Discourse relation classification within a multilingual, cross-framework setting is a challenging task, and the best-performing systems so far have relied on monolingual and mono-framework approaches.In this paper, we introduce transformer-based multilingual models, trained jointly over all datasets—thus covering different languages and discourse frameworks. We demonstrate their ability to outperform single-corpus models and to overcome (to some extent) the disparity among corpora, by relying on linguistic features and generic information about the nature of the datasets. We also compare the performance of different multilingual pretrained models, as well as the encoding of the relation direction, a key component for the task. Our results on the 16 datasets of the DISRPT 2021 benchmark show improvements in accuracy in (almost) all datasets compared to the monolingual models, with at best 65.91% in average accuracy, thus corresponding to a 4% improvement over the state-of-the-art.

Classifying discourse relations is known as a hard task, relying on complex indices. On the other hand, discourse-annotated data is scarce, especially for languages other than English: many corpora, of limited size, exist for several languages but the domain is split between different theoretical frameworks that have a huge impact on the nature of the textual spans to be linked, and the label set used. Moreover, each annotation project implements modifications compared to the theoretical background and other projects. These discrepancies hinder the development of systems taking advantage of all the available data to tackle data sparsity and work on transfer between languages is very limited, almost nonexistent between frameworks, while it could improve our understanding of some theoretical aspects and enhance many applications. In this paper, we propose the first experiments on zero-shot learning for discourse relation classification and investigate several paths in the way source data can be combined, either based on languages, frameworks, or similarity measures. We demonstrate how difficult transfer is for the task at hand, and that the most impactful factor is label set divergence, where the notion of underlying framework possibly conceals crucial disagreements.

In 2023, the third iteration of the DISRPT Shared Task (Discourse Relation Parsing and Treebanking) was held, dedicated to the underlying units used in discourse parsing across formalisms. Following the success of the 2019and 2021 tasks on Elementary Discourse Unit Segmentation, Connective Detection, and Relation Classification, this iteration has added 10 new corpora, including 2 new languages (Thai and Italian) and 3 discourse treebanks annotated in the discourse dependency representation in addition to the previously included frameworks: RST, SDRT, and PDTB. In this paper, we review the data included in the Shared Task, which covers 26 datasets across 13 languages, survey and compare submitted systems, and report on system performance on each task for both annotated and plain-tokenized versions of the data.

This paper presents the results obtained by the MELODI team for the three tasks proposed within the DISRPT 2023 shared task on discourse: segmentation, connective identification, and relation classification. The competition involves corpora in various languages in several underlying frameworks, and proposes two tracks depending on the presence or not of annotations of sentence boundaries and syntactic information. For these three tasks, we rely on a transformer-based architecture, and investigate several optimizations of the models, including hyper-parameter search and layer freezing. For discourse relations, we also explore the use of adapters—a lightweight solution for model fine-tuning—and introduce relation mappings to partially deal with the label set explosion we are facing within the setting of the shared task in a multi-corpus perspective. In the end, we propose one single architecture for segmentation and connectives, based on XLM-RoBERTa large, freezed at lower layers, with new state-of-the-art results for segmentation, and we propose 3 different models for relations, since the task makes it harder to generalize across all corpora.

In French, the placement of the adjective within a noun phrase is subject to variation: it can appear either before or after the noun. We conduct experiments to assess whether transformer-based language models are able to learn the adjective position in noun phrases in French –a position which depends on several linguistic factors. Prior findings have shown that transformer models are insensitive to permutated word order, but in this work, we show that finetuned models are successful at learning and selecting the correct position of the adjective. However, this success can be attributed to the process of finetuning rather than the linguistic knowledge acquired during pretraining, as evidenced by the low accuracy of experiments of classification that make use of pretrained embeddings. Comparing the finetuned models to the choices of native speakers (with a questionnaire), we notice that the models favor context and global syntactic roles, and are weaker with complex structures and fixed expressions.

Aspect is a linguistic concept that describes how an action, event, or state of a verb phrase is situated in time. In this paper, we explore whether different transformer models are capable of identifying aspectual features. We focus on two specific aspectual features: telicity and duration. Telicity marks whether the verb’s action or state has an endpoint or not (telic/atelic), and duration denotes whether a verb expresses an action (dynamic) or a state (stative). These features are integral to the interpretation of natural language, but also hard to annotate and identify with NLP methods. We perform experiments in English and French, and our results show that transformer models adequately capture information on telicity and duration in their vectors, even in their non-finetuned forms, but are somewhat biased with regard to verb tense and word order.

L’aspect du verbe décrit la manière dont une action, un événement ou un état exprimé par un verbe est lié au temps ; la télicité est la propriété d’un syntagme verbal qui présente une action ou un événement comme étant mené à son terme ; la durée distingue les verbes qui expriment une action (dynamique) ou un état (statique). Ces caractéristiques essentielles à l’interprétation du langage naturel, sont également difficiles à annoter et à identifier par les méthodes de TAL. Dans ce travail, nous estimons la capacité de différents modèles de type transformers pré-entraînés (BERT, RoBERTa, XLNet, ALBERT) à prédire la télicité et la durée. Nos résultats montrent que BERT est le plus performant sur les deux tâches, tandis que les modèles XLNet et ALBERT sont les plus faibles. Par ailleurs, les performances de la plupart des modèles sont améliorées lorsqu’on leur fournit en plus la position des verbes. Globalement, notre étude établit que les modèles de type transformers captent en grande partie la télicité et la durée.

Selectional preference is defined as the tendency of a predicate to favor particular arguments within a certain linguistic context, and likewise, reject others that result in conflicting or implausible meanings. The stellar success of contextual word embedding models such as BERT in NLP tasks has led many to question whether these models have learned linguistic information, but up till now, most research has focused on syntactic information. We investigate whether Bert contains information on the selectional preferences of words, by examining the probability it assigns to the dependent word given the presence of a head word in a sentence. We are using word pairs of head-dependent words in five different syntactic relations from the SP-10K corpus of selectional preference (Zhang et al., 2019b), in sentences from the ukWaC corpus, and we are calculating the correlation of the plausibility score (from SP-10K) and the model probabilities. Our results show that overall, there is no strong positive or negative correlation in any syntactic relation, but we do find that certain head words have a strong correlation and that masking all words but the head word yields the most positive correlations in most scenarios –which indicates that the semantics of the predicate is indeed an integral and influential factor for the selection of the argument.

Multilingual, inflectional corpora are a scarce resource in the NLP community, especially corpora with annotated morpheme boundaries. We are evaluating a generated, multilingual inflectional corpus with morpheme boundaries, generated from the English Wiktionary (Metheniti and Neumann, 2018), against the largest, multilingual, high-quality inflectional corpus of the UniMorph project (Kirov et al., 2018). We confirm that the generated Wikinflection corpus is not of such quality as UniMorph, but we were able to extract a significant amount of words from the intersection of the two corpora. Our Wikinflection corpus benefits from the morpheme segmentations of Wiktionary/Wikinflection and from the manually-evaluated morphological feature tags of the UniMorph project, and has 216K lemmas and 5.4M word forms, in a total of 68 languages.