COLING 2025 论文列表

Object categories are typically organized into a multi-granularity taxonomic hierarchy. When classifying categories at different hierarchy levels, traditional uni-modal approaches focus primarily on image features, revealing limitations in complex scenarios. Recent studies integrating Vision-Language Models (VLMs) with class hierarchies have shown promise, yet they fall short of fully exploiting the hierarchical relationships. These efforts are constrained by their inability to perform effectively across varied granularity of categories. To tackle this issue, we propose a novel framework (**HGCLIP**) that effectively combines **CLIP** with a deeper exploitation of the **H**ierarchical class structure via **G**raph representation learning. We explore constructing the class hierarchy into a graph, with its nodes representing the textual or image features of each category. After passing through a graph encoder, the textual features incorporate hierarchical structure information, while the image features emphasize class-aware features derived from prototypes through the attention mechanism. Our approach demonstrates significant improvements on 11 diverse visual recognition benchmarks. Our codes are fully available at https: //github.com/richard-peng-xia/HGCLIP.

Retrieval-Augmented Generation (RAG) has emerged as a common paradigm to use Large Language Models (LLMs) alongside private and up-to-date knowledge bases. In this work, we address the challenges of using LLM-as-a-Judge when evaluating grounded answers generated by RAG systems. To assess the calibration and discrimination capabilities of judge models, we identify 7 generator failure modes and introduce GroUSE (Grounded QA Unitary Scoring of Evaluators), a meta-evaluation benchmark of 144 unit tests. This benchmark reveals that existing automated RAG evaluation frameworks often overlook important failure modes, even when using GPT-4 as a judge. To improve on the current design of automated RAG evaluation frameworks, we propose a novel pipeline and find that while closed models perform well on GroUSE, state-of-the-art open-source judges do not generalize to our proposed criteria, despite strong correlation with GPT-4’s judgement. Our findings suggest that correlation with GPT-4 is an incomplete proxy for the practical performance of judge models and should be supplemented with evaluations on unit tests for precise failure mode detection. We further show that finetuning Llama-3 on GPT-4’s reasoning traces significantly boosts its evaluation capabilities, improving upon both correlation with GPT-4’s evaluations and calibration on reference situations

Complex Table Question Answering involves providing accurate answers to specific questions based on intricate tables that exhibit complex layouts and flexible header locations. Despite considerable progress having been made in the LLM era, the reasoning processes of existing methods are often implicit, feeding the entire table into prompts, making it difficult to effectively filter out irrelevant information in the table. To this end, we propose GraphOTTER that explicitly establishes the reasoning process to pinpoint the correct answers. In particular, GraphOTTER leverages a graph-based representation, transforming the complex table into an undirected graph. It then conducts step-by-step reasoning on the graph, with each step guided by a set of pre-defined intermediate reasoning actions. As such, it constructs a clear reasoning path and effectively identifies the answer to a given question. Comprehensive experiments on two benchmark datasets and two LLM backbones demonstrate the effectiveness of GraphOTTER. Further analysis indicates that its success may be attributed to the ability to efficiently filter out irrelevant information, thereby focusing the reasoning process on the most pertinent data. Our code and experimental datasets are available at https://github.com/JDing0521/GraphOTTER.

In the open-domain multi-document summarization (ODMDS) task, retrieving relevant documents from large repositories and generating coherent summaries are crucial. However, existing methods often treat retrieval and summarization as separate tasks, neglecting the relationships among documents. To address these limitations, we propose an integrated retrieval-summarization framework that captures global document relationships through graph-based clustering, guiding the re-ranking of retrieved documents. This cluster-level thematic information is then used to guide large language models (LLMs) in refining the retrieved documents and generating more accurate, coherent summaries. Experimental results on the ODSUM benchmark demonstrate that our method significantly improves retrieval accuracy and produces summaries that surpass those derived from the oracle documents. These findings highlight the potential of our framework to improve both retrieval and summarization tasks in ODMDS.

Graph representation learning (GRL) in hyperbolic space has gradually emerged as a promising approach. Meanwhile, masking and reconstruction-based (MR-based) methods lead to state-of-the-art self-supervised graph representation. However, existing MR-based methods do not fully consider deep node and structural information. Inspired by the recent active and emerging field of self-supervised learning, we propose a novel node and edge dual-masked self-supervised graph representation learning framework in hyperbolic space, named HDM-GAE. We have designed a graph dual-masked module and a hyperbolic structural self-attention encoder module to mask nodes or edges and perform node aggregation within hyperbolic space, respectively. Comprehensive experiments and ablation studies on real-world multi-category datasets, demonstrate the superiority of our method in downstream tasks such as node classification and link prediction.

Massive open online courses (MOOCs) offer free education globally. Despite this democratization of learning, the massive enrollment in these courses makes it impractical for an instructor to assess every student’s writing assignment. As a result, peer grading, often guided by a straightforward rubric, is the method of choice. While convenient, peer grading often falls short in terms of reliability and validity. In this study, we explore the feasibility of using large language models (LLMs) to replace peer grading in MOOCs. To this end, we adapt the zero-shot chain-of-thought (ZCoT) prompting technique to automate the feedback process once the LLM assigns a score to an assignment. Specifically, to instruct LLMs for grading, we use three distinct prompts based on ZCoT: (1) ZCoT with instructor-provided correct answers, (2) ZCoT with both instructor-provided correct answers and rubrics, and (3) ZCoT with instructor-provided correct answers and LLM-generated rubrics. We tested these prompts in 18 different scenarios using two LLMs—GPT-4 and GPT-3.5—across three MOOCs: Introductory Astronomy, Astrobiology, and the History and Philosophy of Astronomy. Our results show that ZCoT, when augmented with instructor-provided correct answers and rubrics, produces grades that are more aligned with those assigned by instructors compared to peer grading. Finally, our findings indicate a promising potential for automated grading systems in MOOCs, especially in subjects with well-defined rubrics, to improve the learning experience for millions of online learners worldwide.

Federated learning has emerged as a potential solution to overcome the bottleneck posed by the near exhaustion of public text data in training large language models. There are claims that the strategy of exchanging gradients allows using text data including private information. Although recent studies demonstrate that data can be reconstructed from gradients, the threat for text data seems relatively small due to its sensitivity to even a few token errors. However, we propose a novel attack method FET, indicating that it is possible to Fully Expose Text data from gradients. Unlike previous methods that optimize continuous embedding vectors, we directly search for a text sequence with gradients that match the known gradients. First, we infer the total number of tokens and the unique tokens in the target text data from the gradients of the embedding layer. Then we develop a discrete optimization algorithm, which globally explores the solution space and precisely refines the obtained solution, incorporating both global and local search strategies. We also find that gradients of the fully connected layer are dominant, providing sufficient guidance for the optimization process. Our experiments show a significant improvement in attack performance, with an average increase of 39% for TinyBERT-6, 20% for BERT-base and 15% for BERT-large in exact match rates across three datasets. These findings highlight serious privacy risks in text data, suggesting that using smaller models is not an effective privacy-preserving strategy.

Backdoor attacks remain significant security threats to generative large language models (LLMs). Since generative LLMs output sequences of high-dimensional token logits instead of low-dimensional classification logits, most existing backdoor defense methods designed for discriminative models like BERT are ineffective for generative LLMs. Inspired by the observed differences in learning behavior between backdoor and clean mapping in the frequency space, we transform gradients of each training sample, directly influencing parameter updates, into the frequency space. Our findings reveal a distinct separation between the gradients of backdoor and clean samples in the frequency space. Based on this phenomenon, we propose Gradient Clustering in the Frequency Space for Backdoor Sample Filtering (GraCeFul), which leverages sample-wise gradients in the frequency space to effectively identify backdoor samples without requiring retraining LLMs. Experimental results show that GraCeFul outperforms baselines significantly. Notably, GraCeFul exhibits remarkable computational efficiency, achieving nearly 100% recall and F1 scores in identifying backdoor samples, reducing the average success rate of various backdoor attacks to 0% with negligible drops in clean accuracy across multiple free-style question answering datasets. Additionally, GraCeFul generalizes to Llama-2 and Vicuna. The codes are publicly available at https://github.com/ZrW00/GraceFul.

Evaluating the graph comprehension and reasoning abilities of Large Language Models (LLMs) is challenging and often incomplete. Existing benchmarks focus primarily on pure graph understanding, lacking a comprehensive evaluation across all graph types and detailed capability definitions. This paper presents GraCoRe, a benchmark for systematically assessing LLMs’ graph comprehension and reasoning. GraCoRe uses a three-tier hierarchical taxonomy to categorize and test models on pure graph and heterogeneous graphs, subdividing capabilities into 10 distinct areas tested through 19 tasks. Our benchmark includes 11 datasets with 5,140 graphs of varying complexity. We evaluate four closed-source and eight open-source LLMs, conducting thorough analyses from both ability and task perspectives. Key findings reveal that OpenAI o1 model has amazing comprehension and reasoning capabilities, semantic enrichment enhances reasoning performance, node ordering impacts task success, and the ability to process longer texts does not necessarily improve graph comprehension or reasoning.

Large Language Models (LLMs) have been found to memorize and recite some of the textual sequences from their training set verbatim, raising broad concerns about privacy and copyright issues. This Textual Sequence Memorization (TSM) phenomenon leads to a high demand to regulate LLM output to prevent generating certain memorized text that a user wants to be forgotten. However, our empirical study reveals that existing methods for TSM erasure fail to unlearn large numbers of memorized samples without substantially jeopardizing the model utility. To achieve a better trade-off between the effectiveness of TSM erasure and model utility in LLMs, our paper proposes a new method, named Entropy Maximization with Selective Optimization (EMSO), where the model parameters are updated sparsely based on novel optimization and selection criteria, in a manner that does not require additional models or data other than that in the forget set. More specifically, we propose an entropy-based loss that is shown to lead to more stable optimization and better preserves model utility than existing methods. In addition, we propose a contrastive gradient metric that takes both the gradient magnitude and direction into consideration, so as to localize model parameters to update in a sparse model updating scehme. Extensive experiments across three model scales demonstrate that our method excels in handling large-scale forgetting requests while preserving model ability in language generation and understanding.

Geocoding customer addresses and determining precise locations is a crucial component for any e-commerce company. Shipment delivery costs make up a significant portion of overall expenses, and having exact customer locations not only improves operational efficiency but also reduces costs and enhances the customer experience. While state-of-the-art geocoding systems are well-suited for developed countries with structured city layouts and high-quality reference corpora, they are less effective in developing countries like India, where addresses are highly unstructured and reliable reference data is scarce. Recent research has focused on creating geocoding systems tailored for developing nations such as India. In this work, we propose a method to geocode addresses in such environments. We explored various approaches to incorporate geo-spatial relationships using an LLM backbone, which provided insights into how the model learns these relationships both explicitly and implicitly. Our proposed approach outperforms the current state-of-the-art system by 20% in drift accuracy within 100 meters, and the state-of-the-art commercial system by 54%. This has a potential to reduce the incorrect delivery hub assignments by 8% which leads to significant customer experience improvements and business savings.

Generic sentences express generalisations about the world without explicit quantification. Although generics are central to everyday communication, building a precise semantic framework has proven difficult, in part because speakers use generics to generalise properties with widely different statistical prevalence. In this work, we study the implicit quantification and context-sensitivity of generics by leveraging language models as models of language. We create ConGen, a dataset of 2873 naturally occurring generic and quantified sentences in context, and define p-acceptability, a metric based on surprisal that is sensitive to quantification. Our experiments show generics are more context-sensitive than determiner quantifiers and about 20% of naturally occurring generics we analyze express weak generalisations. We also explore how human biases in stereotypes can be observed in language models.

In this work, we introduce the Keep Emotional and Essential Memory (KEEM) dataset, a novel generation-based dataset designed to enhance memory updates in long-term conversational systems. Unlike existing approaches that rely on simple accumulation or operation-based methods, which often result in information conflicts and difficulties in accurately tracking a user’s current state, KEEM dynamically generates integrative memories. This process not only preserves essential factual information but also incorporates emotional context and causal relationships, enabling a more nuanced understanding of user interactions. By seamlessly updating a system’s memory with both emotional and essential data, our approach promotes deeper empathy and enhances the system’s ability to respond meaningfully in open-domain conversations.

Document-level event argument extraction is a crucial task that aims to extract arguments from the entire document, beyond sentence-level analysis. Prior classification-based models still fail to explicitly capture significant relationships and heavily relies on large-scale datasets. In this study, we propose a novel approach called Generation-Augmented and Embedding Fusion. This approach first uses predefined templates and generative language models to produce an embedding capturing role relationship information, then integrates it into the foundational embedding derived from a classification model through a noval embedding fusion mechanism. We conduct the extensive experiments on the RAMS and WikiEvents datasets to demonstrate that our approach is more effective than the baselines, and that it is also data-efficient in low-resource scenarios.

This paper studies the task of generating commonsense reasoning questions (QG) with desired difficulty levels. Compared to traditional shallow questions that can be solved by simple term matching, ours are more challenging. Our answering process requires reasoning over multiple contextual and commonsense clues. That involves advanced comprehension skills, such as abstract semantics learning and missing knowledge inference. Existing work mostly learns to map the given text into questions, lacking a mechanism to control results with the desired complexity. To address this problem, we propose a novel controllable framework. We first derive contextual and commonsense clues involved in reasoning questions from the text. These clues are used to create simple sub-questions. We then aggregate multiple sub-questions to compose complex ones under the guidance of prior reasoning structures. By iterating this process, we can compose a complex QG task based on a series of smaller and simpler QG subtasks. Each subtask serves as a building block for a larger one. Each composition corresponds to an increase in the reasoning step. Moreover, we design a voting verifier to ensure results’ validity from multiple views, including answer consistency, reasoning difficulty, and context correlation. Finally, we can learn the optimal QG model to yield thought-provoking results. Evaluations on two typical datasets validate our method.

Entities are important to understanding literary works, which emphasize characters, plots and environment. The research on entity recognition, especially nested entity recognition in the literary domain is still insufficient partly due to insufficient annotated data. To address this issue, we construct the first Genre-oriented Corpus for Entity Recognition in Chinese Web Novels, namely GenWebNovel, comprising 400 chapters totaling 1,214,283 tokens under two genres, XuanHuan (Eastern Fantasy) and History. Based on the corpus, we analyze the distribution of different types of entities, including person, location, and organization. We also compare the nesting patterns of nested entities between GenWebNovel and the English corpus LitBank. Even though both belong to the literary domain, entities in different genres share few overlaps, making genre adaptation of NER (Named Entity Recognition) a hard problem. We propose a novel method that utilizes a pre-trained language model as an In-context learning example retriever to boost the performance of large language models. Our experiments show that this approach significantly enhances entity recognition, matching state-of-the-art (SOTA) models without requiring additional training data. Our code, dataset, and model are available at https://github.com/hjzhao73/GenWebNovel.

With the prevalence of Large Language Models (LLMs), recent studies have shifted paradigms and leveraged LLMs to tackle the challenging task of Text-to-SQL. Because of the complexity of real world databases, previous works adopt the retrieve-then-generate framework to retrieve relevant database schema and then to generate the SQL query. However, efficient embedding-based retriever suffers from lower retrieval accuracy, and more accurate LLM-based retriever is far more expensive to use, which hinders their applicability for broader applications. To overcome this issue, this paper proposes Gen-SQL, a novel generate-ground-regenerate framework, where we exploit prior knowledge from the LLM to enhance embedding-based retriever and reduce cost. Experiments on several datasets are conducted to demonstrate the effectiveness and scalability of our proposed method. We release our code and data at https://github.com/jieshi10/gensql.

We present GR-NLP-TOOLKIT, an open-source natural language processing (NLP) toolkit developed specifically for modern Greek. The toolkit provides state-of-the-art performance in five core NLP tasks, namely part-of-speech tagging, morphological tagging, dependency parsing, named entity recognition, and Greeklish-to-Greek transliteration. The toolkit is based on pre-trained Transformers, it is freely available, and can be easily installed in Python (pip install gr-nlp-toolkit). It is also accessible through a demonstration platform on HuggingFace, along with a publicly available API for non-commercial use. We discuss the functionality provided for each task, the underlying methods, experiments against comparable open-source toolkits, and future possible enhancements. The toolkit is available at: https://github.com/nlpaueb/gr-nlp-toolkit

Accurately recommending candidate news articles to users has always been the core challenge of news recommendation system. News recommendations often require modeling of user interest to match candidate news. Recent efforts have primarily focused on extracting local subgraph information in a global click graph constructed by the clicked news sequence of all users. However, the computational complexity of extracting global click graph information has hindered the ability to utilize far-reaching linkage which is hidden between two distant nodes in global click graph collaboratively among similar users. To overcome the problem above, we propose a Global-view Long Chain Interests Modeling for news recommendation (GLoCIM), which combines neighbor interest with long chain interest distilled from a global click graph, leveraging the collaboration among similar users to enhance news recommendation. We therefore design a long chain selection algorithm and long chain interest encoder to obtain global-view long chain interest from the global click graph. We design a gated network to integrate long chain interest with neighbor interest to achieve the collaborative interest among similar users. Subsequently we aggregate it with local news category-enhanced representation to generate final user representation. Then candidate news representation can be formed to match user representation to achieve news recommendation. Experimental results on real-world datasets validate the effectiveness of our method to improve the performance of news recommendation.