Simplifying complex texts is essential to ensure equitable access to information, particularly for individuals with cognitive impairments. The Easy-to-Read (ETR) initiative provides a framework to make content more accessible for these individuals. However, manually creating such texts remains time-consuming and resource-intensive. In this work, we investigate the potential of large language models (LLMs) to automate the generation of ETR content. To address the scarcity of aligned corpora and the specific constraints of ETR, we propose a multi-task learning (MTL) approach that trains models jointly on text summarization, text simplification, and ETR generation. We explore two complementary strategies: multi-task retrieval-augmented generation (RAG) for in-context learning (ICL), and MTL-LoRA for parameter-efficient fine-tuning (PEFT). Our experiments with Mistral-7B and LLaMA-3-8B, conducted on ETR-fr, a new high-quality dataset, show that MTL-LoRA consistently outperforms all other strategies in in-domain settings, while the MTL-RAG-based approach achieves better generalization in out-of-domain scenarios. Our code is publicly available at https://github.com/FrLdy/ETR-PEFT-Composition.

This paper explores the impact of incorporating sentiment, emotion, and domain-specific lexicons into a transformer-based model for depression symptom estimation. Lexicon information is added by marking the words in the input transcripts of patient-therapist conversations as well as in social media posts. Overall results show that the introduction of external knowledge within pre-trained language models can be beneficial for prediction performance, while different lexicons show distinct behaviours depending on the targeted task. Additionally, new state-of-the-art results are obtained for the estimation of depression level over patient-therapist interviews.

Automated depression estimation has received significant research attention in recent years as a result of its growing impact on the global community. Within the context of studies based on patient-therapist interview transcripts, most researchers treat the dyadic discourse as a sequence of unstructured sentences, thus ignoring the discourse structure within the learning process. In this paper we propose Multi-view architectures that divide the input transcript into patient and therapist views based on sentence type in an attempt to utilize symmetric discourse structure for improved model performance. Experiments on DAIC-WOZ dataset for binary classification task within depression estimation show advantages of Multi-view architecture over sequential input representations. Our model also outperforms the current state-of-the-art results and provide new SOTA performance on test set of DAIC-WOZ dataset.

This paper addresses the quality of annotations in mental health datasets used for NLP-based depression level estimation from social media texts. While previous research relies on social media-based datasets annotated with binary categories, i.e. depressed or non-depressed, recent datasets such as D2S and PRIMATE aim for nuanced annotations using PHQ-9 symptoms. However, most of these datasets rely on crowd workers without the domain knowledge for annotation. Focusing on the PRIMATE dataset, our study reveals concerns regarding annotation validity, particularly for the lack of interest or pleasure symptom. Through reannotation by a mental health professional, we introduce finer labels and textual spans as evidence, identifying a notable number of false positives. Our refined annotations, to be released under a Data Use Agreement, offer a higher-quality test set for anhedonia detection. This study underscores the necessity of addressing annotation quality issues in mental health datasets, advocating for improved methodologies to enhance NLP model reliability in mental health assessments.

The ever-growing number of people suffering from mental distress has motivated significant research initiatives towards automated depression estimation. Despite the multidisciplinary nature of the task, very few of these approaches include medical professionals in their research process, thus ignoring a vital source of domain knowledge. In this paper, we propose to bring the domain experts back into the loop and incorporate their knowledge within the gold-standard DAIC-WOZ dataset. In particular, we define a novel transformer-based architecture and analyse its performance in light of our expert annotations. Overall findings demonstrate a strong correlation between the psychological tendencies of medical professionals and the behavior of the proposed model, which additionally provides new state-of-the-art results.

Humorous texts can be of different forms such as punchlines, puns, or funny stories. Existing humor classification systems have been dealing with such diverse forms by treating them independently. In this paper, we argue that different forms of humor share a common background either in terms of vocabulary or constructs. As a consequence, it is likely that classification performance can be improved by jointly tackling different humor types. Hence, we design a shared-private multitask architecture following a transfer learning paradigm and perform experiments over four gold standard datasets. Empirical results steadily confirm our hypothesis by demonstrating statistically-significant improvements over baselines and accounting for new state-of-the-art figures for two datasets.

Nous proposons une idée originale pour exploiter les relations entre les classes dans les problèmes multiclasses. Nous définissons deux architectures multitâches de type one-vs-rest qui combinent des ensembles de classifieurs appris dans une configuration multitâche en utilisant des réseaux de neurones. Les expériences menées sur six jeux de données pour la classification des sentiments, des émotions, des thématiques et des relations lexico-sémantiques montrent que nos architectures améliorent constamment les performances par rapport aux stratégies de l’état de l’art de type one-vsrest et concurrencent fortement les autres stratégies multiclasses.

Most social media platforms grant users freedom of speech by allowing them to freely express their thoughts, beliefs, and opinions. Although this represents incredible and unique communication opportunities, it also presents important challenges. Online racism is such an example. In this study, we present a supervised learning strategy to detect racist language on Twitter based on word embedding that incorporate demographic (Age, Gender, and Location) information. Our methodology achieves reasonable classification accuracy over a gold standard dataset (F1=76.3%) and significantly improves over the classification performance of demographic-agnostic models.

In this paper we focus on a particular case of entailment, namely entailment by generality. We argue that there exist various types of implication, a range of different levels of entailment reasoning, based on lexical, syntactic, logical and common sense clues, at different levels of difficulty. We introduce the paradigm of Textual Entailment (TE) by Generality, which can be defined as the entailment from a specific statement towards a relatively more general statement. In this context, the Text T entails the Hypothesis H, and at the same time H is more general than T . We propose an unsupervised and language-independent method to recognize TE by Generality given a case of Text − Hypothesis or T − H where entailment relation holds.

In this work we introduce a particular case of textual entailment (TE), namely Textual Entailment by Generality (TEG). In text, there are different kinds of entailment yielded from different types of implicative reasoning (lexical, syntactic, common sense based), but here we focus just on TEG, which can be defined as an entailment from a specific statement towards a relatively more G general one. Therefore, we have T (G)→ H whenever the premise T entails the hypothesis H, the hypothesis being more general than the premise. We propose an unsupervised and language-independent method to recognize TEGs, given a pair T, H in an entailment relation. We have evaluated our proposal G → H English pairs, where we know through two experiments: (a) Test on T (G)→ H English pairs, where we know that TEG holds; (b) Test on T → H Portuguese pairs, randomly selected with 60% of TEGs and 40% of TE without generality dependency (TEnG).

Some authors (Simard et al.; Melamed; Danielsson & Mühlenbock) have suggested measures of similarity of words in different languages so as to find extra clues for alignment of parallel texts. Cognate words, like ‘Parliament’ and ‘Parlement’, in English and French respectively, provide extra anchors that help to improve the quality of the alignment. In this paper, we will extend an alignment algorithm proposed by Ribeiro et al. using typical contiguous and non-contiguous sequences of characters extracted using a statistically sound method (Dias et al.). With these typical sequences, we are able to find more reliable correspondence points and improve the alignment quality without recurring to heuristics to identify cognates.