENCH, we find that existing techniques systematically fail to identify certain types of lies, especially in settings where it's not possible to determine whether the model lied from the transcript alone. Overall, LIARS' BENCH reveals limitations in prior techniques and provides a practical testbed for guiding progress in lie detection.

Aashray Reddy, Andrew Zagula, Nicholas Saban

Large Language Models (LLMs) continue to exhibit vulnerabilities to jailbreaking attacks: carefully crafted malicious inputs intended to circumvent safety guardrails and elicit harmful responses. As such, we present AutoAdv, a novel framework that automates adversarial prompt generation to systematically evaluate and expose vulnerabilities in LLM safety mechanisms. Our approach leverages a parametric attacker LLM to produce semantically disguised malicious prompts through strategic rewriting techniques, specialized system prompts, and optimized hyperparameter configurations. The primary contribution of our work is a dynamic, multi-turn attack methodology that analyzes failed jailbreak attempts and iteratively generates refined follow-up prompts, leveraging techniques such as roleplaying, misdirection, and contextual manipulation. We quantitatively evaluate attack success rate (ASR) using the StrongREJECT (arXiv:2402.10260 [cs.CL]) framework across sequential interaction turns. Through extensive empirical evaluation of state-of-the-art models--including ChatGPT, Llama, and DeepSeek--we reveal significant vulnerabilities, with our automated attacks achieving jailbreak success rates of up to 86% for harmful content generation. Our findings reveal that current safety mechanisms remain susceptible to sophisticated multi-turn attacks, emphasizing the urgent need for more robust defense strategies.

Christoph Leiter, Jonas Belouadi, Yanran Chen et al.

The NLLG (Natural Language Learning & Generation) arXiv reports assist in navigating the rapidly evolving landscape of NLP and AI research across cs.CL, cs.CV, cs.AI, and cs.LG categories. This fourth installment captures a transformative period in AI history - from January 1, 2023, following ChatGPT's debut, through September 30, 2024. Our analysis reveals substantial new developments in the field - with 45% of the top 40 most-cited papers being new entries since our last report eight months ago and offers insights into emerging trends and major breakthroughs, such as novel multimodal architectures, including diffusion and state space models. Natural Language Processing (NLP; cs.CL) remains the dominant main category in the list of our top-40 papers but its dominance is on the decline in favor of Computer vision (cs.CV) and general machine learning (cs.LG). This report also presents novel findings on the integration of generative AI in academic writing, documenting its increasing adoption since 2022 while revealing an intriguing pattern: top-cited papers show notably fewer markers of AI-generated content compared to random samples. Furthermore, we track the evolution of AI-associated language, identifying declining trends in previously common indicators such as "delve".

Hoa Nguyen, Steffen Eger

Citations are a key ingredient of scientific research to relate a paper to others published in the community. Recently, it has been noted that there is a citation age bias in the Natural Language Processing (NLP) community, one of the currently fastest growing AI subfields, in that the mean age of the bibliography of NLP papers has become ever younger in the last few years, leading to `citation amnesia' in which older knowledge is increasingly forgotten. In this work, we put such claims into perspective by analyzing the bibliography of $\sim$300k papers across 15 different scientific fields submitted to the popular preprint server Arxiv in the time period from 2013 to 2022. We find that all AI subfields (in particular: cs.AI, cs.CL, cs.CV, cs.LG) have similar trends of citation amnesia, in which the age of the bibliography has roughly halved in the last 10 years (from above 12 in 2013 to below 7 in 2022), on average. Rather than diagnosing this as a citation age bias in the NLP community, we believe this pattern is an artefact of the dynamics of these research fields, in which new knowledge is produced in ever shorter time intervals.

Ralph Krüger

Due to the widespread use of data-driven neural machine translation, both by professional translators and layperson users, an adequate machine translation literacy on the part of the users of this technology is becoming more and more important. At the same time, the increasing datafication of both the private and the business sphere requires an adequate data literacy in modern society. The present article takes a closer look at machine translation literacy and data literacy and investigates the interface between the two concepts. This is done to lay the preliminary theoretical foundations for a didactic project aiming to develop didactic resources for teaching data literacy in its machine translation-specific form to students of BA programmes in translation/specialised communication.

Longshen Ou, Xichu Ma, Ye Wang

Despite progress in melody-to-lyric generation, a substantial singability gap remains between machine-generated lyrics and those written by human lyricists. In this work, we aim to narrow this gap by jointly learning both wording and formatting for melody-to-lyric generation. After general-domain pretraining, our model acquires length awareness through an self-supervised stage trained on a large text-only lyric corpus. During supervised melody-to-lyric training, we introduce multiple auxiliary supervision objective informed by musicological findings on melody--lyric relationships, encouraging the model to capture fine-grained prosodic and structural patterns. Compared with naïve fine-tuning, our approach improves adherence to line-count and syllable-count requirements by 3.8% and 21.4% absolute, respectively, without degrading text quality. In human evaluation, it achieves 42.2% and 74.2% relative gains in overall quality over two task-specific baselines, underscoring the importance of formatting-aware training for generating singable lyrics.

Damien Hansen, Emmanuelle Esperança-Rodier, Hervé Blanchon et al.

La traduction automatique neuronale et son adaptation à des domaines spécifiques par le biais de corpus spécialisés ont permis à cette technologie d’intégrer bien plus largement qu’auparavant le métier et la formation des traducteur·trice·s. Si le paradigme neuronal (et le deep learning de manière générale) a ainsi pu investir des domaines parfois insoupçonnés, y compris certains où la créativité est de mise, celui-ci est moins marqué par un gain phénoménal de performance que par une utilisation massive auprès du public et les débats qu’il génère, nombre d’entre eux invoquant couramment le cas littéraire pour (in)valider telle ou telle observation. Pour apprécier la pertinence de cette technologie, et ce faisant surmonter les discours souvent passionnés des opposants et partisans de la traduction automatique, il est toutefois nécessaire de mettre l’outil à l’épreuve, afin de fournir un exemple concret de ce que pourrait produire un système entraîné spécifiquement pour la traduction d’œuvres littéraires. Inscrit dans un projet de recherche plus vaste visant à évaluer l’aide que peuvent fournir les outils informatiques aux traducteurs et traductrices littéraires, cet article propose par conséquent une expérience de traduction automatique de la prose qui n’a plus été tentée pour le français depuis les systèmes probabilistes et qui rejoint un nombre croissant d’études sur le sujet pour d’autres paires de langues. Nous verrons que si les résultats sont encourageants, ceux-ci laissent présager une tout autre manière d’envisager la traduction automatique, plus proche de la traduction humaine assistée par ordinateur que de la post-édition pure, et que l’exemple des œuvres de littérature soulève en outre des réflexions utiles pour la traduction dans son ensemble.

Elissa Nakajima Wickham, Emily Öhman

This paper examines the shift in focus on content policies and user attitudes on the social media platform Reddit. We do this by focusing on comments from general Reddit users from five posts made by admins (moderators) on updates to Reddit Content Policy. All five concern the nature of what kind of content is allowed to be posted on Reddit, and which measures will be taken against content that violates these policies. We use topic modeling to probe how the general discourse for Redditors has changed around limitations on content, and later, limitations on hate speech, or speech that incites violence against a particular group. We show that there is a clear shift in both the contents and the user attitudes that can be linked to contemporary societal upheaval as well as newly passed laws and regulations, and contribute to the wider discussion on hate speech moderation.

Enrique Manjavacas, Lauren Fonteyn

As large language models such as BERT are becoming increasingly popular in Digital Humanities (DH), the question has arisen as to how such models can be made suitable for application to specific textual domains, including that of 'historical text'. Large language models like BERT can be pretrained from scratch on a specific textual domain and achieve strong performance on a series of downstream tasks. However, this is a costly endeavour, both in terms of the computational resources as well as the substantial amounts of training data it requires. An appealing alternative, then, is to employ existing 'general purpose' models (pre-trained on present-day language) and subsequently adapt them to a specific domain by further pre-training. Focusing on the domain of historical text in English, this paper demonstrates that pre-training on domain-specific (i.e. historical) data from scratch yields a generally stronger background model than adapting a present-day language model. We show this on the basis of a variety of downstream tasks, ranging from established tasks such as Part-of-Speech tagging, Named Entity Recognition and Word Sense Disambiguation, to ad-hoc tasks like Sentence Periodization, which are specifically designed to test historically relevant processing.

Lise Volkart, Sabrina Girletti, Johanna Gerlach et al.

We conducted an experiment with translation students to assess the influence of two different post-editing (PE) strategies (reading the source segment or the target segment first) on three aspects: PE time, ratio of corrected errors and number of optional modifications per word. Our results showed that the strategy that is adopted has no influence on the PE time or ratio of corrected errors. However, it does have an influence on the number of optional modifications per word. Two other thought-provoking observations emerged from this study: first, the ratio of corrected errors showed that, on average, students correct only half of the MT errors, which underlines the need for PE practice. Second, the time logs of the experiment showed that when students are not forced to read the source segment first, they tend to neglect the source segment and almost do monolingual PE. This experiment provides new insight relevant to PE teaching as well as the designing of PE environments.

Philippe Lacour, Aurélien Bénel

International audience TraduXio is a digital environment for computer assisted multilingual translation which is web-based, free to use and with an open source code. Its originality is threefold-whereas traditional technologies are limited to two languages (source/target), TraduXio enables the comparison of different versions of the same text in various languages; its concordancer provides relevant and multilingual suggestions through a classification of the source according to the history, genre and author; it uses collaborative devices (privilege management, forums, networks, history of modification, etc.) to promote collective (and distributed) translation. TraduXio is designed to encourage the diversification of language learning and to promote a reappraisal of translation as a professional skill. It can be used in many different ways, by very diverse kind of people. In this presentation, I will present the recent developments of the software (its version 2.1) and illustrate how specific groups (language teaching, social sciences, literature) use it on a regular basis. In this paper, I present the technology but concentrate more on the possible uses of TraduXio, thus focusing on translators' feedback about their experience when working in this digital environment in a truly collaborative way.

Philippe Lacour, Aurélien Bénel

TraduXio is a digital environment for computer assisted multilingual translation which is web-based, free to use and with an open source code. Its originality is threefold-whereas traditional technologies are limited to two languages (source/target), TraduXio enables the comparison of different versions of the same text in various languages; its concordancer provides relevant and multilingual suggestions through a classification of the source according to the history, genre and author; it uses collaborative devices (privilege management, forums, networks, history of modification, etc.) to promote collective (and distributed) translation. TraduXio is designed to encourage the diversification of language learning and to promote a reappraisal of translation as a professional skill. It can be used in many different ways, by very diverse kind of people. In this presentation, I will present the recent developments of the software (its version 2.1) and illustrate how specific groups (language teaching, social sciences, literature) use it on a regular basis. In this paper, I present the technology but concentrate more on the possible uses of TraduXio, thus focusing on translators' feedback about their experience when working in this digital environment in a truly collaborative way.

Abheesht Sharma, Gunjan Chhablani, Harshit Pandey et al.

In this work, we present to the NLP community, and to the wider research community as a whole, an application for the diachronic analysis of research corpora. We open source an easy-to-use tool coined: DRIFT, which allows researchers to track research trends and development over the years. The analysis methods are collated from well-cited research works, with a few of our own methods added for good measure. Succinctly put, some of the analysis methods are: keyword extraction, word clouds, predicting declining/stagnant/growing trends using Productivity, tracking bi-grams using Acceleration plots, finding the Semantic Drift of words, tracking trends using similarity, etc. To demonstrate the utility and efficacy of our tool, we perform a case study on the cs.CL corpus of the arXiv repository and draw inferences from the analysis methods. The toolkit and the associated code are available here: https://github.com/rajaswa/DRIFT.

Rami Botros, Tara N. Sainath, Robert David et al.

Previous works on the Recurrent Neural Network-Transducer (RNN-T) models have shown that, under some conditions, it is possible to simplify its prediction network with little or no loss in recognition accuracy (arXiv:2003.07705 [eess.AS], [2], arXiv:2012.06749 [cs.CL]). This is done by limiting the context size of previous labels and/or using a simpler architecture for its layers instead of LSTMs. The benefits of such changes include reduction in model size, faster inference and power savings, which are all useful for on-device applications. In this work, we study ways to make the RNN-T decoder (prediction network + joint network) smaller and faster without degradation in recognition performance. Our prediction network performs a simple weighted averaging of the input embeddings, and shares its embedding matrix weights with the joint network's output layer (a.k.a. weight tying, commonly used in language modeling arXiv:1611.01462 [cs.LG]). This simple design, when used in conjunction with additional Edit-based Minimum Bayes Risk (EMBR) training, reduces the RNN-T Decoder from 23M parameters to just 2M, without affecting word-error rate (WER).

Jani Marjanen, Jussi Kurunmäki, Lidia Pivovarova et al.

Words with the suffix-ism are reductionist terms that help us navigate complex social issues by using a simple one-word label for them. On the one hand they are often associated with political ideologies, but on the other they are present in many other domains of language, especially culture, science, and religion. This has not always been the case. This paper studies isms in a historical record of digitized newspapers from 1820 to 1917 published in Finland to find out how the language of isms developed historically. We use diachronic word embeddings and affinity propagation clustering to trace how new isms entered the lexicon and how they relate to one another over time. We are able to show how they became more common and entered more and more domains. Still, the uses of isms as traditions for political action and thinking stand out in our analysis.

Wanjiku Nganga, Ikechukwu Achebe

The preservation of languages is critical to maintaining and strengthening the cultures and identities of communities, and this is especially true for under-resourced languages with a predominantly oral culture. Most African languages have a relatively short literary past, and as such the task of dictionary making cannot rely on textual corpora as has been the standard practice in lexicography. This paper emphasizes the significance of the spoken word and the oral tradition as repositories of vocabulary, and argues that spoken word corpora greatly outweigh the value of printed texts for lexicography. We describe a methodology for creating a digital dialectal dictionary for the Igbo language from such a spoken word corpus. We also highlight the language technology tools and resources that have been created to support the transcription of thousands of hours of Igbo speech and the subsequent compilation of these transcriptions into an XML-encoded textual corpus of Igbo dialects. The methodology described in this paper can serve as a blueprint that can be adopted for other under-resourced languages that have predominantly oral cultures.