Guillaume Wisniewski

Maureen de Seyssel, Guillaume Wisniewski, Emmanuel Dupoux · apple-ml

According to the Language Familiarity Effect (LFE), people are better at discriminating between speakers of their native language. Although this cognitive effect was largely studied in the literature, experiments have only been conducted on a limited number of language pairs and their results only show the presence of the effect without yielding a gradual measure that may vary across language pairs. In this work, we show that the computational model of LFE introduced by Thorburn, Feldmand and Schatz (2019) can address these two limitations. In a first experiment, we attest to this model's capacity to obtain a gradual measure of the LFE by replicating behavioural findings on native and accented speech. In a second experiment, we evaluate LFE on a large number of language pairs, including many which have never been tested on humans. We show that the effect is replicated across a wide array of languages, providing further evidence of its universality. Building on the gradual measure of LFE, we also show that languages belonging to the same family yield smaller scores, supporting the idea of an effect of language distance on LFE.

Maureen de Seyssel, Marvin Lavechin, Hadrien Titeux et al. · apple-ml

We present ProsAudit, a benchmark in English to assess structural prosodic knowledge in self-supervised learning (SSL) speech models. It consists of two subtasks, their corresponding metrics, and an evaluation dataset. In the protosyntax task, the model must correctly identify strong versus weak prosodic boundaries. In the lexical task, the model needs to correctly distinguish between pauses inserted between words and within words. We also provide human evaluation scores on this benchmark. We evaluated a series of SSL models and found that they were all able to perform above chance on both tasks, even when evaluated on an unseen language. However, non-native models performed significantly worse than native ones on the lexical task, highlighting the importance of lexical knowledge in this task. We also found a clear effect of size with models trained on more data performing better in the two subtasks.

Lina Conti, Dennis Fucci, Marco Gaido et al.

Unlike text, speech conveys information about the speaker, such as gender, through acoustic cues like pitch. This gives rise to modality-specific bias concerns. For example, in speech translation (ST), when translating from languages with notional gender, such as English, into languages where gender-ambiguous terms referring to the speaker are assigned grammatical gender, the speaker's vocal characteristics may play a role in gender assignment. This risks misgendering speakers, whether through masculine defaults or vocal-based assumptions. Yet, how ST models make these decisions remains poorly understood. We investigate the mechanisms ST models use to assign gender to speaker-referring terms across three language pairs (en-es/fr/it), examining how training data patterns, internal language model (ILM) biases, and acoustic information interact. We find that models do not simply replicate term-specific gender associations from training data, but learn broader patterns of masculine prevalence. While the ILM exhibits strong masculine bias, models can override these preferences based on acoustic input. Using contrastive feature attribution on spectrograms, we reveal that the model with higher gender accuracy relies on a previously unknown mechanism: using first-person pronouns to link gendered terms back to the speaker, accessing gender information distributed across the frequency spectrum rather than concentrated in pitch.

Bingzhi Li, Guillaume Wisniewski, Benoît Crabbé

The long-distance agreement, evidence for syntactic structure, is increasingly used to assess the syntactic generalization of Neural Language Models. Much work has shown that transformers are capable of high accuracy in varied agreement tasks, but the mechanisms by which the models accomplish this behavior are still not well understood. To better understand transformers' internal working, this work contrasts how they handle two superficially similar but theoretically distinct agreement phenomena: subject-verb and object-past participle agreement in French. Using probing and counterfactual analysis methods, our experiments show that i) the agreement task suffers from several confounders which partially question the conclusions drawn so far and ii) transformers handle subject-verb and object-past participle agreements in a way that is consistent with their modeling in theoretical linguistics.

Joachim Minder, Guillaume Wisniewski, Natalie Kübler

This study investigates the capabilities of large language models (LLMs), specifically ChatGPT, in annotating MT outputs based on an error typology. In contrast to previous work focusing mainly on general language, we explore ChatGPT's ability to identify and categorise errors in specialised translations. By testing two different prompts and based on a customised error typology, we compare ChatGPT annotations with human expert evaluations of translations produced by DeepL and ChatGPT itself. The results show that, for translations generated by DeepL, recall and precision are quite high. However, the degree of accuracy in error categorisation depends on the prompt's specific features and its level of detail, ChatGPT performing very well with a detailed prompt. When evaluating its own translations, ChatGPT achieves significantly poorer results, revealing limitations with self-assessment. These results highlight both the potential and the limitations of LLMs for translation evaluation, particularly in specialised domains. Our experiments pave the way for future research on open-source LLMs, which could produce annotations of comparable or even higher quality. In the future, we also aim to test the practical effectiveness of this automated evaluation in the context of translation training, particularly by optimising the process of human evaluation by teachers and by exploring the impact of annotations by LLMs on students' post-editing and translation learning.

Huan Zhang, Yanjian Zhang, Guillaume Wisniewski et al.

We propose a sparsity-aware evolutionary (SAE) framework for model merging that involves iterative pruning-merging cycles to act as a novel mutation operator. We incorporate the sparsity constraints into the score function, which steers the evolutionary process to favor more sparse models, in addition to other conventional performance scores. Interestingly, the by-product of \textit{competition} for sparsity introduces an extra local \textit{attraction} and interplay into the evolutionary process: if one competitor has more zero elements, the other competitor's non-zero elements will occupy those positions, even though the less sparse competitor loses to the more sparse competitor in other positions. The proposed pipeline is evaluated on a variety of large-scale LLM benchmarks. Experiments demonstrate that our approach can improve model merging reliability across multiple benchmarks, and is easy to incorporate due to its simplicity and being orthogonal to most existing approaches.

Lina Conti, Guillaume Wisniewski

Numerous studies have demonstrated the ability of neural language models to learn various linguistic properties without direct supervision. This work takes an initial step towards exploring the less researched topic of how neural models discover linguistic properties of words, such as gender, as well as the rules governing their usage. We propose to use an artificial corpus generated by a PCFG based on French to precisely control the gender distribution in the training data and determine under which conditions a model correctly captures gender information or, on the contrary, appears gender-biased.

Guillaume Wisniewski, Ophélie Lacroix

We compare the performance of a transition-based parser in regards to different annotation schemes. We pro-pose to convert some specific syntactic constructions observed in the universal dependency treebanks into a so-called more standard representation and to evaluate parsing performances over all the languages of the project. We show that the ``standard'' constructions do not lead systematically to better parsing performance and that the scores vary considerably according to the languages.

Maxime Fily, Guillaume Wisniewski, Severine Guillaume et al.

In the highly constrained context of low-resource language studies, we explore vector representations of speech from a pretrained model to determine their level of abstraction with regard to the audio signal. We propose a new unsupervised method using ABX tests on audio recordings with carefully curated metadata to shed light on the type of information present in the representations. ABX tests determine whether the representations computed by a multilingual speech model encode a given characteristic. Three experiments are devised: one on room acoustics aspects, one on linguistic genre, and one on phonetic aspects. The results confirm that the representations extracted from recordings with different linguistic/extra-linguistic characteristics differ along the same lines. Embedding more audio signal in one vector better discriminates extra-linguistic characteristics, whereas shorter snippets are better to distinguish segmental information. The method is fully unsupervised, potentially opening new research avenues for comparative work on under-documented languages.

Lina Conti, Dennis Fucci, Marco Gaido et al.

Contrastive explanations, which indicate why an AI system produced one output (the target) instead of another (the foil), are widely regarded in explainable AI as more informative and interpretable than standard explanations. However, obtaining such explanations for speech-to-text (S2T) generative models remains an open challenge. Drawing from feature attribution techniques, we propose the first method to obtain contrastive explanations in S2T by analyzing how parts of the input spectrogram influence the choice between alternative outputs. Through a case study on gender assignment in speech translation, we show that our method accurately identifies the audio features that drive the selection of one gender over another. By extending the scope of contrastive explanations to S2T, our work provides a foundation for better understanding S2T models.

Yanjian Zhang, Guillaume Wisniewski, Nadi Tomeh et al.

Human reasoning involves different strategies, each suited to specific problems. Prior work shows that large language model (LLMs) tend to favor a single reasoning strategy, potentially limiting their effectiveness in diverse reasoning challenges. In this work, we investigate whether prompting can control LLMs reasoning strategies and assess its impact on logical problem-solving. While our experiments show that no single strategy consistently improves accuracy, performance could be enhanced if models could adaptively choose the optimal strategy. We propose methods to guide LLMs in strategy selection, highlighting new ways to refine their reasoning abilities.

Séverine Guillaume, Guillaume Wisniewski, Alexis Michaud

XLSR-53 a multilingual model of speech, builds a vector representation from audio, which allows for a range of computational treatments. The experiments reported here use this neural representation to estimate the degree of closeness between audio files, ultimately aiming to extract relevant linguistic properties. We use max-pooling to aggregate the neural representations from a "snippet-lect" (the speech in a 5-second audio snippet) to a "doculect" (the speech in a given resource), then to dialects and languages. We use data from corpora of 11 dialects belonging to 5 less-studied languages. Similarity measurements between the 11 corpora bring out greatest closeness between those that are known to be dialects of the same language. The findings suggest that (i) dialect/language can emerge among the various parameters characterizing audio files and (ii) estimates of overall phonetic/phonological closeness can be obtained for a little-resourced or fully unknown language. The findings help shed light on the type of information captured by neural representations of speech and how it can be extracted from these representations

Aurélien Max, Guillaume Wisniewski

Naturally-occurring instances of linguistic phenomena are important both for training and for evaluating automatic processes on text. When available in large quantities, they also prove interesting material for linguistic studies. In this article, we present a new resource built from Wikipedia's revision history, called WiCoPaCo (Wikipedia Correction and Paraphrase Corpus), which contains numerous editings by human contributors, including various corrections and rewritings. We discuss the main motivations for building such a resource, describe how it was built and present initial applications on French.

Guillaume Wisniewski, Lichao Zhu, Nicolas Ballier et al.

This paper aims at identifying the information flow in state-of-the-art machine translation systems, taking as example the transfer of gender when translating from French into English. Using a controlled set of examples, we experiment several ways to investigate how gender information circulates in a encoder-decoder architecture considering both probing techniques as well as interventions on the internal representations used in the MT system. Our results show that gender information can be found in all token representations built by the encoder and the decoder and lead us to conclude that there are multiple pathways for gender transfer.

José Carlos Rosales Núñez, Guillaume Wisniewski, Djamé Seddah

This work explores the capacities of character-based Neural Machine Translation to translate noisy User-Generated Content (UGC) with a strong focus on exploring the limits of such approaches to handle productive UGC phenomena, which almost by definition, cannot be seen at training time. Within a strict zero-shot scenario, we first study the detrimental impact on translation performance of various user-generated content phenomena on a small annotated dataset we developed, and then show that such models are indeed incapable of handling unknown letters, which leads to catastrophic translation failure once such characters are encountered. We further confirm this behavior with a simple, yet insightful, copy task experiment and highlight the importance of reducing the vocabulary size hyper-parameter to increase the robustness of character-based models for machine translation.

José Carlos Rosales Núñez, Djamé Seddah, Guillaume Wisniewski

This work takes a critical look at the evaluation of user-generated content automatic translation, the well-known specificities of which raise many challenges for MT. Our analyses show that measuring the average-case performance using a standard metric on a UGC test set falls far short of giving a reliable image of the UGC translation quality. That is why we introduce a new data set for the evaluation of UGC translation in which UGC specificities have been manually annotated using a fine-grained typology. Using this data set, we conduct several experiments to measure the impact of different kinds of UGC specificities on translation quality, more precisely than previously possible.

Bingzhi Li, Guillaume Wisniewski, Benoit Crabbé

Many recent works have demonstrated that unsupervised sentence representations of neural networks encode syntactic information by observing that neural language models are able to predict the agreement between a verb and its subject. We take a critical look at this line of research by showing that it is possible to achieve high accuracy on this agreement task with simple surface heuristics, indicating a possible flaw in our assessment of neural networks' syntactic ability. Our fine-grained analyses of results on the long-range French object-verb agreement show that contrary to LSTMs, Transformers are able to capture a non-trivial amount of grammatical structure.

Oliver Adams, Benjamin Galliot, Guillaume Wisniewski et al.

This paper reports on progress integrating the speech recognition toolkit ESPnet into Elpis, a web front-end originally designed to provide access to the Kaldi automatic speech recognition toolkit. The goal of this work is to make end-to-end speech recognition models available to language workers via a user-friendly graphical interface. Encouraging results are reported on (i) development of an ESPnet recipe for use in Elpis, with preliminary results on data sets previously used for training acoustic models with the Persephone toolkit along with a new data set that had not previously been used in speech recognition, and (ii) incorporating ESPnet into Elpis along with UI enhancements and a CUDA-supported Dockerfile.