When deriving contextualized word representations from language models, a decision needs to be made on how to obtain one for out-of-vocabulary (OOV) words that are segmented into subwords. What is the best way to represent these words with a single vector, and are these representations of worse quality than those of in-vocabulary words? We carry out an intrinsic evaluation of embeddings from different models on semantic similarity tasks involving OOV words. Our analysis reveals, among other interesting findings, that the quality of representations of words that are split is often, but not always, worse than that of the embeddings of known words. Their similarity values, however, must be interpreted with caution.

The literature on general purpose textual Anomaly Detection is quite sparse, as most textual anomaly detection methods are implemented as out of domain detection in the context of pre-established classification tasks. Notably, in a field where pre-trained representations and models are of common use, the impact of the pre-training data on a task that lacks supervision has not been studied. In this paper, we use the simple setting of k-classes out anomaly detection and search for the best pairing of representation and classifier. We show that well-chosen embeddings allow a simple anomaly detection baseline such as OC-SVM to achieve similar results and even outperform deep state-of-the-art models.

La façon dont nous utilisons les mots est influencée par notre opinion. Nous cherchons à savoir si cela se reflète dans les plongements de mots contextualisés. Par exemple, la représentation d’ « animal » est-elle différente pour les gens qui voudraient abolir les zoos et ceux qui ne le voudraient pas ? Nous explorons cette question du point de vue du changement sémantique des mots. Nos expériences avec des représentations dérivées d’ensembles de données annotés avec les points de vue révèlent des différences minimes, mais significatives, entre postures opposées.

Cet article présente l’approche de l’équipe TTGV dans le cadre de sa participation aux deux tâches proposées lors du DEFT 2023 : l’identification du nombre de réponses supposément justes à un QCM et la prédiction de l’ensemble de réponses correctes parmi les cinq proposées pour une question donnée. Cet article présente les différentes méthodologies mises en oeuvre, explorant ainsi un large éventail d’approches et de techniques pour aborder dans un premier temps la distinction entre les questions appelant une seule ou plusieurs réponses avant de s’interroger sur l’identification des réponses correctes. Nous détaillerons les différentes méthodes utilisées, en mettant en exergue leurs avantages et leurs limites respectives. Ensuite, nous présenterons les résultats obtenus pour chaque approche. Enfin, nous discuterons des limitations intrinsèques aux tâches elles-mêmes ainsi qu’aux approches envisagées dans cette contribution.

Dialog participants sometimes align their linguistic styles, e.g., they use the same words and syntactic constructions as their interlocutors. We propose to investigate the notion of lexico-semantic alignment: to what extent do speakers convey the same meaning when they use the same words? We design measures of lexico-semantic alignment relying on contextualized word representations. We show that they reflect interesting semantic differences between the two sides of a debate and that they can assist in the task of debate’s winner prediction.

Cet article présente l’approche de l’équipe TGV lors de sa participation à la tâche de base de DEFT 2022, dont l’objectif était de prédire automatiquement les notes obtenues par des étudiants sur la base de leurs réponses à des questionnaires. Notre stratégie s’est focalisée sur la mise au point d’une méthode de classification des questions en fonction du type de réponse qu’elles attendent, de manière à pouvoir mener une approche différenciée pour chaque type. Nos trois runs ont consisté en une approche non différenciée, servant de référence, et deux approches différenciées, la première se basant sur la constitution d’un jeu de caractéristiques et la seconde sur le calcul de TF-IDF et de la fonction de hashage. Notre objectif premier était ainsi de vérifier si des approches dédiées à chaque type de questions sont préférables à une approche globale.

The way we use words is influenced by our opinion. We investigate whether this is reflected in contextualized word embeddings. For example, is the representation of “animal” different between people who would abolish zoos and those who would not? We explore this question from a Lexical Semantic Change standpoint. Our experiments with BERT embeddings derived from datasets with stance annotations reveal small but significant differences in word representations between opposing stances.

Our discourses are full of potential lexical ambiguities, due in part to the pervasive use of words having multiple senses. Sometimes, one word may even be used in more than one sense throughout a text. But, to what extent is this true for different kinds of texts? Does the use of polysemous words change when a discourse involves two people, or when speakers have time to plan what to say? We investigate these questions by comparing the polysemy level of texts of different nature, with a focus on spontaneous spoken dialogs; unlike previous work which examines solely scripted, written, monolog-like data. We compare multiple metrics that presuppose different conceptualizations of text polysemy, i.e., they consider the observed or the potential number of senses of words, or their sense distribution in a discourse. We show that the polysemy level of texts varies greatly depending on the kind of text considered, with dialog and spoken discourses having generally a higher polysemy level than written monologs. Additionally, our results emphasize the need for relaxing the popular “one sense per discourse” hypothesis.

Users generate content constantly, leading to new data requiring annotation. Among this data, textual conversations are created every day and come with some specificities: they are mostly private through instant messaging applications, requiring the conversational context to be labeled. These specificities led to several annotation tools dedicated to conversation, and mostly dedicated to dialogue tasks, requiring complex annotation schemata, not always customizable and not taking into account conversation-level labels. In this paper, we present EZCAT, an easy-to-use interface to annotate conversations in a two-level configurable schema, leveraging message-level labels and conversation-level labels. Our interface is characterized by the voluntary absence of a server and accounts management, enhancing its availability to anyone, and the control over data, which is crucial to confidential conversations. We also present our first usage of EZCAT along with our annotation schema we used to annotate confidential customer service conversations. EZCAT is freely available at https://gguibon.github.io/ezcat.

In this paper, we place ourselves in a classification scenario in which the target classes and data type are not accessible during training. We use a meta-learning approach to determine whether or not meta-trained information from common social network data with fine-grained emotion labels can achieve competitive performance on messages labeled with different emotion categories. We leverage few-shot learning to match with the classification scenario and consider metric learning based meta-learning by setting up Prototypical Networks with a Transformer encoder, trained in an episodic fashion. This approach proves to be effective for capturing meta-information from a source emotional tag set to predict previously unseen emotional tags. Even though shifting the data type triggers an expected performance drop, our meta-learning approach achieves decent results when compared to the fully supervised one.

Multimodal sentiment analysis is a trending area of research, and multimodal fusion is one of its most active topic. Acknowledging humans communicate through a variety of channels (i.e visual, acoustic, linguistic), multimodal systems aim at integrating different unimodal representations into a synthetic one. So far, a consequent effort has been made on developing complex architectures allowing the fusion of these modalities. However, such systems are mainly trained by minimising simple losses such as L₁ or cross-entropy. In this work, we investigate unexplored penalties and propose a set of new objectives that measure the dependency between modalities. We demonstrate that our new penalties lead to a consistent improvement (up to 4.3 on accuracy) across a large variety of state-of-the-art models on two well-known sentiment analysis datasets: CMU-MOSI and CMU-MOSEI. Our method not only achieves a new SOTA on both datasets but also produces representations that are more robust to modality drops. Finally, a by-product of our methods includes a statistical network which can be used to interpret the high dimensional representations learnt by the model.

Several recent studies on dyadic human-human interactions have been done on conversations without specific business objectives. However, many companies might benefit from studies dedicated to more precise environments such as after sales services or customer satisfaction surveys. In this work, we place ourselves in the scope of a live chat customer service in which we want to detect emotions and their evolution in the conversation flow. This context leads to multiple challenges that range from exploiting restricted, small and mostly unlabeled datasets to finding and adapting methods for such context. We tackle these challenges by using Few-Shot Learning while making the hypothesis it can serve conversational emotion classification for different languages and sparse labels. We contribute by proposing a variation of Prototypical Networks for sequence labeling in conversation that we name ProtoSeq. We test this method on two datasets with different languages: daily conversations in English and customer service chat conversations in French. When applied to emotion classification in conversations, our method proved to be competitive even when compared to other ones.

Spoken dialogue systems need to be able to handle both multiple languages and multilinguality inside a conversation (e.g in case of code-switching). In this work, we introduce new pretraining losses tailored to learn generic multilingual spoken dialogue representations. The goal of these losses is to expose the model to code-switched language. In order to scale up training, we automatically build a pretraining corpus composed of multilingual conversations in five different languages (French, Italian, English, German and Spanish) from OpenSubtitles, a huge multilingual corpus composed of 24.3G tokens. We test the generic representations on MIAM, a new benchmark composed of five dialogue act corpora on the same aforementioned languages as well as on two novel multilingual tasks (i.e multilingual mask utterance retrieval and multilingual inconsistency identification). Our experiments show that our new losses achieve a better performance in both monolingual and multilingual settings.

Dans cet article nous reproduisons un scénario d’apprentissage selon lequel les données cibles ne sont pas accessibles et seules des données connexes le sont. Nous utilisons une approche par méta-apprentissage afin de déterminer si les méta-informations apprises à partir de messages issus de médias sociaux, finement annotés en émotions, peuvent produire de bonnes performances une fois utilisées sur des messages issus de conversations, étiquetés en émotions avec une granularité différente. Nous mettons à profit l’apprentissage sur quelques exemples (few-shot learning) pour la mise en place de ce scénario. Cette approche se montre efficace pour capturer les méta-informations d’un jeu d’étiquettes émotionnelles pour prédire des étiquettes jusqu’alors inconnues au modèle. Bien que le fait de varier le type de données engendre une baisse de performance, notre approche par méta-apprentissage atteint des résultats décents comparés au référentiel d’apprentissage supervisé.

While being an essential component of spoken language, fillers (e.g. “um” or “uh”) often remain overlooked in Spoken Language Understanding (SLU) tasks. We explore the possibility of representing them with deep contextualised embeddings, showing improvements on modelling spoken language and two downstream tasks — predicting a speaker’s stance and expressed confidence.

Sequence labelling tasks like Dialog Act and Emotion/Sentiment identification are a key component of spoken dialog systems. In this work, we propose a new approach to learn generic representations adapted to spoken dialog, which we evaluate on a new benchmark we call Sequence labellIng evaLuatIon benChmark fOr spoken laNguagE benchmark (SILICONE). SILICONE is model-agnostic and contains 10 different datasets of various sizes. We obtain our representations with a hierarchical encoder based on transformer architectures, for which we extend two well-known pre-training objectives. Pre-training is performed on OpenSubtitles: a large corpus of spoken dialog containing over 2.3 billion of tokens. We demonstrate how hierarchical encoders achieve competitive results with consistently fewer parameters compared to state-of-the-art models and we show their importance for both pre-training and fine-tuning.

Neural language models are usually trained using Maximum-Likelihood Estimation (MLE). The corresponding objective function for MLE is derived from the Kullback-Leibler (KL) divergence between the empirical probability distribution representing the data and the parametric probability distribution output by the model. However, the word frequency discrepancies in natural language make performance extremely uneven: while the perplexity is usually very low for frequent words, it is especially difficult to predict rare words. In this paper, we experiment with several families (alpha, beta and gamma) of power divergences, generalized from the KL divergence, for learning language models with an objective different than standard MLE. Intuitively, these divergences should affect the way the probability mass is spread during learning, notably by prioritizing performances on high or low-frequency words. In addition, we implement and experiment with various sampling-based objectives, where the computation of the output layer is only done on a small subset of the vocabulary. They are derived as power generalizations of a softmax approximated via Importance Sampling, and Noise Contrastive Estimation, for accelerated learning. Our experiments on the Penn Treebank and Wikitext-2 show that these power divergences can indeed be used to prioritize learning on the frequent or rare words, and lead to general performance improvements in the case of sampling-based learning.

Noise-Contrastive Estimation (NCE) is a learning criterion that is regularly used to train neural language models in place of Maximum Likelihood Estimation, since it avoids the computational bottleneck caused by the output softmax. In this paper, we analyse and explain some of the weaknesses of this objective function, linked to the mechanism of self-normalization, by closely monitoring comparative experiments. We then explore several remedies and modifications to propose tractable and efficient NCE training strategies. In particular, we propose to make the scaling factor a trainable parameter of the model, and to use the noise distribution to initialize the output bias. These solutions, yet simple, yield stable and competitive performances in either small and large scale language modelling tasks.

L’estimation contrastive bruitée (NCE) et l’échantillonage par importance (IS) sont des procédures d’entraînement basées sur l’échantillonage, que l’on utilise habituellement à la place de l’estimation du maximum de vraisemblance (MLE) pour éviter le calcul du softmax lorsque l’on entraîne des modèles de langue neuronaux. Dans cet article, nous cherchons à résumer le fonctionnement de ces algorithmes, et leur utilisation dans la littérature du TAL. Nous les comparons expérimentalement, et présentons des manières de faciliter l’entraînement du NCE.

Cet article propose une architecture neuronale pour un modèle de langue à vocabulaire ouvert. Les représentations continues des mots sont calculées à la volée à partir des caractères les composant, gràce à une couche convolutionnelle suivie d’une couche de regroupement (pooling). Cela permet au modèle de représenter n’importe quel mot, qu’il fasse partie du contexte ou soit évalué pour la prédiction. La fonction objectif est dérivée de l’estimation contrastive bruitée (Noise Contrastive Estimation, ou NCE), calculable dans notre cas sans vocabulaire. Nous évaluons la capacité de notre modèle à construire des représentations continues de mots inconnus sur la tâche de traduction automatique IWSLT-2016, de l’Anglais vers le Tchèque, en ré-évaluant les N meilleures hypothèses (N-best reranking). Les résultats expérimentaux permettent des gains jusqu’à 0,7 point BLEU. Ils montrent aussi la difficulté d’utiliser des représentations dérivées des caractères pour la prédiction.

Ce travail cherche à comprendre pourquoi les performances d’un analyseur morpho-syntaxiques chutent fortement lorsque celui-ci est utilisé sur des données hors domaine. Nous montrons à l’aide d’une expérience jouet que ce comportement peut être dû à un phénomène de masquage des caractéristiques lexicalisées par les caractéristiques non lexicalisées. Nous proposons plusieurs modèles essayant de réduire cet effet.

Noise Contrastive Estimation (NCE) is a learning procedure that is regularly used to train neural language models, since it avoids the computational bottleneck caused by the output softmax. In this paper, we attempt to explain some of the weaknesses of this objective function, and to draw directions for further developments. Experiments on a small task show the issues raised by an unigram noise distribution, and that a context dependent noise distribution, such as the bigram distribution, can solve these issues and provide stable and data-efficient learning.

Most of neural language models use different kinds of embeddings for word prediction. While word embeddings can be associated to each word in the vocabulary or derived from characters as well as factored morphological decomposition, these word representations are mainly used to parametrize the input, i.e. the context of prediction. This work investigates the effect of using subword units (character and factored morphological decomposition) to build output representations for neural language modeling. We present a case study on Czech, a morphologically-rich language, experimenting with different input and output representations. When working with the full training vocabulary, despite unstable training, our experiments show that augmenting the output word representations with character-based embeddings can significantly improve the performance of the model. Moreover, reducing the size of the output look-up table, to let the character-based embeddings represent rare words, brings further improvement.

This paper describes LIMSI’s submission to the MT track of IWSLT 2016. We report results for translation from English into Czech. Our submission is an attempt to address the difficulties of translating into a morphologically rich language by paying special attention to the morphology generation on target side. To this end, we propose two ways of improving the morphological fluency of the output: 1. by performing translation and inflection of the target language in two separate steps, and 2. by using a neural language model with characted-based word representation. We finally present the combination of both methods used for our primary system submission.