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.

Storytelling is an integral part of human experience and plays a crucial role in social interactions. Thus, Automatic Story Evaluation (ASE) and Generation (ASG) could benefit society in multiple ways, but they are challenging tasks which require high-level human abilities such as creativity, reasoning, and deep understanding. Meanwhile, Large Language Models (LLMs) now achieve state-of-the-art performance on many NLP tasks. In this paper, we study whether LLMs can be used as substitutes for human annotators for ASE. We perform an extensive analysis of the correlations between LLM ratings, other automatic measures, and human annotations, and we explore the influence of prompting on the results and the explainability of LLM behaviour. Most notably, we find that LLMs outperform current automatic measures for system-level evaluation but still struggle at providing satisfactory explanations for their answers.

This study investigates the consequences of training language models on synthetic data generated by their predecessors, an increasingly prevalent practice given the prominence of powerful generative models. Diverging from the usual emphasis on performance metrics, we focus on the impact of this training methodology on linguistic diversity, especially when conducted recursively over time. To assess this, we adapt and develop a set of novel metrics targeting lexical, syntactic, and semantic diversity, applying them in recursive finetuning experiments across various natural language generation tasks in English. Our findings reveal a consistent decrease in the diversity of the model outputs through successive iterations, especially remarkable for tasks demanding high levels of creativity. This trend underscores the potential risks of training language models on synthetic text, particularly concerning the preservation of linguistic richness. Our study highlights the need for careful consideration of the long-term effects of such training approaches on the linguistic capabilities of language models.

In daily conversations, people often encounter problems prompting conversational repair to enhance mutual understanding. By employing an automatic coreference solver, alongside examining repetition, we identify various linguistic features that distinguish turns when the addressee initiates repair from those when they do not. Our findings reveal distinct patterns that characterize the repair sequence and each type of repair initiation.

We introduce MAFALDA, a benchmark for fallacy classification that merges and unites previous fallacy datasets. It comes with a taxonomy that aligns, refines, and unifies existing classifications of fallacies. We further provide a manual annotation of a part of the dataset together with manual explanations for each annotation. We propose a new annotation scheme tailored for subjective NLP tasks, and a new evaluation method designed to handle subjectivity. We then evaluate several language models under a zero-shot learning setting and human performances on MAFALDA to assess their capability to detect and classify fallacies.

Hedging is a strategy for softening the impact of a statement in conversation. In reducing the strength of an expression, it may help to avoid embarrassment (more technically, “face threat”) to one’s listener. For this reason, it is often found in contexts of instruction, such as tutoring. In this work, we develop a model of hedge generation based on i) fine-tuning state-of-the-art language models trained on human-human tutoring data, followed by ii) reranking to select the candidate that best matches the expected hedging strategy within a candidate pool using a hedge classifier. We apply this method to a natural peer-tutoring corpus containing a significant number of disfluencies, repetitions, and repairs. The results show that generation in this noisy environment is feasible with reranking. By conducting an error analysis for both approaches, we reveal the challenges faced by systems attempting to accomplish both social and task-oriented goals in conversation.

With the increasing amount of problematic peer reviews in top AI conferences, the community is urgently in need of automatic quality control measures. In this paper, we restrict our attention to substantiation — one popular quality aspect indicating whether the claims in a review are sufficiently supported by evidence — and provide a solution automatizing this evaluation process. To achieve this goal, we first formulate the problem as claim-evidence pair extraction in scientific peer reviews, and collect SubstanReview, the first annotated dataset for this task. SubstanReview consists of 550 reviews from NLP conferences annotated by domain experts. On the basis of this dataset, we train an argument mining system to automatically analyze the level of substantiation in peer reviews. We also perform data analysis on the SubstanReview dataset to obtain meaningful insights on peer reviewing quality in NLP conferences over recent years. The dataset is available at https://github.com/YanzhuGuo/SubstanReview.

This paper explores the application of machine learning techniques to predict where hedging occurs in peer-tutoring interactions. The study uses a naturalistic face-to-face dataset annotated for natural language turns, conversational strategies, tutoring strategies, and nonverbal behaviors. These elements are processed into a vector representation of the previous turns, which serves as input to several machine learning models, including MLP and LSTM. The results show that embedding layers, capturing the semantic information of the previous turns, significantly improves the model’s performance. Additionally, the study provides insights into the importance of various features, such as interpersonal rapport and nonverbal behaviors, in predicting hedges by using Shapley values for feature explanation. We discover that the eye gaze of both the tutor and the tutee has a significant impact on hedge prediction. We further validate this observation through a follow-up ablation study.

Depuis l’apparition des systèmes conversationnels, la modélisation des comportements humains constitue un axe de recherche majeur afin de renforcer l’expression des attributs émotionnels de ces systèmes. En nous intéressant aux agents conversationnels génératifs orientés-tâches, nous proposons une nouvelle approche pour rendre la réponse générée plus pertinente au contexte émotionnel de l’interlocuteur. Cette approche consiste à ajouter une étape supplémentaire de prédiction de labels pour conditionner la réponse générée et assurer sa pertinence au contexte socio-émotionnel de l’utilisateur. Nous proposons une formulation de cette nouvelle tâche de prédiction en nous appuyant sur un protocole d’annotation de données que nous avons conçu et implémenté. À travers cet article, nous apportons les contributions suivantes: la formulation de la tâche de prédiction de labels socio-émotionnels et la description du protocole d’annotation associé. Avec cette méthodologie, nous visons à développer des systèmes conversationnels socialement pertinents et indépendants.

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.

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.

Hedges have an important role in the management of rapport. In peer-tutoring, they are notably used by tutors in dyads experiencing low rapport to tone down the impact of instructions and negative feedback. Pursuing the objective of building a tutoring agent that manages rapport with teenagers in order to improve learning, we used a multimodal peer-tutoring dataset to construct a computational framework for identifying hedges. We compared approaches relying on pre-trained resources with others that integrate insights from the social science literature. Our best performance involved a hybrid approach that outperforms the existing baseline while being easier to interpret. We employ a model explainability tool to explore the features that characterize hedges in peer-tutoring conversations, and we identify some novel features, and the benefits of a such a hybrid model approach.

The topic of summarization evaluation has recently attracted a surge of attention due to the rapid development of abstractive summarization systems. However, the formulation of the task is rather ambiguous, neither the linguistic nor the natural language processing communities have succeeded in giving a mutually agreed-upon definition. Due to this lack of well-defined formulation, a large number of popular abstractive summarization datasets are constructed in a manner that neither guarantees validity nor meets one of the most essential criteria of summarization: factual consistency. In this paper, we address this issue by combining state-of-the-art factual consistency models to identify the problematic instances present in popular summarization datasets. We release SummFC, a filtered summarization dataset with improved factual consistency, and demonstrate that models trained on this dataset achieve improved performance in nearly all quality aspects. We argue that our dataset should become a valid benchmark for developing and evaluating summarization systems.

Logical reasoning on natural language is one of the most challenging tasks for deep learning models. There has been an increasing interest in developing new benchmarks to evaluate the reasoning capabilities of language models such as BERT. In parallel, new models based on transformers have emerged to achieve ever better performance on these datasets. However, there is currently no library for logical reasoning that includes such benchmarks and models. This paper introduces LogiTorch, a PyTorch-based library that includes different logical reasoning benchmarks, different models, as well as utility functions such as co-reference resolution. This makes it easy to directly use the preprocessed datasets, to run the models, or to finetune them with different hyperparameters. LogiTorch is open source and can be found on GitHub.

Transformer-based language models achieve state-of-the-art results on several natural language processing tasks. One of these is textual entailment, i.e., the task of determining whether a premise logically entails a hypothesis. However, the models perform poorly on this task when the examples contain negations. In this paper, we propose a new definition of textual entailment that captures also negation. This allows us to develop TINA (Textual Inference with Negation Augmentation), a principled technique for negated data augmentation that can be combined with the unlikelihood loss function.Our experiments with different transformer-based models show that our method can significantly improve the performance of the models on textual entailment datasets with negation – without sacrificing performance on datasets without negation.

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 present the process we used in order to collect new annotations of opinions over the multimodal corpus SEMAINE composed of dyadic interactions. The dataset had already been annotated continuously in two affective dimensions related to the emotions: Valence and Arousal. We annotated the part of SEMAINE called Solid SAL composed of 79 interactions between a user and an operator playing the role of a virtual agent designed to engage a person in a sustained, emotionally colored conversation. We aligned the audio at the word level using the available high-quality manual transcriptions. The annotated dataset contains 5627 speech turns for a total of 73,944 words, corresponding to 6 hours 20 minutes of dyadic interactions. Each interaction has been labeled by three annotators at the speech turn level following a three-step process. This method allows us to obtain a precise annotation regarding the opinion of a speaker. We obtain thus a dataset dense in opinions, with more than 48% of the annotated speech turns containing at least one opinion. We then propose a new baseline for the detection of opinions in interactions improving slightly a state of the art model with RoBERTa embeddings. The obtained results on the database are promising with a F1-score at 0.72.

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.

Research on Automatic Story Generation (ASG) relies heavily on human and automatic evaluation. However, there is no consensus on which human evaluation criteria to use, and no analysis of how well automatic criteria correlate with them. In this paper, we propose to re-evaluate ASG evaluation. We introduce a set of 6 orthogonal and comprehensive human criteria, carefully motivated by the social sciences literature. We also present HANNA, an annotated dataset of 1,056 stories produced by 10 different ASG systems. HANNA allows us to quantitatively evaluate the correlations of 72 automatic metrics with human criteria. Our analysis highlights the weaknesses of current metrics for ASG and allows us to formulate practical recommendations for ASG evaluation.

Learning disentangled representations of textual data is essential for many natural language tasks such as fair classification, style transfer and sentence generation, among others. The existent dominant approaches in the context of text data either rely on training an adversary (discriminator) that aims at making attribute values difficult to be inferred from the latent code or rely on minimising variational bounds of the mutual information between latent code and the value attribute. However, the available methods suffer of the impossibility to provide a fine-grained control of the degree (or force) of disentanglement. In contrast to adversarial methods, which are remarkably simple, although the adversary seems to be performing perfectly well during the training phase, after it is completed a fair amount of information about the undesired attribute still remains. This paper introduces a novel variational upper bound to the mutual information between an attribute and the latent code of an encoder. Our bound aims at controlling the approximation error via the Renyi’s divergence, leading to both better disentangled representations and in particular, a precise control of the desirable degree of disentanglement than state-of-the-art methods proposed for textual data. Furthermore, it does not suffer from the degeneracy of other losses in multi-class scenarios. We show the superiority of this method on fair classification and on textual style transfer tasks. Additionally, we provide new insights illustrating various trade-offs in style transfer when attempting to learn disentangled representations and quality of the generated sentence.

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.

A new metric BaryScore to evaluate text generation based on deep contextualized embeddings (e.g., BERT, Roberta, ELMo) is introduced. This metric is motivated by a new framework relying on optimal transport tools, i.e., Wasserstein distance and barycenter. By modelling the layer output of deep contextualized embeddings as a probability distribution rather than by a vector embedding; this framework provides a natural way to aggregate the different outputs through the Wasserstein space topology. In addition, it provides theoretical grounds to our metric and offers an alternative to available solutions (e.g., MoverScore and BertScore). Numerical evaluation is performed on four different tasks: machine translation, summarization, data2text generation and image captioning. Our results show that BaryScore outperforms other BERT based metrics and exhibits more consistent behaviour in particular for text summarization.

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.

Group cohesion is an emergent phenomenon that describes the tendency of the group members’ shared commitment to group tasks and the interpersonal attraction among them. This paper presents a multimodal analysis of group cohesion using a corpus of multi-party interactions. We utilize 16 two-minute segments annotated with cohesion from the AMI corpus. We define three layers of modalities: non-verbal social cues, dialogue acts and interruptions. The initial analysis is performed at the individual level and later, we combine the different modalities to observe their impact on perceived level of cohesion. Results indicate that occurrence of laughter and interruption are higher in high cohesive segments. We also observe that, dialogue acts and head nods did not have an impact on the level of cohesion by itself. However, when combined there was an impact on the perceived level of cohesion. Overall, the analysis shows that multimodal cues are crucial for accurate analysis of group cohesion.

One of the main challenges in the field of Embodied Conversational Agent (ECA) is to generate socially believable agents. The common strategy for agent behaviour synthesis is to rely on dedicated corpus analysis. Such a corpus is composed of multimedia files of socio-emotional behaviors which have been annotated by external observers. The underlying idea is to identify interaction information for the agent’s socio-emotional behavior by checking whether the intended socio-emotional behavior is actually perceived by humans. Then, the annotations can be used as learning classes for machine learning algorithms applied to the social signals. This paper introduces the POTUS Corpus composed of high-quality audio-video files of political addresses to the American people. Two protagonists are present in this database. First, it includes speeches of former president Barack Obama to the American people. Secondly, it provides videos of these same speeches given by a virtual agent named Rodrigue. The ECA reproduces the original address as closely as possible using social signals automatically extracted from the original one. Both are annotated for social attitudes, providing information about the stance observed in each file. It also provides the social signals automatically extracted from Obama’s addresses used to generate Rodrigue’s ones.

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.

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.

The task of predicting fine grained user opinion based on spontaneous spoken language is a key problem arising in the development of Computational Agents as well as in the development of social network based opinion miners. Unfortunately, gathering reliable data on which a model can be trained is notoriously difficult and existing works rely only on coarsely labeled opinions. In this work we aim at bridging the gap separating fine grained opinion models already developed for written language and coarse grained models developed for spontaneous multimodal opinion mining. We take advantage of the implicit hierarchical structure of opinions to build a joint fine and coarse grained opinion model that exploits different views of the opinion expression. The resulting model shares some properties with attention-based models and is shown to provide competitive results on a recently released multimodal fine grained annotated corpus.

This work aims at characterising verbal alignment processes for improving virtual agent communicative capabilities. We propose computationally inexpensive measures of verbal alignment based on expression repetition in dyadic textual dialogues. Using these measures, we present a contrastive study between Human-Human and Human-Agent dialogues on a negotiation task. We exhibit quantitative differences in the strength and orientation of verbal alignment showing the ability of our approach to characterise important aspects of verbal alignment.

Engagement is an important feature in human-human and human-agent interaction. In this paper, we investigate lexical alignment as a cue of engagement, relying on two different corpora : CID and SEMAINE. Our final goal is to build a virtual conversational character that could use alignment strategies to maintain user’s engagement. To do so, we investigate two alignment processes : shared vocabulary and other-repetitions. A quantitative and qualitative approach is proposed to characterize these aspects in human-human (CID) and human-operator (SEMAINE) interactions. Our results show that these processes are observable in both corpora, indicating a stable pattern that can be further modelled in conversational agents.

EDF utilise les techniques de Text Mining pour optimiser sa relation client, en analysant des réponses aux questions ouvertes d’enquête de satisfaction, et des retranscriptions de conversations issues des centres d’appels. Dans cet article, nous présentons les différentes contraintes applicatives liées à l’utilisation d’outils de text mining pour l’analyse de données clients. Après une analyse des différents outils présents sur le marché, nous avons identifié la technologie Skill CartridgeTM fournie par la société TEMIS comme la plus adaptée à nos besoins. Cette technologie nous permet une modélisation sémantique de concepts liés au motif d’insatisfaction. L’apport de cette modélisation est illustrée pour une tâche de classification de réponses d’enquêtes de satisfaction chargée d’évaluer la fidélité des clients EDF. La modélisation sémantique a permis une nette amélioration des scores de classification (F-mesure = 75,5%) notamment pour les catégories correspondant à la satisfaction et au mécontentement.

The present research focuses on annotation issues in the context of the acoustic detection of fear-type emotions for surveillance applications. The emotional speech material used for this study comes from the previously collected SAFE Database (Situation Analysis in a Fictional and Emotional Database) which consists of audio-visual sequences extracted from movie fictions. A generic annotation scheme was developed to annotate the various emotional manifestations contained in the corpus. The annotation was carried out by two labellers and the two annotations strategies are confronted. It emerges that the borderline between emotion and neutral vary according to the labeller. An acoustic validation by a third labeller allows at analysing the two strategies. Two human strategies are then observed: a first one, context-oriented which mixes audio and contextual (video) information in emotion categorization; and a second one, based mainly on audio information. The k-means clustering confirms the role of audio cues in human annotation strategies. It particularly helps in evaluating those strategies from the point of view of a detection system based on audio cues.