Romaric Besançon

Robin Armingaud, Romaric Besançon

While large general-purpose Transformer-based encoders excel at general language understanding, their performance diminishes in specialized domains like manufacturing due to a lack of exposure to domain-specific terminology and semantics. In this paper, we address this gap by introducing ManufactuBERT, a RoBERTa model continually pretrained on a large-scale corpus curated for the manufacturing domain. We present a comprehensive data processing pipeline to create this corpus from web data, involving an initial domain-specific filtering step followed by a multi-stage deduplication process that removes redundancies. Our experiments show that ManufactuBERT establishes a new state-of-the-art on a range of manufacturing-related NLP tasks, outperforming strong specialized baselines. More importantly, we demonstrate that training on our carefully deduplicated corpus significantly accelerates convergence, leading to a 33\% reduction in training time and computational cost compared to training on the non-deduplicated dataset. The proposed pipeline offers a reproducible example for developing high-performing encoders in other specialized domains. We will release our model and curated corpus at https://huggingface.co/cea-list-ia.

Victor Bocharov, Romaric Besançon, Gaël de Chalendar et al.

In this article, we describe the architecture of the LIMA (Libre Multilingual Analyzer) framework and its recent evolution with the addition of new text analysis modules based on deep neural networks. We extended the functionality of LIMA in terms of the number of supported languages while preserving existing configurable architecture and the availability of previously developed rule-based and statistical analysis components. Models were trained for more than 60 languages on the Universal Dependencies 2.5 corpora, WikiNer corpora, and CoNLL-03 dataset. Universal Dependencies allowed us to increase the number of supported languages and to generate models that could be integrated into other platforms. This integration of ubiquitous Deep Learning Natural Language Processing models and the use of standard annotated collections using Universal Dependencies can be viewed as a new path of interoperability, through the normalization of models and data, that are complementary to a more standard technical interoperability, implemented in LIMA through services available in Docker containers on Docker Hub.

Clément Cornet, Héloïse Aumaître, Romaric Besançon et al.

Die studies are fundamental to quantifying ancient monetary production, providing insights into the relationship between coinage, politics, and history. The process requires tedious manual work, which limits the size of the corpora that can be studied. Few works have attempted to automate this task, and none have been properly released and evaluated from a computer vision perspective. We propose a fully automatic approach that introduces several innovations compared to previous methods. We rely on fast and robust local descriptors matching that is set automatically. Second, the core of our proposal is a clustering-based approach that uses an intrinsic metric (that does not need the ground truth labels) to determine its critical hyper-parameters. We validate the approach on two corpora of Greek coins, propose an automatic implementation and evaluation of previous baselines, and show that our approach significantly outperforms them.

Robin Armingaud, Romaric Besançon

Relation Extraction (RE) is a fundamental task in Natural Language Processing, and its document-level variant poses significant challenges, due to complex interactions between entities across sentences. While supervised models have achieved strong results in fully resourced settings, their behavior with limited training data remains insufficiently studied. We introduce GLiDRE, a new compact model for document-level relation extraction, designed to work efficiently in both supervised and few-shot settings. Experiments in both low-resource supervised training and few-shot meta-learning benchmarks show that our approach outperforms existing methods in data-constrained scenarios, establishing a new state-of-the-art in few-shot document-level relation extraction. Our code will be publicly available.

Clément Cornet, Romaric Besançon, Hervé Le Borgne

Sparse autoencoders (SAEs) have emerged as a powerful technique for extracting human-interpretable features from neural networks activations. Previous works compared different models based on SAE-derived features but those comparisons have been restricted to models within the same modality. We propose a novel indicator allowing quantitative comparison of models across SAE features, and use it to conduct a comparative study of visual, textual and multimodal encoders. We also propose to quantify the Comparative Sharedness of individual features between different classes of models. With these two new tools, we conduct several studies on 21 encoders of the three types, with two significantly different sizes, and considering generalist and domain specific datasets. The results allow to revisit previous studies at the light of encoders trained in a multimodal context and to quantify to which extent all these models share some representations or features. They also suggest that visual features that are specific to VLMs among vision encoders are shared with text encoders, highlighting the impact of text pretraining. The code is available at https://github.com/CEA-LIST/SAEshareConcepts

Jesús Lovón-Melgarejo, Jose G. Moreno, Romaric Besançon et al.

Since Pretrained Language Models (PLMs) are the cornerstone of the most recent Information Retrieval (IR) models, the way they encode semantic knowledge is particularly important. However, little attention has been given to studying the PLMs' capability to capture hierarchical semantic knowledge. Traditionally, evaluating such knowledge encoded in PLMs relies on their performance on a task-dependent evaluation approach based on proxy tasks, such as hypernymy detection. Unfortunately, this approach potentially ignores other implicit and complex taxonomic relations. In this work, we propose a task-agnostic evaluation method able to evaluate to what extent PLMs can capture complex taxonomy relations, such as ancestors and siblings. The evaluation is based on intrinsic properties that capture the hierarchical nature of taxonomies. Our experimental evaluation shows that the lexico-semantic knowledge implicitly encoded in PLMs does not always capture hierarchical relations. We further demonstrate that the proposed properties can be injected into PLMs to improve their understanding of hierarchy. Through evaluations on taxonomy reconstruction, hypernym discovery and reading comprehension tasks, we show that the knowledge about hierarchy is moderately but not systematically transferable across tasks.

Clément Cornet, Romaric Besançon, Hervé Le Borgne

We propose an alternative to sparse autoencoders (SAEs) as a simple and effective unsupervised method for extracting interpretable concepts from neural networks. The core idea is to cluster differences in activations, which we formally justify within a discriminant analysis framework. To enhance the diversity of extracted concepts, we refine the approach by weighting the clustering using the skewness of activations. The method aligns with Deleuze's modern view of concepts as differences. We evaluate the approach across five models and three modalities (vision, language, and audio), measuring concept quality, diversity, and consistency. Our results show that the proposed method achieves concept quality surpassing prior unsupervised SAE variants while approaching supervised baselines, and that the extracted concepts enable steering of a model's inner representations, demonstrating their causal influence on downstream behavior.

Omar Adjali, Romaric Besançon, Olivier Ferret et al.

In many information extraction applications, entity linking (EL) has emerged as a crucial task that allows leveraging information about named entities from a knowledge base. In this paper, we address the task of multimodal entity linking (MEL), an emerging research field in which textual and visual information is used to map an ambiguous mention to an entity in a knowledge base (KB). First, we propose a method for building a fully annotated Twitter dataset for MEL, where entities are defined in a Twitter KB. Then, we propose a model for jointly learning a representation of both mentions and entities from their textual and visual contexts. We demonstrate the effectiveness of the proposed model by evaluating it on the proposed dataset and highlight the importance of leveraging visual information when it is available.