Marco Dinarelli

Titouan Parcollet, Ha Nguyen, Solene Evain et al.

Self-supervised learning (SSL) is at the origin of unprecedented improvements in many different domains including computer vision and natural language processing. Speech processing drastically benefitted from SSL as most of the current domain-related tasks are now being approached with pre-trained models. This work introduces LeBenchmark 2.0 an open-source framework for assessing and building SSL-equipped French speech technologies. It includes documented, large-scale and heterogeneous corpora with up to 14,000 hours of heterogeneous speech, ten pre-trained SSL wav2vec 2.0 models containing from 26 million to one billion learnable parameters shared with the community, and an evaluation protocol made of six downstream tasks to complement existing benchmarks. LeBenchmark 2.0 also presents unique perspectives on pre-trained SSL models for speech with the investigation of frozen versus fine-tuned downstream models, task-agnostic versus task-specific pre-trained models as well as a discussion on the carbon footprint of large-scale model training. Overall, the newly introduced models trained on 14,000 hours of French speech outperform multilingual and previous LeBenchmark SSL models across the benchmark but also required up to four times more energy for pre-training.

Lorenzo Lupo, Marco Dinarelli, Laurent Besacier

A straightforward approach to context-aware neural machine translation consists in feeding the standard encoder-decoder architecture with a window of consecutive sentences, formed by the current sentence and a number of sentences from its context concatenated to it. In this work, we propose an improved concatenation approach that encourages the model to focus on the translation of the current sentence, discounting the loss generated by target context. We also propose an additional improvement that strengthen the notion of sentence boundaries and of relative sentence distance, facilitating model compliance to the context-discounted objective. We evaluate our approach with both average-translation quality metrics and contrastive test sets for the translation of inter-sentential discourse phenomena, proving its superiority to the vanilla concatenation approach and other sophisticated context-aware systems.

Lorenzo Lupo, Marco Dinarelli, Laurent Besacier

Context-aware translation can be achieved by processing a concatenation of consecutive sentences with the standard Transformer architecture. This paper investigates the intuitive idea of providing the model with explicit information about the position of the sentences contained in the concatenation window. We compare various methods to encode sentence positions into token representations, including novel methods. Our results show that the Transformer benefits from certain sentence position encoding methods on English to Russian translation if trained with a context-discounted loss (Lupo et al., 2022). However, the same benefits are not observed in English to German. Further empirical efforts are necessary to define the conditions under which the proposed approach is beneficial.

Ryan Whetten, Titouan Parcollet, Marco Dinarelli et al.

Self-supervised learning (SSL) has transformed speech processing, yet its reliance on massive pre-training datasets remains a bottleneck. While robustness is often attributed to scale and diversity, the role of the data distribution is less understood. We systematically examine how curated subsets of pre-training data influence Automatic Speech Recognition (ASR) performance. Surprisingly, optimizing for acoustic, speaker, or linguistic diversity yields no clear improvements over random sampling. Instead, we find that prioritizing the longest utterances achieves superior ASR results while using only half the original dataset, reducing pre-training time by 24% on a large corpora. These findings suggest that for pre-training speech SSL models, data length is a more critical factor than either data diversity or overall data quantity for performance and efficiency, offering a new perspective for data selection strategies in SSL speech processing.

Elisa Gugliotta, Marco Dinarelli

In this paper we present the final result of a project on Tunisian Arabic encoded in Arabizi, the Latin-based writing system for digital conversations. The project led to the creation of two integrated and independent resources: a corpus and a NLP tool created to annotate the former with various levels of linguistic information: word classification, transliteration, tokenization, POS-tagging, lemmatization. We discuss our choices in terms of computational and linguistic methodology and the strategies adopted to improve our results. We report on the experiments performed in order to outline our research path. Finally, we explain why we believe in the potential of these resources for both computational and linguistic researches. Keywords: Tunisian Arabizi, Annotated Corpus, Neural Network Architecture

Marco Dinarelli, Marco Naguib, François Portet

Recent advances in spoken language understanding benefited from Self-Supervised models trained on large speech corpora. For French, the LeBenchmark project has made such models available and has led to impressive progress on several tasks including spoken language understanding. These advances have a non-negligible cost in terms of computation time and energy consumption. In this paper, we compare several learning strategies trying to reduce such cost while keeping competitive performance. At the same time we propose an extensive analysis where we measure the cost of our models in terms of training time and electric energy consumption, hopefully promoting a comprehensive evaluation procedure. The experiments are performed on the FSC and MEDIA corpora, and show that it is possible to reduce the learning cost while maintaining state-of-the-art performance and using SSL models.

Ryan Whetten, Titouan Parcollet, Adel Moumen et al.

Self-Supervised Learning (SSL) has proven to be effective in various domains, including speech processing. However, SSL is computationally and memory expensive. This is in part due the quadratic complexity of multi-head self-attention (MHSA). Alternatives for MHSA have been proposed and used in the speech domain, but have yet to be investigated properly in an SSL setting. In this work, we study the effects of replacing MHSA with recent state-of-the-art alternatives that have linear complexity, namely, HyperMixing, Fastformer, SummaryMixing, and Mamba. We evaluate these methods by looking at the speed, the amount of VRAM consumed, and the performance on the SSL MP3S benchmark. Results show that these linear alternatives maintain competitive performance compared to MHSA while, on average, decreasing VRAM consumption by around 20% to 60% and increasing speed from 7% to 65% for input sequences ranging from 20 to 80 seconds.

Marco Naguib, François Portet, Marco Dinarelli

Recent advances in spoken language understanding benefited from Self-Supervised models trained on large speech corpora. For French, the LeBenchmark project has made such models available and has led to impressive progress on several tasks including spoken language understanding. These advances have a non-negligible cost in terms of computation time and energy consumption. In this paper, we compare several learning strategies aiming at reducing such cost while keeping competitive performances. The experiments are performed on the MEDIA corpus, and show that it is possible to reduce the learning cost while maintaining state-of-the-art performances.

Phuong-Hang Le, Valentin Pelloin, Arnault Chatelain et al.

We release Pantagruel models, a new family of self-supervised encoder models for French text and speech. Instead of predicting modality-tailored targets such as textual tokens or speech units, Pantagruel learns contextualized target representations in the feature space, allowing modality-specific encoders to capture linguistic and acoustic regularities more effectively. Separate models are pre-trained on large-scale French corpora, including Wikipedia, OSCAR and CroissantLLM for text, together with MultilingualLibriSpeech, LeBenchmark, and INA-100k for speech. INA-100k is a newly introduced 100,000-hour corpus of French audio derived from the archives of the Institut National de l'Audiovisuel (INA), the national repository of French radio and television broadcasts, providing highly diverse audio data. We evaluate Pantagruel across a broad range of downstream tasks spanning both modalities, including those from the standard French benchmarks such as FLUE or LeBenchmark. Across these tasks, Pantagruel models show competitive or superior performance compared to strong French baselines such as CamemBERT, FlauBERT, and LeBenchmark2.0, while maintaining a shared architecture that can seamlessly handle either speech or text inputs. These results confirm the effectiveness of feature-space self-supervised objectives for French representation learning and highlight Pantagruel as a robust foundation for multimodal speech-text understanding.

Ryan Whetten, Titouan Parcollet, Marco Dinarelli et al.

Self-Supervised Learning (SSL) has proven to be useful in various speech tasks. However, these methods are generally very demanding in terms of data, memory, and computational resources. BERT-based Speech pre-Training with Random-projection Quantizer (BEST-RQ), is an SSL method that has shown great performance on Automatic Speech Recognition (ASR) while being simpler than other SSL methods, such as wav2vec 2.0. Despite BEST-RQ's great performance, details are lacking in the original paper, such as the amount of GPU/TPU hours used in pre-training, and there is no official easy-to-use open-source implementation. Furthermore, BEST-RQ has not been evaluated on other downstream tasks aside from ASR and speech translation. In this work, we describe a re-implementation of a Random-projection quantizer and perform a preliminary study with a comparison to wav2vec 2.0 on four downstream tasks. We discuss the details and differences of our implementation. We show that a random projection quantizer can achieve similar downstream performance as wav2vec 2.0 while decreasing training time by over a factor of two.

Mariam Nakhlé, Marco Dinarelli, Raheel Qader et al.

Despite the strong research interest in document-level Machine Translation (MT), the test sets dedicated to this task are still scarce. The existing test sets mainly cover topics from the general domain and fall short on specialised domains, such as legal and financial. Also, in spite of their document-level aspect, they still follow a sentence-level logic that does not allow for including certain linguistic phenomena such as information reorganisation. In this work, we aim to fill this gap by proposing a novel test set: DOLFIN. The dataset is built from specialised financial documents, and it makes a step towards true document-level MT by abandoning the paradigm of perfectly aligned sentences, presenting data in units of sections rather than sentences. The test set consists of an average of 1950 aligned sections for five language pairs. We present a detailed data collection pipeline that can serve as inspiration for aligning new document-level datasets. We demonstrate the usefulness and quality of this test set by evaluating a number of models. Our results show that the test set is able to discriminate between context-sensitive and context-agnostic models and shows the weaknesses when models fail to accurately translate financial texts. The test set is made public for the community.

Ryan Whetten, Lucas Maison, Titouan Parcollet et al.

In Self-Supervised Learning (SSL), pre-training and evaluation are resource intensive. In the speech domain, current indicators of the quality of SSL models during pre-training, such as the loss, do not correlate well with downstream performance. Consequently, it is often difficult to gauge the final downstream performance in a cost efficient manner during pre-training. In this work, we propose unsupervised efficient methods that give insights into the quality of the pre-training of SSL speech models, namely, measuring the cluster quality and rank of the embeddings of the SSL model. Results show that measures of cluster quality and rank correlate better with downstream performance than the pre-training loss with only one hour of unlabeled audio, reducing the need for GPU hours and labeled data in SSL model evaluation.

Solene Evain, Ha Nguyen, Hang Le et al.

Self-Supervised Learning (SSL) using huge unlabeled data has been successfully explored for image and natural language processing. Recent works also investigated SSL from speech. They were notably successful to improve performance on downstream tasks such as automatic speech recognition (ASR). While these works suggest it is possible to reduce dependence on labeled data for building efficient speech systems, their evaluation was mostly made on ASR and using multiple and heterogeneous experimental settings (most of them for English). This questions the objective comparison of SSL approaches and the evaluation of their impact on building speech systems. In this paper, we propose LeBenchmark: a reproducible framework for assessing SSL from speech. It not only includes ASR (high and low resource) tasks but also spoken language understanding, speech translation and emotion recognition. We also focus on speech technologies in a language different than English: French. SSL models of different sizes are trained from carefully sourced and documented datasets. Experiments show that SSL is beneficial for most but not all tasks which confirms the need for exhaustive and reliable benchmarks to evaluate its real impact. LeBenchmark is shared with the scientific community for reproducible research in SSL from speech.

Lorenzo Lupo, Marco Dinarelli, Laurent Besacier

Multi-encoder models are a broad family of context-aware neural machine translation systems that aim to improve translation quality by encoding document-level contextual information alongside the current sentence. The context encoding is undertaken by contextual parameters, trained on document-level data. In this work, we discuss the difficulty of training these parameters effectively, due to the sparsity of the words in need of context (i.e., the training signal), and their relevant context. We propose to pre-train the contextual parameters over split sentence pairs, which makes an efficient use of the available data for two reasons. Firstly, it increases the contextual training signal by breaking intra-sentential syntactic relations, and thus pushing the model to search the context for disambiguating clues more frequently. Secondly, it eases the retrieval of relevant context, since context segments become shorter. We propose four different splitting methods, and evaluate our approach with BLEU and contrastive test sets. Results show that it consistently improves learning of contextual parameters, both in low and high resource settings.

Elisa Gugliotta, Marco Dinarelli, Olivier Kraif

In this paper we propose a multi-task sequence prediction system, based on recurrent neural networks and used to annotate on multiple levels an Arabizi Tunisian corpus. The annotation performed are text classification, tokenization, PoS tagging and encoding of Tunisian Arabizi into CODA* Arabic orthography. The system is learned to predict all the annotation levels in cascade, starting from Arabizi input. We evaluate the system on the TIGER German corpus, suitably converting data to have a multi-task problem, in order to show the effectiveness of our neural architecture. We show also how we used the system in order to annotate a Tunisian Arabizi corpus, which has been afterwards manually corrected and used to further evaluate sequence models on Tunisian data. Our system is developed for the Fairseq framework, which allows for a fast and easy use for any other sequence prediction problem.

Elisa Gugliotta, Marco Dinarelli

This article describes the constitution process of the first morpho-syntactically annotated Tunisian Arabish Corpus (TArC). Arabish, also known as Arabizi, is a spontaneous coding of Arabic dialects in Latin characters and arithmographs (numbers used as letters). This code-system was developed by Arabic-speaking users of social media in order to facilitate the writing in the Computer-Mediated Communication (CMC) and text messaging informal frameworks. There is variety in the realization of Arabish amongst dialects, and each Arabish code-system is under-resourced, in the same way as most of the Arabic dialects. In the last few years, the focus on Arabic dialects in the NLP field has considerably increased. Taking this into consideration, TArC will be a useful support for different types of analyses, computational and linguistic, as well as for NLP tools training. In this article we will describe preliminary work on the TArC semi-automatic construction process and some of the first analyses we developed on TArC. In addition, in order to provide a complete overview of the challenges faced during the building process, we will present the main Tunisian dialect characteristics and their encoding in Tunisian Arabish.

Marco Dinarelli, Nikita Kapoor, Bassam Jabaian et al.

End-to-end architectures have been recently proposed for spoken language understanding (SLU) and semantic parsing. Based on a large amount of data, those models learn jointly acoustic and linguistic-sequential features. Such architectures give very good results in the context of domain, intent and slot detection, their application in a more complex semantic chunking and tagging task is less easy. For that, in many cases, models are combined with an external language model to enhance their performance. In this paper we introduce a data efficient system which is trained end-to-end, with no additional, pre-trained external module. One key feature of our approach is an incremental training procedure where acoustic, language and semantic models are trained sequentially one after the other. The proposed model has a reasonable size and achieves competitive results with respect to state-of-the-art while using a small training dataset. In particular, we reach 24.02% Concept Error Rate (CER) on MEDIA/test while training on MEDIA/train without any additional data.

Marco Dinarelli, Loïc Grobol

We propose a neural architecture with the main characteristics of the most successful neural models of the last years: bidirectional RNNs, encoder-decoder, and the Transformer model. Evaluation on three sequence labelling tasks yields results that are close to the state-of-the-art for all tasks and better than it for some of them, showing the pertinence of this hybrid architecture for this kind of tasks.

Marco Dinarelli, Loïc Grobol

During the last couple of years, Recurrent Neural Networks (RNN) have reached state-of-the-art performances on most of the sequence modelling problems. In particular, the "sequence to sequence" model and the neural CRF have proved to be very effective in this domain. In this article, we propose a new RNN architecture for sequence labelling, leveraging gated recurrent layers to take arbitrarily long contexts into account, and using two decoders operating forward and backward. We compare several variants of the proposed solution and their performances to the state-of-the-art. Most of our results are better than the state-of-the-art or very close to it and thanks to the use of recent technologies, our architecture can scale on corpora larger than those used in this work.

Marco Dinarelli, Yoann Dupont, Isabelle Tellier

Understanding spoken language is a highly complex problem, which can be decomposed into several simpler tasks. In this paper, we focus on Spoken Language Understanding (SLU), the module of spoken dialog systems responsible for extracting a semantic interpretation from the user utterance. The task is treated as a labeling problem. In the past, SLU has been performed with a wide variety of probabilistic models. The rise of neural networks, in the last couple of years, has opened new interesting research directions in this domain. Recurrent Neural Networks (RNNs) in particular are able not only to represent several pieces of information as embeddings but also, thanks to their recurrent architecture, to encode as embeddings relatively long contexts. Such long contexts are in general out of reach for models previously used for SLU. In this paper we propose novel RNNs architectures for SLU which outperform previous ones. Starting from a published idea as base block, we design new deep RNNs achieving state-of-the-art results on two widely used corpora for SLU: ATIS (Air Traveling Information System), in English, and MEDIA (Hotel information and reservation in France), in French.

Yoann Dupont, Marco Dinarelli, Isabelle Tellier

In the last few years, Recurrent Neural Networks (RNNs) have proved effective on several NLP tasks. Despite such great success, their ability to model \emph{sequence labeling} is still limited. This lead research toward solutions where RNNs are combined with models which already proved effective in this domain, such as CRFs. In this work we propose a solution far simpler but very effective: an evolution of the simple Jordan RNN, where labels are re-injected as input into the network, and converted into embeddings, in the same way as words. We compare this RNN variant to all the other RNN models, Elman and Jordan RNN, LSTM and GRU, on two well-known tasks of Spoken Language Understanding (SLU). Thanks to label embeddings and their combination at the hidden layer, the proposed variant, which uses more parameters than Elman and Jordan RNNs, but far fewer than LSTM and GRU, is more effective than other RNNs, but also outperforms sophisticated CRF models.

Marco Dinarelli, Isabelle Tellier

In this paper we study different types of Recurrent Neural Networks (RNN) for sequence labeling tasks. We propose two new variants of RNNs integrating improvements for sequence labeling, and we compare them to the more traditional Elman and Jordan RNNs. We compare all models, either traditional or new, on four distinct tasks of sequence labeling: two on Spoken Language Understanding (ATIS and MEDIA); and two of POS tagging for the French Treebank (FTB) and the Penn Treebank (PTB) corpora. The results show that our new variants of RNNs are always more effective than the others.