Machine Translation (MT) is widely employed to address resource scarcity in low-resource languages by translating data from high-resource languages. While sentiment preservation in translation has long been studied, a critical but underexplored factor is the role of cultural alignment between source and target languages. In this paper, we hypothesize that semantic labels drift or are altered during MT due to cultural divergence. Through a series of experiments across culturally sensitive and neutral domains, we establish three key findings: (1) MT systems, including modern Large Language Models (LLMs), induce label drift during translation, particularly in culturally sensitive domains; (2) unlike earlier statistical MT tools, LLMs encode cultural knowledge, and leveraging this knowledge can amplify label drift; and (3) cultural similarity or dissimilarity between source and target languages is a crucial determinant of label preservation. Our findings highlight that neglecting cultural factors in MT not only undermines label fidelity but also risks misinterpretation and cultural conflict in downstream applications. We release our codebase to facilitate future research in cross-cultural translation: https://github.com/mohsinulkabir14/label_drift

When multiple normative frameworks yield equally valid answers, which framework should a language model select, and can it answer correctly within it? We formalize this as normative pluralism and study it in Islamic finance with a four-choice taxonomy that decomposes framework selection from within-framework correctness. The decomposition exposes a failure mode invisible to two-choice instruments, the stereotype trap: a model leans toward the culturally appropriate framework but produces factually wrong answers within it. Across twelve models, two languages, and fifty demographic signals, cultural cues redirect framework selection at severe cost: redirection and correctness are inversely coupled (r=−0.89 Arabic, r=−0.81 English). Under the strongest signal, large open-weight models reach 97% Islamic-frame selection, a result two-choice instruments would report as near-perfect alignment, yet 57-66% of those responses misquote the Islamic rule. We argue that cultural default bias is competence-conditioned routing: models default to the framework where they are most accurate, and the cues prescribed as mitigation expose the competence gap they were meant to correct.

Large language models (LLMs) have been widely applied across various domains of finance. Since their training data are largely derived from human-authored corpora, LLMs may inherit a range of human biases. Behavioral biases can lead to instability and uncertainty in decision-making, particularly when processing financial information. However, existing research on LLM bias has mainly focused on direct questioning or simplified, general-purpose settings, with limited consideration of the complex real-world financial environments and high-risk, context-sensitive, multilingual financial misinformation detection tasks (MFMD). In this work, we propose MFMDScen, a comprehensive benchmark for evaluating behavioral biases of LLMs in MFMD across diverse economic scenarios. In collaboration with financial experts, we construct three types of complex financial scenarios: (i) role- and personality-based, (ii) role- and region-based, and (iii) role-based scenarios incorporating ethnicity and religious beliefs. We further develop a multilingual financial misinformation dataset covering English, Chinese, Greek, and Bengali. By integrating these scenarios with misinformation claims, MFMDScen enables a systematic evaluation of 22 mainstream LLMs. Our findings reveal that pronounced behavioral biases persist across both commercial and open-source models. This project is available at https://github.com/lzw108/FMD.

Cross-cultural competence in large language models (LLMs) requires understanding and adapting Culture-Specific Items (CSIs) across varying cultural contexts. However, progress in evaluating this capability remains limited by the lack of high-quality CSI-annotated corpora with parallel cross-cultural sentence pairs. We introduce XCR-Bench, a Cross(X)-Cultural Reasoning Benchmark containing 4.1k parallel sentences and 1,098 CSIs across three reasoning tasks. XCR-Bench integrates Newmark’s CSI framework with Hall’s Triad of Culture, enabling evaluation across levels of cultural visibility: from observable practices to implicit social norms and values. Experiments on eight multilingual LLMs show that state-of-the-art models exhibit consistent weaknesses in identifying and adapting specific categories of CSIs, revealing a gap between surface-level recall and explicit cultural reasoning. Performance declines significantly in culturally sensitive categories and at deeper cultural levels (p<0.005, 8/8 models). We further find that adaptation quality varies systematically across target cultures and across Bengali regional variants, revealing systematic regional lexical and cultural preference patterns and ethno-religious biases in model outputs, even within a single linguistic setting. We publicly release the corpus and code to support future research on cross-cultural NLP.

Large language models are increasingly regarded as strong translators, yet it remains unclear what their success establishes beyond overall translation quality. This question is especially consequential in Chinese web novel translation, where sentence-level cultural meaning shapes tone, characterization, and reader experience. Existing evaluation provides only indirect evidence, as benchmarks emphasize overall quality and available metrics do not explicitly assess the cultural and pragmatic phenomena underlying reader experience. We introduce DITING (a mythological creature known for discerning truth), the first comprehensive evaluation framework for this problem, supported by an expert-annotated benchmark of over 18K Chinese-English sentence pairs, a reasoning-driven multi-agent evaluator, AgentEval, and a 1,440-pair meta-evaluation benchmark, MetricAlign. AgentEval achieves the strongest correlation with human judgments among eight representative automatic metrics. Applying DITING to 14 translation systems shows that current progress is selective rather than general: frontier LLMs lead overall, but their gains concentrate in a subset of dimensions, while zero-pronoun resolution remains the clearest bottleneck. Overall, our results show that aggregate rankings substantially overstate how broadly sentence-level cultural meaning has been solved in web novel translation. All the code and data in https://github.com/whunextgen/DITING.

Despite the promise of large language models (LLMs) in finance, their capabilities for financial misinformation detection (FMD) remain largely unexplored. To evaluate the capabilities of LLMs in FMD task, we introduce the financial misinformation detection shared task featured at COLING FinNLP-FNP-LLMFinLegal-2024, FMD Challenge. This challenge aims to evaluate the ability of LLMs to verify financial misinformation while generating plausible explanations. In this paper, we provide an overview of this task and dataset, summarize participants’ methods, and present their experimental evaluations, highlighting the effectiveness of LLMs in addressing the FMD task. To the best of our knowledge, the FMD Challenge is one of the first challenges for assessing LLMs in the field of FMD. Therefore, we provide detailed observations and draw conclusions for the future development of this field.

Current research on bias in language models (LMs) predominantly focuses on data quality, with significantly less attention paid to model architecture and temporal influences of data. Even more critically, few studies systematically investigate the origins of bias. We propose a methodology grounded in comparative behavioral theory to interpret the complex interaction between training data and model architecture in bias propagation during language modeling. Building on recent work that relates transformers to n-gram LMs, we evaluate how data, model design choices, and temporal dynamics affect bias propagation. Our findings reveal that: (1) n-gram LMs are highly sensitive to context window size in bias propagation, while transformers demonstrate architectural robustness; (2) the temporal provenance of training data significantly affects bias; and (3) different model architectures respond differentially to controlled bias injection, with certain biases (e.g. sexual orientation) being disproportionately amplified. As language models become ubiquitous, our findings highlight the need for a holistic approach- tracing bias to its origins across both data and model dimensions, not just symptoms, to mitigate harm.

A large number of studies rely on closed-style multiple-choice surveys to evaluate cultural alignment in Large Language Models (LLMs). In this work, we challenge this constrained evaluation paradigm and explore more realistic, unconstrained approaches. Using the World Values Survey (WVS) and Hofstede Cultural Dimensions as case studies, we demonstrate that LLMs exhibit stronger cultural alignment in less constrained settings, where responses are not forced. Additionally, we show that even minor changes, such as reordering survey choices, lead to inconsistent outputs, exposing the limitations of closed-style evaluations. Our findings advocate for more robust and flexible evaluation frameworks that focus on specific cultural proxies, encouraging more nuanced and accurate assessments of cultural alignment in LLMs.

Large Language Models (LLMs) have emerged as one of the most important breakthroughs in natural language processing (NLP) for their impressive skills in language generation and other language-specific tasks. Though LLMs have been evaluated in various tasks, mostly in English, they have not yet undergone thorough evaluation in under-resourced languages such as Bengali (Bangla). To this end, this paper introduces BenLLM-Eval, which consists of a comprehensive evaluation of LLMs to benchmark their performance in the low-resourced Bangla language. In this regard, we select various important and diverse Bangla NLP tasks, such as text summarization, question answering, paraphrasing, natural language inference, text classification, and sentiment analysis for zero-shot evaluation of popular LLMs, namely, ChatGPT, LLaMA-2, and Claude-2. Our experimental results demonstrate that while in some Bangla NLP tasks, zero-shot LLMs could achieve performance on par, or even better than current SOTA fine-tuned models; in most tasks, their performance is quite poor (with the performance of open-source LLMs like LLaMA-2 being significantly bad) in comparison to the current SOTA results. Therefore, it calls for further efforts to develop a better understanding of LLMs in low-resource languages like Bangla.

The analysis of consumer sentiment, as expressed through reviews, can provide a wealth of insight regarding the quality of a product. While the study of sentiment analysis has been widely explored in many popular languages, relatively less attention has been given to the Bangla language, mostly due to a lack of relevant data and cross-domain adaptability. To address this limitation, we present BanglaBook, a large-scale dataset of Bangla book reviews consisting of 158,065 samples classified into three broad categories: positive, negative, and neutral. We provide a detailed statistical analysis of the dataset and employ a range of machine learning models to establish baselines including SVM, LSTM, and Bangla-BERT. Our findings demonstrate a substantial performance advantage of pre-trained models over models that rely on manually crafted features, emphasizing the necessity for additional training resources in this domain. Additionally, we conduct an in-depth error analysis by examining sentiment unigrams, which may provide insight into common classification errors in under-resourced languages like Bangla. Our codes and data are publicly available at https://github.com/mohsinulkabir14/BanglaBook.

This paper introduces a novel informal Bangla word embedding for designing a cost-efficient solution for the task “Violence Inciting Text Detection” which focuses on developing classification systems to categorize violence that can potentially incite further violent actions. We propose a semi-supervised learning approach by training an informal Bangla FastText embedding, which is further fine-tuned on lightweight models on task specific dataset and yielded competitive results to our initial method using BanglaBERT, which secured the 7th position with an f1-score of 73.98%. We conduct extensive experiments to assess the efficiency of the proposed embedding and how well it generalizes in terms of violence classification, along with it’s coverage on the task’s dataset. Our proposed Bangla IFT embedding achieved a competitive macro average F1 score of 70.45%. Additionally, we provide a detailed analysis of our findings, delving into potential causes of misclassification in the detection of violence-inciting text.

The art of mathematical reasoning stands as a fundamental pillar of intellectual progress and is a central catalyst in cultivating human ingenuity. Researchers have recently published a plethora of works centered around the task of solving Math Word Problems (MWP) — a crucial stride towards general AI. These existing models are susceptible to dependency on shallow heuristics and spurious correlations to derive the solution expressions. In order to ameliorate this issue, in this paper, we propose a framework for MWP solvers based on the generation of linguistic variants of the problem text. The approach involves solving each of the variant problems and electing the predicted expression with the majority of the votes. We use DeBERTa (Decoding-enhanced BERT with disentangled attention) as the encoder to leverage its rich textual representations and enhanced mask decoder to construct the solution expressions. Furthermore, we introduce a challenging dataset, ParaMAWPS, consisting of paraphrased, adversarial, and inverse variants of selectively sampled MWPs from the benchmark Mawps dataset. We extensively experiment on this dataset along with other benchmark datasets using some baseline MWP solver models. We show that training on linguistic variants of problem statements and voting on candidate predictions improve the mathematical reasoning and robustness of the model. We make our code and data publicly available.

Sarcasm is a form of figurative language that serves as a humorous tool for mockery and ridicule. We present a novel architecture for sarcasm generation with emoji from a non-sarcastic input sentence in English. We divide the generation task into two sub tasks: one for generating textual sarcasm and another for collecting emojis associated with those sarcastic sentences. Two key elements of sarcasm are incorporated into the textual sarcasm generation task: valence reversal and semantic incongruity with context, where the context may involve shared commonsense or general knowledge between the speaker and their audience. The majority of existing sarcasm generation works have focused on this textual form. However, in the real world, when written texts fall short of effectively capturing the emotional cues of spoken and face-to-face communication, people often opt for emojis to accurately express their emotions. Due to the wide range of applications of emojis, incorporating appropriate emojis to generate textual sarcastic sentences helps advance sarcasm generation. We conclude our study by evaluating the generated sarcastic sentences using human judgement. All the codes and data used in this study has been made publicly available.