Rémi Lebret

Yasmine Karoui, Rémi Lebret, Negar Foroutan et al.

Vision-Language Pre-training (VLP) has advanced the performance of many vision-language tasks, such as image-text retrieval, visual entailment, and visual reasoning. The pre-training mostly utilizes lexical databases and image queries in English. Previous work has demonstrated that the pre-training in English does not transfer well to other languages in a zero-shot setting. However, multilingual pre-trained language models (MPLM) have excelled at a variety of single-modal language tasks. In this paper, we propose a simple yet efficient approach to adapt VLP to unseen languages using MPLM. We utilize a cross-lingual contextualized token embeddings alignment approach to train text encoders for non-English languages. Our approach does not require image input and primarily uses machine translation, eliminating the need for target language data. Our evaluation across three distinct tasks (image-text retrieval, visual entailment, and natural language visual reasoning) demonstrates that this approach outperforms the state-of-the-art multilingual vision-language models without requiring large parallel corpora. Our code is available at https://github.com/Yasminekaroui/CliCoTea.

Sepideh Mamooler, Rémi Lebret, Stéphane Massonnet et al.

Active Learning (AL) is a powerful tool for learning with less labeled data, in particular, for specialized domains, like legal documents, where unlabeled data is abundant, but the annotation requires domain expertise and is thus expensive. Recent works have shown the effectiveness of AL strategies for pre-trained language models. However, most AL strategies require a set of labeled samples to start with, which is expensive to acquire. In addition, pre-trained language models have been shown unstable during fine-tuning with small datasets, and their embeddings are not semantically meaningful. In this work, we propose a pipeline for effectively using active learning with pre-trained language models in the legal domain. To this end, we leverage the available unlabeled data in three phases. First, we continue pre-training the model to adapt it to the downstream task. Second, we use knowledge distillation to guide the model's embeddings to a semantically meaningful space. Finally, we propose a simple, yet effective, strategy to find the initial set of labeled samples with fewer actions compared to existing methods. Our experiments on Contract-NLI, adapted to the classification task, and LEDGAR benchmarks show that our approach outperforms standard AL strategies, and is more efficient. Furthermore, our pipeline reaches comparable results to the fully-supervised approach with a small performance gap, and dramatically reduced annotation cost. Code and the adapted data will be made available.

Mohammadreza Banaei, Klaudia Bałazy, Artur Kasymov et al.

Recent transformer language models achieve outstanding results in many natural language processing (NLP) tasks. However, their enormous size often makes them impractical on memory-constrained devices, requiring practitioners to compress them to smaller networks. In this paper, we explore offline compression methods, meaning computationally-cheap approaches that do not require further fine-tuning of the compressed model. We challenge the classical matrix factorization methods by proposing a novel, better-performing autoencoder-based framework. We perform a comprehensive ablation study of our approach, examining its different aspects over a diverse set of evaluation settings. Moreover, we show that enabling collaboration between modules across layers by compressing certain modules together positively impacts the final model performance. Experiments on various NLP tasks demonstrate that our approach significantly outperforms commonly used factorization-based offline compression methods.

İlker Gül, Rémi Lebret, Karl Aberer

Stance detection, a key task in natural language processing, determines an author's viewpoint based on textual analysis. This study evaluates the evolution of stance detection methods, transitioning from early machine learning approaches to the groundbreaking BERT model, and eventually to modern Large Language Models (LLMs) such as ChatGPT, LLaMa-2, and Mistral-7B. While ChatGPT's closed-source nature and associated costs present challenges, the open-source models like LLaMa-2 and Mistral-7B offers an encouraging alternative. Initially, our research focused on fine-tuning ChatGPT, LLaMa-2, and Mistral-7B using several publicly available datasets. Subsequently, to provide a comprehensive comparison, we assess the performance of these models in zero-shot and few-shot learning scenarios. The results underscore the exceptional ability of LLMs in accurately detecting stance, with all tested models surpassing existing benchmarks. Notably, LLaMa-2 and Mistral-7B demonstrate remarkable efficiency and potential for stance detection, despite their smaller sizes compared to ChatGPT. This study emphasizes the potential of LLMs in stance detection and calls for more extensive research in this field.

Negar Foroutan, Angelika Romanou, Matin Ansaripour et al.

Documents are fundamental to preserving and disseminating information, often incorporating complex layouts, tables, and charts that pose significant challenges for automatic document understanding (DU). While vision-language large models (VLLMs) have demonstrated improvements across various tasks, their effectiveness in processing long-context vision inputs remains unclear. This paper introduces WikiMixQA, a benchmark comprising 1,000 multiple-choice questions (MCQs) designed to evaluate cross-modal reasoning over tables and charts extracted from 4,000 Wikipedia pages spanning seven distinct topics. Unlike existing benchmarks, WikiMixQA emphasizes complex reasoning by requiring models to synthesize information from multiple modalities. We evaluate 12 state-of-the-art vision-language models, revealing that while proprietary models achieve ~70% accuracy when provided with direct context, their performance deteriorates significantly when retrieval from long documents is required. Among these, GPT-4-o is the only model exceeding 50% accuracy in this setting, whereas open-source models perform considerably worse, with a maximum accuracy of 27%. These findings underscore the challenges of long-context, multi-modal reasoning and establish WikiMixQA as a crucial benchmark for advancing document understanding research.

Tim Poštuvan, Jiaxuan You, Mohammadreza Banaei et al.

One of the most important factors that contribute to the success of a machine learning model is a good training objective. Training objective crucially influences the model's performance and generalization capabilities. This paper specifically focuses on graph neural network training objective for link prediction, which has not been explored in the existing literature. Here, the training objective includes, among others, a negative sampling strategy, and various hyperparameters, such as edge message ratio which controls how training edges are used. Commonly, these hyperparameters are fine-tuned by complete grid search, which is very time-consuming and model-dependent. To mitigate these limitations, we propose Adaptive Grid Search (AdaGrid), which dynamically adjusts the edge message ratio during training. It is model agnostic and highly scalable with a fully customizable computational budget. Through extensive experiments, we show that AdaGrid can boost the performance of the models up to $1.9\%$ while being nine times more time-efficient than a complete search. Overall, AdaGrid represents an effective automated algorithm for designing machine learning training objectives.

Saibo Geng, Rémi Lebret, Karl Aberer

Deep learning-based Natural Language Processing methods, especially transformers, have achieved impressive performance in the last few years. Applying those state-of-the-art NLP methods to legal activities to automate or simplify some simple work is of great value. This work investigates the value of domain adaptive pre-training and language adapters in legal NLP tasks. By comparing the performance of language models with domain adaptive pre-training on different tasks and different dataset splits, we show that domain adaptive pre-training is only helpful with low-resource downstream tasks, thus far from being a panacea. We also benchmark the performance of adapters in a typical legal NLP task and show that they can yield similar performance to full model tuning with much smaller training costs. As an additional result, we release LegalRoBERTa, a RoBERTa model further pre-trained on legal corpora.

Klaudia Bałazy, Mohammadreza Banaei, Rémi Lebret et al.

The adoption of Transformer-based models in natural language processing (NLP) has led to great success using a massive number of parameters. However, due to deployment constraints in edge devices, there has been a rising interest in the compression of these models to improve their inference time and memory footprint. This paper presents a novel loss objective to compress token embeddings in the Transformer-based models by leveraging an AutoEncoder architecture. More specifically, we emphasize the importance of the direction of compressed embeddings with respect to original uncompressed embeddings. The proposed method is task-agnostic and does not require further language modeling pre-training. Our method significantly outperforms the commonly used SVD-based matrix-factorization approach in terms of initial language model Perplexity. Moreover, we evaluate our proposed approach over SQuAD v1.1 dataset and several downstream tasks from the GLUE benchmark, where we also outperform the baseline in most scenarios. Our code is public.

Mohammadreza Banaei, Rémi Lebret, Karl Aberer

This paper presents our approach for SwissText & KONVENS 2020 shared task 2, which is a multi-stage neural model for Swiss German (GSW) identification on Twitter. Our model outputs either GSW or non-GSW and is not meant to be used as a generic language identifier. Our architecture consists of two independent filters where the first one favors recall, and the second one filter favors precision (both towards GSW). Moreover, we do not use binary models (GSW vs. not-GSW) in our filters but rather a multi-class classifier with GSW being one of the possible labels. Our model reaches F1-score of 0.982 on the test set of the shared task.

Fangyu Liu, Rémi Lebret, Didier Orel et al.

We propose an automated image selection system to assist photo editors in selecting suitable images for news articles. The system fuses multiple textual sources extracted from news articles and accepts multilingual inputs. It is equipped with char-level word embeddings to help both modeling morphologically rich languages, e.g. German, and transferring knowledge across nearby languages. The text encoder adopts a hierarchical self-attention mechanism to attend more to both keywords within a piece of text and informative components of a news article. We extensively experiment with our system on a large-scale text-image database containing multimodal multilingual news articles collected from Swiss local news media websites. The system is compared with multiple baselines with ablation studies and is shown to beat existing text-image retrieval methods in a weakly-supervised learning setting. Besides, we also offer insights on the advantage of using multiple textual sources and multilingual data.

Fábio Perez, Rémi Lebret, Karl Aberer

In this work, we introduce a novel framework that employs cluster annotation to boost active learning by reducing the number of human interactions required to train deep neural networks. Instead of annotating single samples individually, humans can also label clusters, producing a higher number of annotated samples with the cost of a small label error. Our experiments show that the proposed framework requires 82% and 87% less human interactions for CIFAR-10 and EuroSAT datasets respectively when compared with the fully-supervised training while maintaining similar performance on the test set.

Hamza Harkous, Kassem Fawaz, Rémi Lebret et al.

Privacy policies are the primary channel through which companies inform users about their data collection and sharing practices. These policies are often long and difficult to comprehend. Short notices based on information extracted from privacy policies have been shown to be useful but face a significant scalability hurdle, given the number of policies and their evolution over time. Companies, users, researchers, and regulators still lack usable and scalable tools to cope with the breadth and depth of privacy policies. To address these hurdles, we propose an automated framework for privacy policy analysis (Polisis). It enables scalable, dynamic, and multi-dimensional queries on natural language privacy policies. At the core of Polisis is a privacy-centric language model, built with 130K privacy policies, and a novel hierarchy of neural-network classifiers that accounts for both high-level aspects and fine-grained details of privacy practices. We demonstrate Polisis' modularity and utility with two applications supporting structured and free-form querying. The structured querying application is the automated assignment of privacy icons from privacy policies. With Polisis, we can achieve an accuracy of 88.4% on this task. The second application, PriBot, is the first freeform question-answering system for privacy policies. We show that PriBot can produce a correct answer among its top-3 results for 82% of the test questions. Using an MTurk user study with 700 participants, we show that at least one of PriBot's top-3 answers is relevant to users for 89% of the test questions.

Amit Gupta, Rémi Lebret, Hamza Harkous et al.

We propose a novel, semi-supervised approach towards domain taxonomy induction from an input vocabulary of seed terms. Unlike all previous approaches, which typically extract direct hypernym edges for terms, our approach utilizes a novel probabilistic framework to extract hypernym subsequences. Taxonomy induction from extracted subsequences is cast as an instance of the minimumcost flow problem on a carefully designed directed graph. Through experiments, we demonstrate that our approach outperforms stateof- the-art taxonomy induction approaches across four languages. Importantly, we also show that our approach is robust to the presence of noise in the input vocabulary. To the best of our knowledge, no previous approaches have been empirically proven to manifest noise-robustness in the input vocabulary.

Amit Gupta, Rémi Lebret, Hamza Harkous et al.

We propose a simple, yet effective, approach towards inducing multilingual taxonomies from Wikipedia. Given an English taxonomy, our approach leverages the interlanguage links of Wikipedia followed by character-level classifiers to induce high-precision, high-coverage taxonomies in other languages. Through experiments, we demonstrate that our approach significantly outperforms the state-of-the-art, heuristics-heavy approaches for six languages. As a consequence of our work, we release presumably the largest and the most accurate multilingual taxonomic resource spanning over 280 languages.

Rémi Lebret, Ronan Collobert

Recently, there has been a lot of effort to represent words in continuous vector spaces. Those representations have been shown to capture both semantic and syntactic information about words. However, distributed representations of phrases remain a challenge. We introduce a novel model that jointly learns word vector representations and their summation. Word representations are learnt using the word co-occurrence statistical information. To embed sequences of words (i.e. phrases) with different sizes into a common semantic space, we propose to average word vector representations. In contrast with previous methods which reported a posteriori some compositionality aspects by simple summation, we simultaneously train words to sum, while keeping the maximum information from the original vectors. We evaluate the quality of the word representations on several classical word evaluation tasks, and we introduce a novel task to evaluate the quality of the phrase representations. While our distributed representations compete with other methods of learning word representations on word evaluations, we show that they give better performance on the phrase evaluation. Such representations of phrases could be interesting for many tasks in natural language processing.

Rémi Lebret, Pedro O. Pinheiro, Ronan Collobert

Generating a novel textual description of an image is an interesting problem that connects computer vision and natural language processing. In this paper, we present a simple model that is able to generate descriptive sentences given a sample image. This model has a strong focus on the syntax of the descriptions. We train a purely bilinear model that learns a metric between an image representation (generated from a previously trained Convolutional Neural Network) and phrases that are used to described them. The system is then able to infer phrases from a given image sample. Based on caption syntax statistics, we propose a simple language model that can produce relevant descriptions for a given test image using the phrases inferred. Our approach, which is considerably simpler than state-of-the-art models, achieves comparable results in two popular datasets for the task: Flickr30k and the recently proposed Microsoft COCO.

Rémi Lebret, Ronan Collobert

The bag-of-words (BOW) model is the common approach for classifying documents, where words are used as feature for training a classifier. This generally involves a huge number of features. Some techniques, such as Latent Semantic Analysis (LSA) or Latent Dirichlet Allocation (LDA), have been designed to summarize documents in a lower dimension with the least semantic information loss. Some semantic information is nevertheless always lost, since only words are considered. Instead, we aim at using information coming from n-grams to overcome this limitation, while remaining in a low-dimension space. Many approaches, such as the Skip-gram model, provide good word vector representations very quickly. We propose to average these representations to obtain representations of n-grams. All n-grams are thus embedded in a same semantic space. A K-means clustering can then group them into semantic concepts. The number of features is therefore dramatically reduced and documents can be represented as bag of semantic concepts. We show that this model outperforms LSA and LDA on a sentiment classification task, and yields similar results than a traditional BOW-model with far less features.

Rémi Lebret, Ronan Collobert

Recent works on word representations mostly rely on predictive models. Distributed word representations (aka word embeddings) are trained to optimally predict the contexts in which the corresponding words tend to appear. Such models have succeeded in capturing word similarties as well as semantic and syntactic regularities. Instead, we aim at reviving interest in a model based on counts. We present a systematic study of the use of the Hellinger distance to extract semantic representations from the word co-occurence statistics of large text corpora. We show that this distance gives good performance on word similarity and analogy tasks, with a proper type and size of context, and a dimensionality reduction based on a stochastic low-rank approximation. Besides being both simple and intuitive, this method also provides an encoding function which can be used to infer unseen words or phrases. This becomes a clear advantage compared to predictive models which must train these new words.

Rémi Lebret, Ronan Collobert

Word embeddings resulting from neural language models have been shown to be successful for a large variety of NLP tasks. However, such architecture might be difficult to train and time-consuming. Instead, we propose to drastically simplify the word embeddings computation through a Hellinger PCA of the word co-occurence matrix. We compare those new word embeddings with some well-known embeddings on NER and movie review tasks and show that we can reach similar or even better performance. Although deep learning is not really necessary for generating good word embeddings, we show that it can provide an easy way to adapt embeddings to specific tasks.