Irina Illina

Tulika Bose, Nikolaos Aletras, Irina Illina et al.

Hate speech classifiers exhibit substantial performance degradation when evaluated on datasets different from the source. This is due to learning spurious correlations between words that are not necessarily relevant to hateful language, and hate speech labels from the training corpus. Previous work has attempted to mitigate this problem by regularizing specific terms from pre-defined static dictionaries. While this has been demonstrated to improve the generalizability of classifiers, the coverage of such methods is limited and the dictionaries require regular manual updates from human experts. In this paper, we propose to automatically identify and reduce spurious correlations using attribution methods with dynamic refinement of the list of terms that need to be regularized during training. Our approach is flexible and improves the cross-corpora performance over previous work independently and in combination with pre-defined dictionaries.

Dana Ruiter, Liane Reiners, Ashwin Geet D'Sa et al.

Even though hate speech (HS) online has been an important object of research in the last decade, most HS-related corpora over-simplify the phenomenon of hate by attempting to label user comments as "hate" or "neutral". This ignores the complex and subjective nature of HS, which limits the real-life applicability of classifiers trained on these corpora. In this study, we present the M-Phasis corpus, a corpus of ~9k German and French user comments collected from migration-related news articles. It goes beyond the "hate"-"neutral" dichotomy and is instead annotated with 23 features, which in combination become descriptors of various types of speech, ranging from critical comments to implicit and explicit expressions of hate. The annotations are performed by 4 native speakers per language and achieve high (0.77 <= k <= 1) inter-annotator agreements. Besides describing the corpus creation and presenting insights from a content, error and domain analysis, we explore its data characteristics by training several classification baselines.

Tulika Bose, Nikolaos Aletras, Irina Illina et al.

State-of-the-art approaches for hate-speech detection usually exhibit poor performance in out-of-domain settings. This occurs, typically, due to classifiers overemphasizing source-specific information that negatively impacts its domain invariance. Prior work has attempted to penalize terms related to hate-speech from manually curated lists using feature attribution methods, which quantify the importance assigned to input terms by the classifier when making a prediction. We, instead, propose a domain adaptation approach that automatically extracts and penalizes source-specific terms using a domain classifier, which learns to differentiate between domains, and feature-attribution scores for hate-speech classes, yielding consistent improvements in cross-domain evaluation.

Tulika Bose, Irina Illina, Dominique Fohr

The concerning rise of hateful content on online platforms has increased the attention towards automatic hate speech detection, commonly formulated as a supervised classification task. State-of-the-art deep learning-based approaches usually require a substantial amount of labeled resources for training. However, annotating hate speech resources is expensive, time-consuming, and often harmful to the annotators. This creates a pressing need to transfer knowledge from the existing labeled resources to low-resource hate speech corpora with the goal of improving system performance. For this, neighborhood-based frameworks have been shown to be effective. However, they have limited flexibility. In our paper, we propose a novel training strategy that allows flexible modeling of the relative proximity of neighbors retrieved from a resource-rich corpus to learn the amount of transfer. In particular, we incorporate neighborhood information with Optimal Transport, which permits exploiting the geometry of the data embedding space. By aligning the joint embedding and label distributions of neighbors, we demonstrate substantial improvements over strong baselines, in low-resource scenarios, on different publicly available hate speech corpora.

Yaya Sy, Dioula Doucouré, Christophe Cerisara et al.

Speakers of under-represented languages face both a language barrier, as most online knowledge is in a few dominant languages, and a modality barrier, since information is largely text-based while many languages are primarily oral. We address this for French-Wolof by training the first bilingual speech-text Matryoshka embedding model, enabling efficient retrieval of French text from Wolof speech queries without relying on a costly ASR-translation pipelines. We introduce large-scale data curation pipelines and new benchmarks, compare modeling strategies, and show that modality fusion within a frozen text Matryoshka model performs best. Although trained only for retrieval, the model generalizes well to other tasks, such as speech intent detection, indicating the learning of general semantic representations. Finally, we analyze cost-accuracy trade-offs across Matryoshka dimensions and ranks, showing that information is concentrated only in a few components, suggesting potential for efficiency improvements.

Tian Huang, Tom Bourgeade, Irina Illina

Objective Structured Clinical Examinations (OSCEs) are the standard method for assessing medical students' clinical and communication skills through structured patient interviews. In France, however, the organization of training sessions is limited by human and logistical constraints, restricting students' access to repeated practice and structured feedback. Recent advances in Natural Language Processing (NLP) and Large Language Models (LLMs) now offer the opportunity to automatically evaluate such medical interviews, thereby alleviating the need for human examiners during training. Yet, real French OSCE annotated transcripts remain extremely scarce, limiting reproducible research and reliable benchmarking. To address these challenges, we investigate the use of LLMs for both generating and evaluating French OSCE dialogues in a low-resource context. We introduce a controlled pipeline that produces synthetic doctor-patient interview transcripts guided by scenario-specific evaluation criteria, combining ideal and perturbed performances to simulate varying student skill levels. The resulting dialogues are automatically silver-labeled through an LLM-assisted framework supporting adjustable evaluation strictness. Benchmarking multiple open-source and proprietary LLMs shows that mid-size models ($\le$32B parameters) achieve accuracies comparable to GPT-4o ($\sim$90\%) on synthetic data, highlighting the feasibility of locally deployable, privacy-preserving evaluation systems for medical education.

Raphaël Bagat, Irina Illina, Emmanuel Vincent

Automatic Speech Recognition (ASR) systems, despite large multilingual training, struggle in out-of-domain and low-resource scenarios where labeled data is scarce. We propose BEARD (BEST-RQ Encoder Adaptation with Re-training and Distillation), a novel framework designed to adapt Whisper's encoder using unlabeled data. Unlike traditional self-supervised learning methods, BEARD uniquely combines a BEST-RQ objective with knowledge distillation from a frozen teacher encoder, ensuring the encoder's complementarity with the pre-trained decoder. Our experiments focus on the ATCO2 corpus from the challenging Air Traffic Control (ATC) communications domain, characterized by non-native speech, noise, and specialized phraseology. Using about 5,000 hours of untranscribed speech for BEARD and 2 hours of transcribed speech for fine-tuning, the proposed approach significantly outperforms previous baseline and fine-tuned model, achieving a relative improvement of 12% compared to the fine-tuned model. To the best of our knowledge, this is the first work to use a self-supervised learning objective for domain adaptation of Whisper.

Yaya Sy, Christophe Cerisara, Irina Illina

Pruning large pre-trained transformers for low-resource languages is challenging, as it often requires massive retraining data to recover performance. For instance, Distill-Whisper prunes Whisper by 40% and retrains on 21,000 hours of speech, far beyond what is available for most languages. Can Whisper be made lighter and faster for edge devices in data-scarce settings? Focusing on Bambara with only 32h of speech-to-text data, we propose a new pruning recipe. Instead of vocabulary pruning, which is unsuitable due to frequent code-switching by Bambara speakers, we compress the embeddings with low-rank decomposition and feature distillation. Rather than removing layers, we merge them to limit performance loss. The final model preserves 90% of the original performance while being 48% smaller and 2.15x faster on a MacBook Air M1.

Yaya Sy, Dioula Doucouré, Christophe Cerisara et al.

We present our journey in training a speech language model for Wolof, an underrepresented language spoken in West Africa, and share key insights. We first emphasize the importance of collecting large-scale, spontaneous, high-quality unsupervised speech data, and show that continued pretraining HuBERT on this dataset outperforms both the base model and African-centric models on ASR. We then integrate this speech encoder into a Wolof LLM to train the first Speech LLM for this language, extending its capabilities to tasks such as speech translation. Furthermore, we explore training the Speech LLM to perform multi-step Chain-of-Thought before transcribing or translating. Our results show that the Speech LLM not only improves speech recognition but also performs well in speech translation. The models and the code will be openly shared.

Yaya Sy, Christophe Cerisara, Irina Illina

Current LLM structured pruning methods typically involve two steps: (1) compression with calibration data and (2) costly continued pretraining on billions of tokens to recover lost performance. This second step is necessary as the first significantly impacts model accuracy. Prior research suggests pretrained Transformer weights aren't inherently low-rank, unlike their activations, which may explain this drop. Based on this observation, we propose Lillama, a compression method that locally distills activations with low-rank weights. Using SVD for initialization and a joint loss combining teacher and student activations, we accelerate convergence and reduce memory use with local gradient updates. Lillama compresses Mixtral-8x7B within minutes on a single A100 GPU, removing 10 billion parameters while retaining over 95% of its original performance. Phi-2 3B can be compressed by 40% with just 13 million calibration tokens, resulting in a small model that competes with recent models of similar size. The method generalizes well to non-transformer architectures, compressing Mamba-3B by 20% while maintaining 99% performance.

Nicolas Furnon, Romain Serizel, Slim Essid et al.

Speech enhancement promises higher efficiency in ad-hoc microphone arrays than in constrained microphone arrays thanks to the wide spatial coverage of the devices in the acoustic scene. However, speech enhancement in ad-hoc microphone arrays still raises many challenges. In particular, the algorithms should be able to handle a variable number of microphones, as some devices in the array might appear or disappear. In this paper, we propose a solution that can efficiently process the spatial information captured by the different devices of the microphone array, while being robust to a link failure. To do this, we use an attention mechanism in order to put more weight on the relevant signals sent throughout the array and to neglect the redundant or empty channels.

Nicolas Zampieri, Irina Illina, Dominique Fohr

The task of automatically detecting hate speech in social media is gaining more and more attention. Given the enormous volume of content posted daily, human monitoring of hate speech is unfeasible. In this work, we propose new word-level features for automatic hate speech detection (HSD): multiword expressions (MWEs). MWEs are lexical units greater than a word that have idiomatic and compositional meanings. We propose to integrate MWE features in a deep neural network-based HSD framework. Our baseline HSD system relies on Universal Sentence Encoder (USE). To incorporate MWE features, we create a three-branch deep neural network: one branch for USE, one for MWE categories, and one for MWE embeddings. We conduct experiments on two hate speech tweet corpora with different MWE categories and with two types of MWE embeddings, word2vec and BERT. Our experiments demonstrate that the proposed HSD system with MWE features significantly outperforms the baseline system in terms of macro-F1.

Dominique Fohr, Irina Illina

The word error rate (WER) of an automatic speech recognition (ASR) system increases when a mismatch occurs between the training and the testing conditions due to the noise, etc. In this case, the acoustic information can be less reliable. This work aims to improve ASR by modeling long-term semantic relations to compensate for distorted acoustic features. We propose to perform this through rescoring of the ASR N-best hypotheses list. To achieve this, we train a deep neural network (DNN). Our DNN rescoring model is aimed at selecting hypotheses that have better semantic consistency and therefore lower WER. We investigate two types of representations as part of input features to our DNN model: static word embeddings (from word2vec) and dynamic contextual embeddings (from BERT). Acoustic and linguistic features are also included. We perform experiments on the publicly available dataset TED-LIUM mixed with real noise. The proposed rescoring approaches give significant improvement of the WER over the ASR system without rescoring models in two noisy conditions and with n-gram and RNNLM.

Nicolas Furnon, Romain Serizel, Irina Illina et al.

Multichannel processing is widely used for speech enhancement but several limitations appear when trying to deploy these solutions to the real-world. Distributed sensor arrays that consider several devices with a few microphones is a viable alternative that allows for exploiting the multiple devices equipped with microphones that we are using in our everyday life. In this context, we propose to extend the distributed adaptive node-specific signal estimation approach to a neural networks framework. At each node, a local filtering is performed to send one signal to the other nodes where a mask is estimated by a neural network in order to compute a global multi-channel Wiener filter. In an array of two nodes, we show that this additional signal can be efficiently taken into account to predict the masks and leads to better speech enhancement performances than when the mask estimation relies only on the local signals.

Ashwin Geet D'Sa, Irina Illina, Dominique Fohr

The spectacular expansion of the Internet has led to the development of a new research problem in the field of natural language processing: automatic toxic comment detection, since many countries prohibit hate speech in public media. There is no clear and formal definition of hate, offensive, toxic and abusive speeches. In this article, we put all these terms under the umbrella of "toxic" speech. The contribution of this paper is the design of binary classification and regression-based approaches aiming to predict whether a comment is toxic or not. We compare different unsupervised word representations and different DNN based classifiers. Moreover, we study the robustness of the proposed approaches to adversarial attacks by adding one (healthy or toxic) word. We evaluate the proposed methodology on the English Wikipedia Detox corpus. Our experiments show that using BERT fine-tuning outperforms feature-based BERT, Mikolov's and fastText representations with different DNN classifiers.

Imran Sheikh, Irina Illina, Dominique Fohr et al.

Many Proper Names (PNs) are Out-Of-Vocabulary (OOV) words for speech recognition systems used to process diachronic audio data. To help recovery of the PNs missed by the system, relevant OOV PNs can be retrieved out of the many OOVs by exploiting semantic context of the spoken content. In this paper, we propose two neural network models targeted to retrieve OOV PNs relevant to an audio document: (a) Document level Continuous Bag of Words (D-CBOW), (b) Document level Continuous Bag of Weighted Words (D-CBOW2). Both these models take document words as input and learn with an objective to maximise the retrieval of co-occurring OOV PNs. With the D-CBOW2 model we propose a new approach in which the input embedding layer is augmented with a context anchor layer. This layer learns to assign importance to input words and has the ability to capture (task specific) key-words in a bag-of-word neural network model. With experiments on French broadcast news videos we show that these two models outperform the baseline methods based on raw embeddings from LDA, Skip-gram and Paragraph Vectors. Combining the D-CBOW and D-CBOW2 models gives faster convergence during training.