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.

Multiword expression (MWE) identification in tweets is a complex task due to the complex linguistic nature of MWEs combined with the non-standard language use in social networks. MWE features were shown to be helpful for hate speech detection (HSD). In this article, we present joint experiments on these two related tasks on English Twitter data: first we focus on the MWE identification task, and then we observe the influence of MWE-based features on the HSD task. For MWE identification, we compare the performance of two systems: lexicon-based and deep neural networks-based (DNN). We experimentally evaluate seven configurations of a state-of-the-art DNN system based on recurrent networks using pre-trained contextual embeddings from BERT. The DNN-based system outperforms the lexicon-based one thanks to its superior generalisation power, yielding much better recall. For the HSD task, we propose a new DNN architecture for incorporating MWE features. We confirm that MWE features are helpful for the HSD task. Moreover, the proposed DNN architecture beats previous MWE-based HSD systems by 0.4 to 1.1 F-measure points on average on four Twitter HSD corpora.

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.

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.

L’identification des expressions polylexicales (EP) dans les tweets est une tâche difficile en raison de la nature linguistique complexe des EP combinée à l’utilisation d’un langage non standard. Dans cet article, nous présentons cette tâche d’identification sur des données anglaises de Twitter. Nous comparons les performances de deux systèmes : un utilisant un dictionnaire et un autre des réseaux de neurones. Nous évaluons expérimentalement sept configurations d’un système état de l’art fondé sur des réseaux neuronaux récurrents utilisant des embeddings contextuels générés par BERT. Le système fondé sur les réseaux neuronaux surpasse l’approche dictionnaire, collecté automatiquement à partir des EP dans des corpus, grâce à son pouvoir de généralisation supérieur.

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.

Rapidly changing social media content calls for robust and generalisable abuse detection models. However, the state-of-the-art supervised models display degraded performance when they are evaluated on abusive comments that differ from the training corpus. We investigate if the performance of supervised models for cross-corpora abuse detection can be improved by incorporating additional information from topic models, as the latter can infer the latent topic mixtures from unseen samples. In particular, we combine topical information with representations from a model tuned for classifying abusive comments. Our performance analysis reveals that topic models are able to capture abuse-related topics that can transfer across corpora, and result in improved generalisability.

The state-of-the-art abusive language detection models report great in-corpus performance, but underperform when evaluated on abusive comments that differ from the training scenario. As human annotation involves substantial time and effort, models that can adapt to newly collected comments can prove to be useful. In this paper, we investigate the effectiveness of several Unsupervised Domain Adaptation (UDA) approaches for the task of cross-corpora abusive language detection. In comparison, we adapt a variant of the BERT model, trained on large-scale abusive comments, using Masked Language Model (MLM) fine-tuning. Our evaluation shows that the UDA approaches result in sub-optimal performance, while the MLM fine-tuning does better in the cross-corpora setting. Detailed analysis reveals the limitations of the UDA approaches and emphasizes the need to build efficient adaptation methods for this task.

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.

Research on hate speech classification has received increased attention. In real-life scenarios, a small amount of labeled hate speech data is available to train a reliable classifier. Semi-supervised learning takes advantage of a small amount of labeled data and a large amount of unlabeled data. In this paper, label propagation-based semi-supervised learning is explored for the task of hate speech classification. The quality of labeling the unlabeled set depends on the input representations. In this work, we show that pre-trained representations are label agnostic, and when used with label propagation yield poor results. Neural network-based fine-tuning can be adopted to learn task-specific representations using a small amount of labeled data. We show that fully fine-tuned representations may not always be the best representations for the label propagation and intermediate representations may perform better in a semi-supervised setup.

Dans cet article nous proposons une méthode d’adaptation du lexique, destinée à améliorer les systèmes de la reconnaissance automatique de la parole (SRAP) des locuteurs non natifs. En effet, la reconnaissance automatique souffre d’une chute significative de ses performances quand elle est utilisée pour reconnaître la parole des locuteurs non natifs, car les phonèmes de la langue étrangère sont fréquemment mal prononcés par ces locuteurs. Pour prendre en compte ce problème de prononciations erronées, notre approche propose d’intégrer les prononciations non natives dans le lexique et par la suite d’utiliser ce lexique enrichi pour la reconnaissance. Pour réaliser notre approche nous avons besoin d’un petit corpus de parole non native et de sa transcription. Pour générer les prononciations non natives, nous proposons de tenir compte des correspondances graphèmes-phonèmes en vue de générer de manière automatique des règles de création de nouvelles prononciations. Ces nouvelles prononciations seront ajoutées au lexique. Nous présentons une évaluation de notre méthode sur un corpus de locuteurs non natifs français s’exprimant en anglais.

Malgré les avancés spectaculaires ces dernières années, les systèmes de Reconnaissance Automatique de Parole (RAP) commettent encore des erreurs, surtout dans des environnements bruités. Pour améliorer la RAP, nous proposons de se diriger vers une contextualisation d’un système RAP, car les informations sémantiques sont importantes pour la performance de la RAP. Les systèmes RAP actuels ne prennent en compte principalement que les informations lexicales et syntaxiques. Pour modéliser les informations sémantiques, nous proposons de détecter les mots de la phrase traitée qui pourraient avoir été mal reconnus et de proposer des mots correspondant mieux au contexte. Cette analyse sémantique permettra de réévaluer les N meilleures hypothèses de transcription (N-best). Nous utilisons les embeddings Word2Vec et BERT. Nous avons évalué notre méthodologie sur le corpus des conférences TED (TED-LIUM). Les résultats montrent une amélioration significative du taux d’erreur mots en utilisant la méthodologie proposée.

La démonstration de résumé et de traduction automatique de vidéos résulte de nos travaux dans le projet AMIS. L’objectif du projet était d’aider un voyageur à comprendre les nouvelles dans un pays étranger. Pour cela, le projet propose de résumer et traduire automatiquement une vidéo en langue étrangère (ici, l’arabe). Un autre objectif du projet était aussi de comparer les opinions et sentiments exprimés dans plusieurs vidéos comparables. La démonstration porte sur l’aspect résumé, transcription et traduction. Les exemples montrés permettront de comprendre et mesurer qualitativement les résultats du projet.

Automatic speech recognition for Arabic is a very challenging task. Despite all the classical techniques for Automatic Speech Recognition (ASR), which can be efficiently applied to Arabic speech recognition, it is essential to take into consideration the language specificities to improve the system performance. In this article, we focus on Modern Standard Arabic (MSA) speech recognition. We introduce the challenges related to Arabic language, namely the complex morphology nature of the language and the absence of the short vowels in written text, which leads to several potential vowelization for each graphemes, which is often conflicting. We develop an ASR system for MSA by using Kaldi toolkit. Several acoustic and language models are trained. We obtain a Word Error Rate (WER) of 14.42 for the baseline system and 12.2 relative improvement by rescoring the lattice and by rewriting the output with the right Z hamoza above or below Alif.

The IFCASL corpus is a French-German bilingual phonetic learner corpus designed, recorded and annotated in a project on individualized feedback in computer-assisted spoken language learning. The motivation for setting up this corpus was that there is no phonetically annotated and segmented corpus for this language pair of comparable of size and coverage. In contrast to most learner corpora, the IFCASL corpus incorporate data for a language pair in both directions, i.e. in our case French learners of German, and German learners of French. In addition, the corpus is complemented by two sub-corpora of native speech by the same speakers. The corpus provides spoken data by about 100 speakers with comparable productions, annotated and segmented on the word and the phone level, with more than 50% manually corrected data. The paper reports on inter-annotator agreement and the optimization of the acoustic models for forced speech-text alignment in exercises for computer-assisted pronunciation training. Example studies based on the corpus data with a phonetic focus include topics such as the realization of /h/ and glottal stop, final devoicing of obstruents, vowel quantity and quality, pitch range, and tempo.

Out-Of-Vocabulary (OOV) words missed by Large Vocabulary Continuous Speech Recognition (LVCSR) systems can be recovered with the help of topic and semantic context of the OOV words captured from a diachronic text corpus. In this paper we investigate how the choice of documents for the diachronic text corpora affects the retrieval of OOV Proper Names (PNs) relevant to an audio document. We first present our diachronic French broadcast news datasets, which highlight the motivation of our study on OOV PNs. Then the effect of using diachronic text data from different sources and a different time span is analysed. With OOV PN retrieval experiments on French broadcast news videos, we conclude that a diachronic corpus with text from different sources leads to better retrieval performance than one relying on text from single source or from a longer time span.

This work proposes a new confidence measure for evaluating text-to-speech alignment systems outputs, which is a key component for many applications, such as semi-automatic corpus anonymization, lips syncing, film dubbing, corpus preparation for speech synthesis and speech recognition acoustic models training. This confidence measure exploits deep neural networks that are trained on large corpora without direct supervision. It is evaluated on an open-source spontaneous speech corpus and outperforms a confidence score derived from a state-of-the-art text-to-speech aligner. We further show that this confidence measure can be used to fine-tune the output of this aligner and improve the quality of the resulting alignment.

We present the design of a corpus of native and non-native speech for the language pair French-German, with a special emphasis on phonetic and prosodic aspects. To our knowledge there is no suitable corpus, in terms of size and coverage, currently available for the target language pair. To select the target L1-L2 interference phenomena we prepare a small preliminary corpus (corpus1), which is analyzed for coverage and cross-checked jointly by French and German experts. Based on this analysis, target phenomena on the phonetic and phonological level are selected on the basis of the expected degree of deviation from the native performance and the frequency of occurrence. 14 speakers performed both L2 (either French or German) and L1 material (either German or French). This allowed us to test, recordings duration, recordings material, the performance of our automatic aligner software. Then, we built corpus2 taking into account what we learned about corpus1. The aims are the same but we adapted speech material to avoid too long recording sessions. 100 speakers will be recorded. The corpus (corpus1 and corpus2) will be prepared as a searchable database, available for the scientific community after completion of the project.

Speech-text alignment tools are frequently used in speech technology and research. In this paper, we propose a GPL software CoALT (Comparing Automatic Labelling Tools) for comparing two automatic labellers or two speech-text alignment tools, ranking them and displaying statistics about their differences. The main feature of CoALT is that a user can define its own criteria for evaluating and comparing the speech-text alignment tools since the required quality for labelling depends on the targeted application. Beyond ranking, our tool provides useful statistics for each labeller and above all about their differences and can emphasize the drawbacks and advantages of each labeller. We have applied our software for the French and English languages but it can be used for another language by simply defining the list of the phonetic symbols and optionally a set of phonetic rules. In this paper we present the usage of the software for comparing two automatic labellers on the corpus TIMIT. Moreover, as automatic labelling tools are configurable (number of GMMs, phonetic lexicon, acoustic parameterisation), we then present how CoALT allows to determine the best parameters for our automatic labelling tool.