Alignment techniques often inadvertently induce sycophancy in LLMs.While prior studies studied this behaviour in direct-answer settings, the role of Chain-of-Thought (CoT) reasoning remains under-explored: does it serve as a logical constraint that mitigates sycophancy, or a tool for post-hoc rationalization that masks it?We evaluate a range of models across objective and subjective tasks to investigate the issue.Results show that reasoning generally reduces sycophancy in final decisions but also masks sycophancy in some samples, where models construct deceptive justifications through logical inconsistencies, calculation errors, and onesided arguments, etc.Furthermore, LLMs are more prone to sycophancy in subjective tasks and under authority-bias.Our mechanistic analysis on three open-source models reveals that the tendency of sycophancy is dynamic during the reasoning process rather than being predetermined at the input stage 1 .Nostalgebraist.

Do large language models (LLMs) truly acquire embodied cognition and cultural conventions from text?We introduce demonstratives, fundamental spatial expressions like "this/that" in English and "这/那" in Chinese, as a novel probe for grounded knowledge.Using 6,400 responses from 320 native speakers, we establish a human baseline: English speakers reliably distinguish proximal-distal referents but struggle with perspective-taking, while Chinese speakers switch perspectives fluently but tolerate distal ambiguity.In contrast, five state-ofthe-art LLMs fail to inherently understand the proximal-distal contrast and show no cultural differences, defaulting to English-centric reasoning.Our study contributes (i) a new task, based on demonstratives, as a new lens for evaluating embodied cognition and cultural conventions; (ii) empirical evidence of cross-cultural asymmetries in human interpretation; (iii) a new perspective on the egocentric-sociocentric debate, showing both orientations coexist but vary across languages; and (iv) a call to address individual variation in future model design.Model Gpt5.

Autoregressive Large Language Models (LLMs) demonstrate exceptional performance in language understanding and generation. However, their application in text embedding tasks has been relatively slow, along with the analysis of their semantic representation in probing tasks, due to the constraints of the unidirectional attention mechanism. This paper aims to explore whether such constraints can be overcome by enabling bidirectional attention in LLMs. We tested different variants of the Llama architecture through additional training steps, progressively enabling bidirectional attention and unsupervised/supervised contrastive learning. Our results show that bidirectional attention improves the LLMs’ ability to represent subsequent context but weakens their utilization of preceding context, while contrastive learning training can help to maintain both abilities.

Effectively resolving phonological ambiguities is crucial for robust natural language processing, as these ambiguities are pervasive in tasks ranging from speech-to-text, spelling correction, to offensive language detection.However, current Large Language Models (LLMs) frequently struggle to resolve such ambiguities.To address this challenge, we present a framework to enhances LLMs' phonological capability through a multiple-stage training approach.Our method begins with supervised fine-tuning on well-constructed datasets, including three subtask datasets designed to enhance the model's foundational phonological knowledge, along with a synthetic dataset of step-by-step reasoning chains.Following this, we apply reinforcement learning to incentivize and stabilize its reasoning.Results show that our framework enables the base model to achieve relatively comparable performance to a much larger model.Our ablation studies reveal that subtask datasets and the synthetic dataset can simultaneously impact as complementary modular enhancers to strengthen LLMs' integrated application 1 .

Genta Indra Winata, Frederikus Hudi, Patrick Amadeus Irawan, David Anugraha, Rifki Afina Putri, Wang Yutong, Adam Nohejl, Ubaidillah Ariq Prathama, Nedjma Ousidhoum, Afifa Amriani, Anar Rzayev, Anirban Das, Ashmari Pramodya, Aulia Adila, Bryan Wilie, Candy Olivia Mawalim, Cheng Ching Lam, Daud Abolade, Emmanuele Chersoni, Enrico Santus, Fariz Ikhwantri, Garry Kuwanto, Hanyang Zhao, Haryo Akbarianto Wibowo, Holy Lovenia, Jan Christian Blaise Cruz, Jan Wira Gotama Putra, Junho Myung, Lucky Susanto, Maria Angelica Riera Machin, Marina Zhukova, Michael Anugraha, Muhammad Farid Adilazuarda, Natasha Christabelle Santosa, Peerat Limkonchotiwat, Raj Dabre, Rio Alexander Audino, Samuel Cahyawijaya, Shi-Xiong Zhang, Stephanie Yulia Salim, Yi Zhou, Yinxuan Gui, David Ifeoluwa Adelani, En-Shiun Annie Lee, Shogo Okada, Ayu Purwarianti, Alham Fikri Aji, Taro Watanabe, Derry Tanti Wijaya, Alice Oh, Chong-Wah Ngo. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

Noun-noun compounds interpretation is the task where a model is given one of such constructions, and it is asked to provide a paraphrase, making the semantic relation between the nouns explicit, as in carrot cake is "a cake made of carrots."Such a task requires the ability to understand the implicit structured representation of the compound meaning.In this paper, we test to what extent the recent Large Language Models can interpret the semantic relation between the constituents of lexicalized English compounds and whether they can abstract from such semantic knowledge to predict the semantic relation between the constituents of similar but novel compounds by relying on analogical comparisons (e.g., carrot dessert).We test both Surprisal metrics and prompt-based methods to see whether i.) they can correctly predict the relation between constituents, and ii.) the semantic representation of the relation is robust to paraphrasing.Using a dataset of lexicalized and annotated noun-noun compounds, we find that LLMs can infer some semantic relations better than others (with a preference for compounds involving concrete concepts).When challenged to perform abstractions and transfer their interpretations to semantically similar but novel compounds, LLMs show serious limitations 1 .

The field of Distributional Semantics has recently undergone important changes, with the contextual representations produced by Transformers taking the place of static word embeddings models. Noticeably, previous studies comparing the two types of vectors have only focused on the English language and a limited number of models. In our study, we present a comparative evaluation of static and contextualized distributional models for Mandarin Chinese, focusing on a range of intrinsic tasks. Our results reveal that static models remain stronger for some of the classical tasks that consider word meaning independent of context, while contextualized models excel in identifying semantic relations between word pairs and in the categorization of words into abstract semantic classes.

Financial sentiment analysis plays a pivotal role in understanding market dynamics and investor sentiment. In this paper, we propose the Supervised Cross-Momentum Contrast (SuCroMoCo) framework, a novel approach for financial sentiment analysis. SuCroMoCo leverages supervised contrastive learning and cross-momentum contrast to align financial text representations with prototypical representations based on sentiment categories. This alignment greatly improves classification performance, addressing the limitations of pre-trained language models (PLMs) in fully grasping the intricate nature of financial text. Through extensive experiments, we demonstrate that SuCroMoCo outperforms existing PLMs-based approaches and Large Language Models (LLMs) on diverse benchmark datasets. The code and datasets can be found in https://github.com/PengBO-O/SuCroMoCo.

Ellipsis is a linguistic phenomenon characterized by the omission of one or more sentence elements. Solving such a linguistic construction is not a trivial issue in natural language processing since it involves the retrieval of non-overtly expressed verbal material, which might in turn require the model to integrate human-like syntactic and semantic knowledge. In this paper, we explored the issue of how the prototypicality of event participants affects the ability of Language Models (LMs) to handle elliptical sentences and to identify the omitted arguments at different degrees of thematic fit, ranging from highly typical participants to semantically anomalous ones. With this purpose in mind, we built ELLie, the first dataset composed entirely of utterances containing different types of elliptical constructions, and structurally suited for evaluating the effect of argument thematic fit in solving ellipsis and reconstructing the missing element. Our tests demonstrated that the probability scores assigned by the models are higher for typical events than for atypical and impossible ones in different elliptical contexts, confirming the influence of prototypicality of the event participants in interpreting such linguistic structures. Finally, we conducted a retrieval task of the elided verb in the sentence in which the low performance of LMs highlighted a considerable difficulty in reconstructing the correct event

Adverse Drug Event (ADE) extraction is one of the core tasks in digital pharmacovigilance, especially when applied to informal texts. This task has been addressed by the Natural Language Processing community using large pre-trained language models, such as BERT. Despite the great number of Transformer-based architectures used in the literature, it is unclear which of them has better performances and why. Therefore, in this paper we perform an extensive evaluation and analysis of 19 Transformer-based models for ADE extraction on informal texts. We compare the performance of all the considered models on two datasets with increasing levels of informality (forums posts and tweets). We also combine the purely Transformer-based models with two commonly-used additional processing layers (CRF and LSTM), and analyze their effect on the models performance. Furthermore, we use a well-established feature importance technique (SHAP) to correlate the performance of the models with a set of features that describe them: model category (AutoEncoding, AutoRegressive, Text-to-Text), pre-training domain, training from scratch, and model size in number of parameters. At the end of our analyses, we identify a list of take-home messages that can be derived from the experimental data.

Medical term normalization consists in mapping a piece of text to a large number of output classes.Given the small size of the annotated datasets and the extremely long tail distribution of the concepts, it is of utmost importance to develop models that are capable to generalize to scarce or unseen concepts.An important attribute of most target ontologies is their hierarchical structure.In this paper we introduce a simple and effective learning strategy that leverages such information to enhance the generalizability of both discriminative and generative models.The evaluation shows that the proposed strategy produces state-of-the-art performance on seen concepts and consistent improvements on unseen ones, allowing also for efficient zero-shot knowledge transfer across text typologies and datasets.

Beatrice Portelli, Edoardo Lenzi, Emmanuele Chersoni, Giuseppe Serra, Enrico Santus. Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume. 2021.

Word embeddings are vectorial semantic representations built with either counting or predicting techniques aimed at capturing shades of meaning from word co-occurrences. Since their introduction, these representations have been criticized for lacking interpretable dimensions. This property of word embeddings limits our understanding of the semantic features they actually encode. Moreover, it contributes to the “black box” nature of the tasks in which they are used, since the reasons for word embedding performance often remain opaque to humans. In this contribution, we explore the semantic properties encoded in word embeddings by mapping them onto interpretable vectors, consisting of explicit and neurobiologically motivated semantic features (Binder et al. 2016). Our exploration takes into account different types of embeddings, including factorized count vectors and predict models (Skip-Gram, GloVe, etc.), as well as the most recent contextualized representations (i.e., ELMo and BERT). In our analysis, we first evaluate the quality of the mapping in a retrieval task, then we shed light on the semantic features that are better encoded in each embedding type. A large number of probing tasks is finally set to assess how the original and the mapped embeddings perform in discriminating semantic categories. For each probing task, we identify the most relevant semantic features and we show that there is a correlation between the embedding performance and how they encode those features. This study sets itself as a step forward in understanding which aspects of meaning are captured by vector spaces, by proposing a new and simple method to carve human-interpretable semantic representations from distributional vectors.

Abstract Machine learning methods, especially deep learning models, have achieved impressive performance in various natural language processing tasks including sentiment analysis. However, deep learning models are more demanding for training data. Data augmentation techniques are widely used to generate new instances based on modifications to existing data or relying on external knowledge bases to address annotated data scarcity, which hinders the full potential of machine learning techniques. This paper presents our work using part‐of‐speech (POS) focused lexical substitution for data augmentation (PLSDA) to enhance the performance of machine learning algorithms in sentiment analysis. We exploit POS information to identify words to be replaced and investigate different augmentation strategies to find semantically related substitutions when generating new instances. The choice of POS tags as well as a variety of strategies such as semantic‐based substitution methods and sampling methods are discussed in detail. Performance evaluation focuses on the comparison between PLSDA and two previous lexical substitution‐based data augmentation methods, one of which is thesaurus‐based, and the other is lexicon manipulation based. Our approach is tested on five English sentiment analysis benchmarks: SST‐2, MR, IMDB, Twitter, and AirRecord. Hyperparameters such as the candidate similarity threshold and number of newly generated instances are optimized. Results show that six classifiers (SVM, LSTM, BiLSTM‐AT, bidirectional encoder representations from transformers [BERT], XLNet, and RoBERTa) trained with PLSDA achieve accuracy improvement of more than 0.6% comparing to two previous lexical substitution methods averaged on five benchmarks. Introducing POS constraint and well‐designed augmentation strategies can improve the reliability of lexical data augmentation methods. Consequently, PLSDA significantly improves the performance of sentiment analysis algorithms.

While neural embeddings represent a popular choice for word representation in a wide variety of NLP tasks, their usage for thematic fit modeling has been limited, as they have been reported to lag behind syntax-based count models. In this paper, we propose a complete evaluation of count models and word embeddings on thematic fit estimation, by taking into account a larger number of parameters and verb roles and introducing also dependency-based embeddings in the comparison. Our results show a complex scenario, where a determinant factor for the performance seems to be the availability to the model of reliable syntactic information for building the distributional representations of the roles.

Automatic Chinese irony detection is a challenging task, and it has a strong impact on linguistic research. However, Chinese irony detection often lacks labeled benchmark datasets. In this paper, we introduce Ciron, the first Chinese benchmark dataset available for irony detection for machine learning models. Ciron includes more than 8.7K posts, collected from Weibo, a micro blogging platform. Most importantly, Ciron is collected with no pre-conditions to ensure a much wider coverage. Evaluation on seven different machine learning classifiers proves the usefulness of Ciron as an important resource for Chinese irony detection.

Abstract Most compositional distributional semantic models represent sentence meaning with a single vector. In this paper, we propose a structured distributional model (SDM) that combines word embeddings with formal semantics and is based on the assumption that sentences represent events and situations. The semantic representation of a sentence is a formal structure derived from discourse representation theory and containing distributional vectors. This structure is dynamically and incrementally built by integrating knowledge about events and their typical participants, as they are activated by lexical items. Event knowledge is modelled as a graph extracted from parsed corpora and encoding roles and relationships between participants that are represented as distributional vectors. SDM is grounded on extensive psycholinguistic research showing that generalized knowledge about events stored in semantic memory plays a key role in sentence comprehension.We evaluate SDMon two recently introduced compositionality data sets, and our results show that combining a simple compositionalmodel with event knowledge constantly improves performances, even with dif ferent types of word embeddings.

In this paper, we introduce a new distributional method for modeling predicateargument thematic fit judgments.We use a syntax-based DSM to build a prototypical representation of verb-specific roles: for every verb, we extract the most salient second order contexts for each of its roles (i.e. the most salient dimensions of typical role fillers), and then we compute thematic fit as a weighted overlap between the top features of candidate fillers and role prototypes.Our experiments show that our method consistently outperforms a baseline re-implementing a state-of-theart system, and achieves better or comparable results to those reported in the literature for the other unsupervised systems.Moreover, it provides an explicit representation of the features characterizing verbspecific semantic roles.

Several studies on sentence processing suggest that the mental lexicon keeps track of the mutual expectations between words.Current DSMs, however, represent context words as separate features, thereby loosing important information for word expectations, such as word interrelations.In this paper, we present a DSM that addresses this issue by defining verb contexts as joint syntactic dependencies.We test our representation in a verb similarity task on two datasets, showing that joint contexts achieve performances comparable to single dependencies or even better.Moreover, they are able to overcome the data sparsity problem of joint feature spaces, in spite of the limited size of our training corpus.