COLING 2025 Papers

The rise of large language models (LLMs) has created a need for advanced benchmarking systems beyond traditional setups. To this end, we introduce QUENCH, a novel text-based English Quizzing Benchmark manually curated and transcribed from YouTube quiz videos. QUENCH possesses masked entities and rationales for the LLMs to predict via generation. At the intersection of world knowledge, geographical context, and common sense reasoning, QUENCH helps assess world knowledge and deduction capabilities of LLMs via a zero-shot, open-domain quizzing setup. We perform an extensive evaluation on 7 LLMs and 4 metrics, investigating the influence of model size, prompting style, geographical context, and gold-labeled rationale generation. The benchmarking concludes with an error analysis of various types of generative errors to which the LLMs are prone.

In this paper, we introduce QABISAR, a novel framework for statutory article retrieval, to overcome the semantic mismatch problem when modeling each query-article pair in isolation, making it hard to learn representation that can effectively capture multi-faceted information. QABISAR leverages bipartite interactions between queries and articles to capture diverse aspects inherent in them. Further, we employ knowledge distillation to transfer enriched query representations from the graph network into the query bi-encoder, to capture the rich semantics present in the graph representations, despite absence of graph-based supervision for unseen queries during inference. Our experiments on a real-world expert-annotated dataset demonstrate its effectiveness.

Chinese Semantic Error Recognition (CSER) has always been a weak link in Chinese language processing due to the complexity and obscureness of Chinese semantics. Existing research has gradually focused on leveraging pre-trained models to perform CSER. Although some researchers have attempted to integrate syntax information into the pre-trained language model, it requires training the models from scratch, which is time-consuming and laborious. Furthermore, despite the existence of datasets for CSER, the constrained size of these datasets impairs the performance of the models. Thus, in order to address the difficulty posed by a limited sample set and the need of annotating samples with semantic-level errors, we propose a Pseudo-label Data Construction method for CSER (PDC-CSER), generating pseudo-labels for augmented samples based on perplexity and model respectively, which overcomes the difficulty of constructing pseudo-label data containing semantic-level errors and ensures the quality of pseudo-labels. Moreover, we propose a CSER method with the Dependency Syntactic Attention mechanism (CSER-DSA) to explicitly infuse dependency syntactic information only in the fine-tuning stage, achieving robust performance, and simultaneously reducing substantial computing power and time cost. Results demonstrate that the pseudo-label technology PDC-CSER and the semantic error recognition method CSER-DSA surpass the existing models

Text-to-speech (TTS) has seen significant advancements in high-quality, expressive speech synthesis. However, achieving diverse and natural prosody in synthesized speech remains challenging. In this paper, we propose ProsodyFlow, an end-to-end TTS model that integrates large self-supervised speech models and conditional flow matching to model prosodic features effectively. Our approach involves using a speech LLM to extract acoustic features, mapping these features into a prosody latent space, and then employing conditional flow matching to generate prosodic vectors conditioned on the input text. Experiments on the LJSpeech dataset show that ProsodyFlow improves synthesis quality and efficiency compared to existing models, achieving more prosodic and expressive speech synthesizing.

The rapid advancements in Large Language Models (LLMs) have revolutionized natural language processing (NLP) and adjacent fields, yet fine-tuning these models for specific tasks remains computationally expensive and risks degrading pre-learned features. To address these challenges, we propose Propulsion, a novel parameter-efficient fine-tuning (PEFT) method designed to optimize task-specific performance while drastically reducing computational overhead. Inspired by the concept of controlled adjustments in physical motion, Propulsion selectively re-scales specific dimensions of a pre-trained model, guiding output predictions toward task objectives without modifying the model’s parameters. By introducing lightweight, trainable Propulsion parameters at the pre-trained layer, we minimize the number of parameters updated during fine-tuning, thus preventing the overfitting or overwriting of existing knowledge. Our theoretical analysis, supported by Neural Tangent Kernel (NTK) theory, shows that Propulsion approximates the performance of full fine-tuning with far fewer trainable parameters. Empirically, Propulsion reduces the parameter count from 355.3 million to a mere 0.086 million—achieving over a 10x reduction compared to standard approaches like LoRA—while maintaining competitive performance across benchmarks.

Propaganda plays a critical role in shaping public opinion and fueling disinformation. While existing research primarily focuses on identifying propaganda techniques, it lacks the ability to capture the broader motives and the impacts of such content. To address these challenges, we introduce PropaInsight, a conceptual framework grounded in foundational social science research, which systematically dissects propaganda into techniques, arousal appeals, and underlying intent. PropaInsight offers a more granular understanding of how propaganda operates across different contexts. Additionally, we present PropaGaze, a novel dataset that combines human-annotated data with high-quality synthetic data generated through a meticulously designed pipeline. Our experiments show that off-the-shelf LLMs struggle with propaganda analysis, but PropaGaze significantly improves performance. Fine-tuned Llama-7B-Chat achieves 203.4% higher text span IoU in technique identification and 66.2% higher BertScore in appeal analysis compared to 1-shot GPT-4-Turbo. Moreover, PropaGaze complements limited human-annotated data in data-sparse and cross-domain scenarios, demonstrating its potential for comprehensive and generalizable propaganda analysis.

Recent advancements in large language models (LLMs) have significantly enhanced their coding capabilities. However, existing benchmarks predominantly focused on simplified or isolated aspects of coding, such as single-file code generation or repository issue debugging, falling short of measuring the full spectrum of challenges raised by real-world programming activities. In this case study, we explore the performance of LLMs across the entire software development lifecycle with DevEval, encompassing stages including software design, environment setup, implementation, acceptance testing, and unit testing. DevEval features four programming languages, multiple domains, high-quality data collection, and carefully designed and verified metrics for each task. Empirical studies show that current LLMs, including GPT-4, fail to solve the challenges presented within DevEval. Our findings offer actionable insights for the future development of LLMs toward real-world programming applications.

Large Language Model (LLM)-based Task-Oriented Dialogue (TOD) systems show promising performance in helping users achieve specific goals in a zero-shot setting. However, existing systems engage with users in a reactive manner, relying on a basic single-query mechanism with the knowledge base and employing passive policy planning. The proactive TOD systems, which can provide potentially helpful information and plan cross-domain multi-task dialogue policies, have not been well studied. In addition, effective evaluation methods are also lacking. To address these issues, we propose ProTOD, a novel LLM-based proactive TOD framework designed to improve system proactivity and goal completion. First, we design an adaptive exploratory retrieval mechanism to dynamically navigate domain knowledge. Second, we introduce a two-stage passive-to-proactive policy planner that effectively organizes knowledge and actions relationship. Finally, we develop two distinct user simulators with different personalities to simulate real-world interactions and propose a new error measure called Human-targeted Policy Edit Rate (HPER) for evaluation. Experimental results show that ProTOD achieves state-of-the-art (SOTA) performance, improving goal completion rates by 10% while significantly enhancing the proactive engagement.

Activation sparsity refers to the existence of considerable weakly-contributed elements among activation outputs, serving as a promising paradigm for accelerating model inference. Nevertheless, most large language models (LLMs) adopt activation functions without intrinsic activation sparsity (e.g., GELU and Swish). Some recent efforts have explored introducing ReLU or its variants as the substitutive activation function to pursue activation sparsity and acceleration, but few can simultaneously obtain high activation sparsity and comparable model performance. This paper introduces a simple and effective method named “ProSparse” to sparsify LLMs while achieving both targets. Specifically, after introducing ReLU activation, ProSparse adopts progressive sparsity regularization with a factor smoothly increasing for multiple stages. This can enhance activation sparsity and mitigate performance degradation by avoiding radical shifts in activation distributions. With ProSparse, we obtain high sparsity of 89.32% for LLaMA2-7B, 88.80% for LLaMA2-13B, and 87.89% for end-size MiniCPM-1B, respectively, with comparable performance to their original Swish-activated versions. These present the most sparsely activated models among open-source LLaMA versions and competitive end-size models. Inference acceleration experiments further demonstrate the significant practical acceleration potential of LLMs with higher activation sparsity, obtaining up to 4.52x inference speedup.

This paper proposes a method to estimate the performance of pretrained models fine-tuned with a larger dataset from the result with a smaller dataset. Specifically, we demonstrate that when a pretrained model is fine-tuned, its classification performance increases at the same overall rate, regardless of the original dataset size, as the number of epochs increases. Subsequently, we verify that an approximate formula based on this trend can be used to predict the performance when the model is trained with ten times or more training data, even when the initial training dataset is limited. Our results show that this approach can help resource-limited companies develop machine-learning models.

Addressing the disparity between predictions and actual results can enable individuals to expand their thought processes and stimulate self-reflection, thus promoting accurate planning. In this research, we present **PreAct**, an agent framework that integrates **pre**diction, **rea**soning, and **act**ion. By utilizing the information derived from predictions, the large language model (LLM) agent can provide a wider range and more strategically focused reasoning. This leads to more efficient actions that aid the agent in accomplishing intricate tasks. Our experimental results show that PreAct surpasses the ReAct method in completing complex tasks and that PreAct’s performance can be further improved when paired with other memory or selection strategy techniques. We presented the model with varying quantities of historical predictions and discovered that these predictions consistently enhance LLM planning. The variances in single-step reasoning between PreAct and ReAct indicate that PreAct indeed has benefits in terms of diversity and strategic orientation over ReAct.

Nowadays, research of Text Classification (TC) based on graph neural networks (GNNs) is on the rise. Both inductive methods and transductive methods have made significant progress. For transductive methods, the semantic interaction between texts plays a crucial role in the learning of effective text representations. However, it is difficult to perform inductive learning while modeling interactions between texts on the graph. To give a universal solution, we propose the graph neural network based on pre-trained semantic interaction called PaSIG. Firstly, we construct a text-word heterogeneity graph and design an asymmetric structure to ensure one-way message passing from words to the test texts. Meanwhile, we use the context representation capability of the pre-trained language model to construct node features that contain classification semantic information. Afterward, we explore the adaptative aggregation methods with a gated fusion mechanism. Extensive experiments on five datasets have shown the effectiveness of PaSIG, with the accuracy exceeding the baseline by 2.7% on average. While achieving state-of-the-art performance, we have also taken measures of subgraph sampling and intermediate state preservation to achieve fast inference.

This work introduces PrahokBART, a compact pre-trained sequence-to-sequence model trained from scratch for Khmer using carefully curated Khmer and English corpora. We focus on improving the pre-training corpus quality and addressing the linguistic issues of Khmer, which are ignored in existing multilingual models, by incorporating linguistic components such as word segmentation and normalization. We evaluate PrahokBART on three generative tasks: machine translation, text summarization, and headline generation, where our results demonstrate that it outperforms mBART50, a strong multilingual pre-trained model. Additionally, our analysis provides insights into the impact of each linguistic module and evaluates how effectively our model handles space during text generation, which is crucial for the naturalness of texts in Khmer.

Research on text simplification has been ongoing for many years, yet document simplification remains a significant challenge due to the need to address complex factors such as technical terminology, metaphors, and overall coherence. In this work, we introduce a novel multi-agent framework AgentSimp for document simplification, based on large language models. This framework simulates the collaborative efforts of a team of human experts through the roles played by multiple agents, effectively meeting the intricate demands of document simplification. We investigate two communication strategies among agents (pipeline-style and synchronous) and two document reconstruction strategies (Direct and Iterative). According to both automatic evaluation metrics and human evaluation results, AgentSimp produces simplified documents that are more thoroughly simplified and more coherent across various articles and styles.

Differing from sentiment transfer, positive reframing seeks to substitute negative perspectives with positive expressions while preserving the original meaning. With the emergence of pre-trained language models (PLMs), it is possible to achieve acceptable results by fine-tuning PLMs. Nevertheless, generating fluent, diverse and task-constrained reframing text remains a significant challenge. To tackle this issue, a **m**ulti-**s**trategy **o**ptimization **f**ramework (MSOF) is proposed in this paper. Starting from the objective of positive reframing, we first design positive sentiment reward and content preservation reward to encourage the model to transform the negative expressions of the original text while ensuring the integrity and consistency of the semantics. Then, different decoding optimization approaches are introduced to improve the quality of text generation. Finally, based on the modeling formula of positive reframing, we propose a multi-dimensional re-ranking method that further selects candidate sentences from three dimensions: strategy consistency, text similarity and fluency. Extensive experiments on two Seq2Seq PLMs, BART and T5, demonstrate our framework achieves significant improvements on unconstrained and controlled positive reframing tasks.

Transformers with causal attention can solve tasks that require positional information without using positional encodings. In this work, we propose and investigate a new hypothesis about how positional information can be stored without using explicit positional encoding. We observe that nearby embeddings are more similar to each other than faraway embeddings, allowing the transformer to potentially reconstruct the positions of tokens. We show that this pattern can occur in both the trained and the randomly initialized Transformer models with causal attention and no positional encodings over a common range of hyperparameters.

This study examines how Transformer language models utilise lexico-phrasal information to interpret the polysemy of the Korean adverbial postposition -(u)lo. We analysed the attention weights of both a Korean pre-trained BERT model and a fine-tuned version. Results show a general reduction in attention weights following fine-tuning, alongside changes in the lexico-phrasal information used, depending on the specific function of -(u)lo. These findings suggest that, while fine-tuning broadly affects a model’s syntactic sensitivity, it may also alter its capacity to leverage lexico-phrasal features according to the function of the target word.

The growing volume of scientific literature in polymer science presents a significant challenge for researchers attempting to extract and annotate domain-specific entities, such as polymer names, material properties, and related information. Manual annotation of these documents is both time-consuming and prone to error due to the complexity of scientific language. To address this, we introduce PolyMinder, an automated support system designed to assist polymer scientists in extracting and annotating polymer-related entities and their relationships from scientific documents. The system utilizes recent advanced Named Entity Recognition (NER) and Relation Extraction (RE) models tailored to the polymer domain. PolyMinder streamlines the annotation process by providing a web-based interface where users can visualize, verify, and refine the extracted information before finalizing the annotations. The system’s source code is made publicly available to facilitate further research and development in this field. Our system can be accessed through the following URL: https://www.jaist.ac.jp/is/labs/nguyen-lab/systems/polyminder

The task of text-to-image generation has encountered significant challenges when applied to literary works, especially poetry. Poems are a distinct form of literature, with meanings that frequently transcend beyond the literal words. To address this shortcoming, we propose a PoemToPixel framework designed to generate images that visually represent the inherent meanings of poems. Our approach incorporates the concept of prompt tuning in our image generation framework to ensure that the resulting images closely align with the poetic content. In addition, we propose the PoeKey algorithm, which extracts three key elements in the form of emotions, visual elements, and themes from poems to form instructions which are subsequently provided to a diffusion model for generating corresponding images. Furthermore, to expand the diversity of the poetry dataset across different genres and ages, we introduce MiniPo, a novel multimodal dataset comprising 1001 children’s poems and images. Leveraging this dataset alongside PoemSum, we conducted both quantitative and qualitative evaluations of image generation using our PoemToPixel framework. This paper demonstrates the effectiveness of our approach and offers a fresh perspective on generating images from literary sources. The code and dataset used in this work are publicly available.