International audience

Late-interaction models such as ColBERT offer competitive performance across various retrieval tasks but require storing a dense embedding for each document token, leading to a substantial index storage overhead. Past works address this by attempting to prune low-importance token embeddings based on statistical and empirical measures, but they often either lack formal grounding or are ineffective. To address these shortcomings, we introduce a framework grounded in hyperspace geometry and cast token pruning as a Voronoi cell estimation problem in the embedding space. By interpreting each token's influence as a measure of its Voronoi region, our approach enables principled pruning that retains retrieval quality while reducing index size. Through our experiments, we demonstrate that this approach serves not only as a competitive pruning strategy but also as a valuable tool for improving and interpreting token-level behavior within dense retrieval systems.

Multi-label text classification (MLTC) is an essential task in NLP applications.Traditional methods require extensive labeled data and are limited to fixed label sets.Extracting labels with large language models (LLMs) is more effective and universal, but incurs high computational costs.In this work, we introduce a distillation-based T5 generalist model for zero-shot MLTC and few-shot fine-tuning.Our model accommodates variable label sets with general domain-agnostic pretraining, while modeling dependency between labels.Experiments show that our approach outperforms baselines of similar size on three few-shot tasks.Our code is available at repository.

In this paper, we propose a novel method for joint entity and relation extraction from unstructured text by framing it as a conditional sequence generation problem. In contrast to conventional generative information extraction models that are left-to-right token-level generators, our approach is \textit{span-based}. It generates a linearized graph where nodes represent text spans and edges represent relation triplets. Our method employs a transformer encoder-decoder architecture with pointing mechanism on a dynamic vocabulary of spans and relation types. Our model can capture the structural characteristics and boundaries of entities and relations through span representations while simultaneously grounding the generated output in the original text thanks to the pointing mechanism. Evaluation on benchmark datasets validates the effectiveness of our approach, demonstrating competitive results. Code is available at https://github.com/urchade/ATG.

Information extraction (IE) from text documents is an important NLP task that includes entity, relation, and event extraction. These tasks are often addressed jointly as a graph generation problem, where entities and event triggers represent nodes and where relations and event arguments represent edges. Most existing systems use local classifiers for nodes and edges, trained using cross-entropy loss, and employ inference strategies such as beam search to approximate the optimal graph structure. These approaches typically suffer from exposure bias due to the discrepancy between training and decoding. In this paper, we tackle this problem by casting graph generation as auto-regressive sequence labeling and making its training aware of the decoding procedure by using a differentiable version of beam search. We evaluate the effectiveness of our approach through extensive experiments conducted on the ACE05 and ConLL04 datasets across diverse languages. Our experimental findings affirm that our model outperforms its non-decoding-aware version for all datasets employed. Furthermore, we conduct ablation studies that emphasize the effectiveness of aligning training and inference. Additionally, we introduce a novel quantification of exposure bias within this context, providing valuable insights into the functioning of our model.

As pretrained language model emerge and consistently develop, prompt-based training has become a well-studied paradigm to improve the exploitation of models for many natural language processing tasks. Furthermore, prompting demonstrates great performance compared to conventional fine-tuning in scenarios with limited annotated data, such as zero-shot or few-shot situations. Verbalizers are crucial in this context, as they help interpret masked word distributions generated by language models into output predictions. This study introduces a benchmarking approach to assess three common baselines of verbalizers for topic classification in few-shot learning scenarios. Additionally, we find that increasing the number of label words for automatic label word searching enhances model performance. Moreover, we investigate the effectiveness of template assembling with various aggregation strategies to develop stronger classifiers that outperform models trained with individual templates. Our approach achieves comparable results to prior research while using significantly fewer resources. Our code is available at https://github.com/quang-anh-nguyen/verbalizer_benchmark.git.

Urchade Zaratiana, Nadi Tomeh, Pierre Holat, Thierry Charnois. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2024.

We present a parsing model for projective dependency trees which takes advantage of the existence of complementary dependency annotations for a language.This is the case for Arabic with the availability of CATiB and UD treebanks.Our system performs syntactic parsing according to both annotation types jointly as a sequence of arc-creating operations following the Easy-First approach, and partially created trees for one annotation type are also available to the other as features for the score function.This method gives error reduction of 9.9% on CATiB and 6.1% on UD compared to a single-task baseline, and ablation tests show that the main contribution of this reduction is given by sharing tree representation between tasks, and not simply sharing BiLSTM layers as is usually performed in NLP multitask systems.

We explore the consequences of representing token segmentations as hierarchical structures (trees) for the task of Multiword Expression (MWE) recognition, in isolation or in combination with dependency parsing.We propose a novel representation of token segmentation as trees on tokens, resembling dependency trees.Given this new representation, we present and evaluate two different architectures to combine MWE recognition and dependency parsing in the easy-first framework: a pipeline and a joint system, both taking advantage of lexical and syntactic dimensions.We experimentally validate that MWE recognition significantly helps syntactic parsing.

Sequential pattern mining is one of the most studied and challenging tasks in data mining. However, the extension of well-known methods from many other classical patterns to sequences is not a trivial task. In this paper we study the notion of δ-freeness for sequences. While this notion has extensively been discussed for itemsets, this work is the first to extend it to sequences. We define an efficient algorithm devoted to the extraction of δ-free sequential patterns. Furthermore, we show the advantage of the δ-free sequences and highlight their importance when building sequence classifiers, and we show how they can be used to address the feature selection problem in statistical classifiers, as well as to build symbolic classifiers which optimizes both accuracy and earliness of predictions.