China

This paper describes the submission of the HITS team to the DISRPT 2025 shared task. The shared task includes three sub-tasks: (1) discourse unit segmentation across formalisms, (2) cross-lingual discourse connective identification, and (3) cross-formalism discourse relation classification. This paper presents our strategies for the DISRPT 2025 Shared Task. In Task 1, our approach involves fine-tuning through multilingual joint training on linguistically motivated language groups. We incorporated two key techniques to improve model performance: a weighted loss function to address the task&#39;s significant class imbalance and Fast Gradient Method (FGM) adversarial training to boost the model&#39;s robustness.

In task 2, our approach involves building an ensemble of three encoder models whose embeddings are smartly fused together with a multi-head attention layer. We also add Part-Of-Speech tags and dependency relations present in the training file as linguistic features. A CRF layer is added after the classification layer to account for dependencies between adjacent labels. To account for label imbalance, we use focal loss and label smoothing. This ensures our model is robust and flexible enough to handle different languages.

In task 3, we use two-stage fine-tuning framework designed to transfer the nuanced reasoning capabilities of a very large &quot;teacher&quot; model to a compact &quot;student&quot; model so that the smaller model can learn complex discourse relationships. The fine-tuning process follows a curriculum learning framework. In such a framework the model learns to perform increasingly harder tasks. In our case, the model first learns to look at the discourse units and then predict the label followed by looking at Chain-Of-Thought reasoning for harder examples. This way it can learn to internalise such reasoning and increase prediction accuracy on the harder samples.

EMNLP 2025

HITS at DISRPT 2025: Discourse Segmentation, Connective Detection, and Relation Classification

multilingual llm

discourse

classification

## Welcome!
"I am excited to welcome you to this year’s edition of the Conference on Empirical Methods in Natural Language Processing! Importantly, it marks the 30th edition of EMNLP. With over 8,000 submissions, more than 3,000 accepted papers, and thousands of attendees, we have come a long way from that first
workshop, which had 14 accepted papers. As the field looks ahead, Suzhou is the fitting location for celebrating this milestone: rooted in a long literary tradition, yet modern and forward-looking, and home to a large share of the NLP community."<br>

*Message from the General Chair, Dirk Hovy*





<br>

Celebrate 30 Years of EMNLP! 
EMNLP 2025 will be held in Suzhou, China from November 5th to November 9th, 2025.

This paper describes the submission of the HITS team to the DISRPT 2025 shared task. The shared task includes three sub-tasks: (1) discourse unit segmentation across formalisms, (2) cross-lingual discourse connective identification, and (3) cross-formalism discourse relation classification. This paper presents our strategies for the DISRPT 2025 Shared Task. In Task 1, our approach involves fine-tuning through multilingual joint training on linguistically motivated language groups. We incorporated two key techniques to improve model performance: a weighted loss function to address the task's significant class imbalance and Fast Gradient Method (FGM) adversarial training to boost the model's robustness.

In task 2, our approach involves building an ensemble of three encoder models whose embeddings are smartly fused together with a multi-head attention layer. We also add Part-Of-Speech tags and dependency relations present in the training file as linguistic features. A CRF layer is added after the classification layer to account for dependencies between adjacent labels. To account for label imbalance, we use focal loss and label smoothing. This ensures our model is robust and flexible enough to handle different languages.

In task 3, we use two-stage fine-tuning framework designed to transfer the nuanced reasoning capabilities of a very large "teacher" model to a compact "student" model so that the smaller model can learn complex discourse relationships. The fine-tuning process follows a curriculum learning framework. In such a framework the model learns to perform increasingly harder tasks. In our case, the model first learns to look at the discourse units and then predict the label followed by looking at Chain-Of-Thought reasoning for harder examples. This way it can learn to internalise such reasoning and increase prediction accuracy on the harder samples.

workshop paper

The work presented here describes our participation in DISRPT 2025 shared task in three tasks, Task1: Discourse Unit Segmentation across Formalisms, Task 2: Discourse Connective Identification across Languages and Task 3: Discourse Relation Classification across Formalisms. We have fine-tuned XLM-RoBERTa, a language model to address these three tasks. We have come up with one single multilingual language model for each task. Our system handles data in both the formats .conllu and .tok and different discourse formalisms. We have obtained encouraging results. The performance on test data in the three tasks is similar to the results obtained for the development data.

SeCoRel: Multilingual Discourse Analysis in DISRPT 2025

In this work we examine LLMs' ability to ask clarification questions in task-oriented dialogues that follow the asynchronous instruction-giver/instruction-follower format. We present a new corpus that combines two existing annotations of the Minecraft Dialogue Corpus --- one for reference and ambiguity in reference, and one for SDRT including clarifications --- into a single common format providing the necessary information to experiment with clarifications and their relation to ambiguity. With this corpus we compare LLM actions with original human-generated clarification questions, examining how both humans and LLMs act in the case of ambiguity. We find that there is only a weak link between ambiguity and humans producing clarification questions in these dialogues, and low correlation between humans and LLMs. Humans hardly ever produce clarification questions for referential ambiguity, but often do so for task-based uncertainty. Conversely, LLMs produce more clarification questions for referential ambiguity, but less so for task uncertainty. We question if LLMs' ability to ask clarification questions is predicated on their recent ability to simulate reasoning, and test this with different reasoning approaches, finding that reasoning does appear to increase question frequency and relevancy.

Referential ambiguity and clarification requests: comparing human and LLM behaviour

In this paper, we analyse coreference annotation of the German language, focussing on the phenomenon of simplification, that is, the tendency to use words and constructions that are assumed to be easier perceived, understood, or produced. Simplification is one of the tools used by language users in order to optimise communication effectively. We are interested in how simplification is reflected in coreference in two different language products exposed to the phenomena of simplification: simultaneous interpreting and Easy German. For this, we automatically annotate simplified texts with coreference. We then evaluate the outputs of automatic annotation. In addition, we also look into quantitative distributions of some coreference features. Our findings show that although the language products under analysis diverge in terms of simplification driving factors, they share some specific coreference features. We also show that this specificity may cause annotation errors in simplified language, e.g. in non-nominal or split antecedents.

Coreference in simplified German: Linguistic features and challenges of automatic annotation

We present the submissions of our team to the Unconstrained and LLM tracks of the Computational Models of Reference, Anaphora and Coreference (CRAC2025) shared task, where we ended respectively in the fifth and the first place, but nevertheless with similar scores: average CoNLL-F1 scores of 61.57 and 62.96 on the test set, but with very large differences in computational cost. Indeed, the classical pair-wise resolution system submitted to the Unconstrained track obtained similar performance but with less than 10\% of the computational cost. Reflecting on this fact, we point out problems that we ran into using generative AI to perform coreference resolution. We explain how the framework of text generation stands in the way of a reliable text-global coreference representation. Nonetheless, we realize there are many potential improvements of our LLM-system; we discuss them at the end of this article.

GLaRef@CRAC2025: Should we transform coreference resolution into a text generation task?

This paper presents our submission to the CRAC 2025 Shared Task on Multilingual Coreference Resolution in the LLM track. We propose a prompt-based few-shot coreference resolution system where the final inference is performed by Grok-3 using in-context learning. The core of our methodology is a difficulty- aware sample selection pipeline that leverages Gemini Flash 2.0 to compute semantic diffi- culty metrics, including mention dissimilarity and pronoun ambiguity. By identifying and selecting the most challenging training sam- ples for each language, we construct highly informative prompts to guide Grok-3 in predict- ing coreference chains and reconstructing zero anaphora. Our approach secured 3rd place in the CRAC 2025 shared task.

Few-Shot Coreference Resolution with Semantic Difficulty Metrics and In-Context Learning

This paper describes our approach to the CRAC 2025 Shared Task on Multilingual Coreference Resolution. We compete in the LLM track, where the systems are limited to generative text-to-text approaches. Our system is based on Llama 3.1-8B, fine-tuned to tag the document with coreference annotations. We have made one significant modification to the text format provided by the organizers: The model relies on the syntactic head for mention span representation. Additionally, we use joint pre-training, and we train the model to generate empty nodes. We provide an in-depth analysis of the performance of our models, which reveals several implementation problems. Although our system ended up in last place, we achieved the best performance on 10 datasets out of 22 within the track. By fixing the discovered problems in the post-evaluation phase, we improved our results substantially, outperforming all the systems in the LLM track and even some unconstrained track systems.

Fine-Tuned Llama for Multilingual Text-to-Text Coreference Resolution

In this work, we present our system, which ranked second in the CRAC 2025 Shared Task on Multilingual Coreference Resolution (LLM Track). For multilingual coreference resolution, our system mainly uses long-context large language models (LLMs) in a few-shot in-context learning setting. Among the various approaches we explored, few-shot prompting proved to be the most effective, particularly due to the complexity of the task and the availability of high-quality data with referential relationships provided as part of the competition. We employed Gemini 2.5 Pro, one of the best available closed-source long-context LLMs at the time of submission. Our system achieved a CoNLL F1 score of 61.74 on the mini-testset, demonstrating that performance improves significantly with the number of few-shot examples provided, thanks to the model's extended context window. While this approach comes with trade-offs in terms of inference cost and response latency, it highlights the potential of long-context LLMs for tackling multilingual coreference without task-specific fine-tuning. Although direct comparisons with traditional supervised systems are not straightforward, our findings provide valuable insights and open avenues for future work, particularly in expanding support for low-resource languages.

Few-Shot Multilingual Coreference Resolution Using Long-Context Large Language Models

We present CorPipe 25, the winning entry to the CRAC 2025 Shared Task on Multilingual Coreference Resolution. This fourth iteration of the shared task introduces a new LLM track alongside the original unconstrained track, features reduced development and test sets to lower computational requirements, and includes additional datasets. CorPipe 25 represents a complete reimplementation of our previous systems, migrating from TensorFlow to PyTorch. Our system significantly outperforms all other submissions in both the LLM and unconstrained tracks by a substantial margin of 8 percentage points. The source code and trained models are publicly available at https://github.com/ufal/crac2025-corpipe.

CorPipe at CRAC 2025: Evaluating Multilingual Encoders for Multilingual Coreference Resolution

Gundel et al.'s Givenness Hierarchy remains one of the most influental frameworks of Information Status to this date, and has been employed in different technical contexts to account for context-sensitive and hearer-tailored language in human-machine interaction and natural language processing as well as as a topic of linguistic inquiry. At the same time, the data basis upon which this theory has been developed remains relatively thin. Although its applicability to a broad array of languages has been repeatedly confirmed, the empirical evidence presented for certain phenomena, and in particular, with respect to demonstrative determiners and demonstrative pronouns did not always reach conventional levels of statistical significance. In this paper, we provide an empirical, corpus-based re-assessment of two seminal papers for the Givenness Hierarchy, Gundel et al. (1990) and Gundel et al. (1993), where we aim to replicate their findings on the basis of corpora with coreference annotation for their original sample of languages, i.e., Arabic, Chinese, English, Japanese, Korean, Russian and Spanish. We describe the operationalization of Gundel et al.'s `cognitive statuses', their approximation by means of anaphoric relations, the preprocessing of diverse and heterogeneous corpora and evaluate Gundel et al.'s claims. Our contribution is three-fold: We evaluate the Givenness Hierarchy against quantitative data at a scale that allows to assess statistical significance, we discuss challenges and problems encountered in the process, in the preprocessing and in the interpretation of the diverse corpora, we provide two generalizations: a procedure for bootstrapping Givenness Hierarchies for other languages, and possible cross-linguistically applicable tendencies in the systems of referring expressions.

Revisiting the Givenness Hierarchy. A Corpus-Based Evaluation

We tackle the task of mention detection for pair-programming dialogue, a setting which adds several challenges to the task due to the characteristics of natural dialogue, the dynamic environment of the dialogue task, and the domain-specific vocabulary and structures. We compare recent variants of the Llama and GPT families and explore different prompt and context engineering approaches. While aspects like hesitations and references to read-out code and variable names made the task challenging, GPT 4.1 approximated human performance when we provided few-shot examples similar to the inference text and corrected formatting errors.

Downloads

Next from EMNLP 2025

SeCoRel: Multilingual Discourse Analysis in DISRPT 2025

Similar lecture

SeCoRel: Multilingual Discourse Analysis in DISRPT 2025

.css-70qvj9{display:-webkit-box;display:-webkit-flex;display:-ms-flexbox;display:flex;-webkit-align-items:center;-webkit-box-align:center;-ms-flex-align:center;align-items:center;}Downloads

Next from EMNLP 2025

SeCoRel: Multilingual Discourse Analysis in DISRPT 2025

Similar lecture

SeCoRel: Multilingual Discourse Analysis in DISRPT 2025

Downloads