Regularization, Semi-supervision, and Supervision for a Plausible Attention-Based Explanation

Published in 28th International Conference on Applications of Natural Language to Information Systems (NLDB) 2023, 2023

Abstract

Attention mechanism is contributing to the majority of recent advances in machine learning for natural language processing. Additionally, it results in an attention map that shows the proportional influence of each input in its decision. Empirical studies postulate that attention maps can be provided as an explanation for model output. However, it is still questionable to ask whether this explanation helps regular people to understand and accept the model output (the plausibility of the explanation). Recent studies show that attention weights in RNN encoders are hardly plausible because they spread on input tokens. We thus propose three additional constraints to the learning objective function to improve the plausibility of the attention map: regularization to increase the attention weight sparsity, semi-supervision to supervise the map by a heuristic and supervision by human annotation. Results show that all techniques can improve the attention map plausibility at some level. We also observe that specific instructions for human annotation might have a negative effect on classification performance. Beyond the attention map, results on text classification tasks also show that the contextualization layer plays a crucial role in finding the right space for finding plausible tokens, no matter how constraints bring the gain.

Citation: Duc Hau Nguyen, Cyrielle Mallart, Guillaume Gravier & Pascale Sébillot. Regularization, Semi-supervision, and Supervision for a Plausible Attention-Based Explanation. In Natural Language Processing and Information Systems (NLDB), pp. 285–298, 2023.
Download Paper

Géométrie de l’auto-attention en classification : Quand la géométrie remplace l’attention

Published in [JEP/TALN/RECITAL](https://aclanthology.org/venues/jeptalnrecital/), 2023

Abstract (En)

Several studies have highlighted the anisotropy of embeddings produced by a BERT model within a sentence, meaning their concentration in a given direction, especially in a classification task. In this article, we seek to better understand this phenomenon and how this convergence is constructed by closely analyzing the geometric properties of embeddings, keys, and values in a self-attention layer. We show that the direction in which the embeddings align characterizes the class membership of the sentence. We then study the intrinsic functioning of the self-attention layer and the mechanisms at play between keys and values to ensure the construction of an anisotropic representation. This construction occurs progressively as multiple layers are stacked. It also proves to be robust to external constraints on the distribution of attention weights, which the model compensates for by adjusting the values and keys.

Citation: Loïc Fosse, Duc Hau Nguyen, Pascale Sébillot, and Guillaume Gravier. Géométrie de l’auto-attention en classification : quand la géométrie remplace l’attention. In Conférence sur le Traitement Automatique des Langues Naturelles, pp. 137-15, 2023.
Download Paper

Hybridation Des Approches Symboliques et Apprentissage Profond Pour La Reconnaissance Des Entité Dans Les Signatures de Mail.

Published in 2023 Extraction et Gestion des Connaissances (EGC), 2023

Hybridation Des Approches Symboliques et Apprentissage Profond Pour La Reconnaissance Des Entité Dans Les Signatures de Mail.

Citation: Duc Hau Nguyen, Nicolas Fouqué, Victor Klötzer, Hugo Thomas. Hybridation des approches symboliques et apprentissage profond pour la reconnaissance des entités dans les signatures de mail. In Atelier TextMine @ Extraction et Gestion des Connaissances, 2023
Download Paper

Filtrage et régularisation pour améliorer la plausibilité des poids d’attention dans la tâche d’inférence en langue naturelle

Published in [JEP/TALN/RECITAL](https://aclanthology.org/venues/jeptalnrecital/), 2022

Abstract (En)

We study the plausibility of the attention mechanism for a sentence inference task (entailment), i.e., its ability to provide a plausible explanation for a human regarding the relationship between two sentences. Based on the Explanable Stanford Natural Language Inference (e-SNLI) corpus, it has been shown that attention weights are generally not plausible in practice and tend not to focus on important tokens. In this paper, different approaches are suggested to make attention weights more plausible, relying on masks derived from morphosyntactic analysis or using regularization to enforce sparsity. We demonstrate that these strategies significantly improve the plausibility of attention weights and prove to be more effective than traditional saliency-map approaches.

Citation: Duc Hau Nguyen, Guillaume Gravier, and Pascale Sébillot. Filtrage et régularisation pour améliorer la plausibilité des poids d’attention dans la tâche d’inférence en langue naturelle. Traitement Automatique des Langues Naturelles (TALN), pp. 95–103, 2022.
Download Paper

Géométrie de l’auto-Attention en classification : Quand la géométrie remplace l’attention

Published in [JEP/TALN/RECITAL](https://aclanthology.org/venues/jeptalnrecital/), 2022

Abstract (En)

We study the statistical properties of embeddings in transformer models for French. We rely on an analysis of variance, intra-sentence cosine similarities, and the effective rank of embeddings at different levels of a transformer, for both pre-trained models and models adapted for text classification. We show that the pre-trained FlauBERT and CamemBERT models exhibit very different behaviors, even though both tend to generate anisotropic representations, i.e., concentrating in a cone within the embedding space, as observed for English. Adaptation to text classification alters the models’ behavior, particularly in the final layers, and strongly promotes alignment of the embeddings, also reducing the effective dimensionality of the space in the end. We also highlight a link between the convergence of embeddings within a sentence and text classification, a link whose nature remains challenging to grasp.

Citation: Loïc Fosse, Duc-Hau Nguyen, Pascale Sébillot, and Guillaume Gravier. Une étude statistique des plongements dans les modèles transformers pour le français. In Conference Traitement Automatique des Langues Naturelles (TALN), pp. 247–256, 2022.
Download Paper

Detecting Fake News Conspiracies with Multitask and Prompt-Based Learning

Published in , 2021

Abstract

We present in this paper our participation to the task of fake news conspiracy theories detection from tweets. We rely on a variant of BERT-based classification approach to devise a first classification method for the three different tasks. Moreover, we propose a multitask learning approach to perform the three different tasks at once. Finally, we developed a prompt-based approach to generate classifications thanks to a TinyBERT pre-trained model. Our experimental results show the multitask model to be the best on the three tasks.

Citation: Cheikh Brahim El Vaigh, Thomas Girault, Cyrielle Mallart, and Duc Hau Nguyen. Detecting Fake News Conspiracies with Multitask and Prompt-Based Learning. In Workshop MediaEval Multimedia Evaluation Benchmark, 2021.
Download Paper

A Study of the Plausibility of Attention between RNN Encoders in Natural Language Inference

Published in 20th IEEE International Conference on Machine Learning and Applications (ICMLA), 2021

Abstract

Citation: Duc Hau Nguyen, Guillaume Gravier, Pascale Sébillot. A Study of the Plausibility of Attention between RNN Encoders in Natural Language Inference. In IEEE International Conference on Machine Learning and Applications (ICMLA), pp. 1623–1629, 2021.
Download Paper