This paper presents DISRPT, a multilingual, multi-domain, and cross-framework benchmark dataset for discourse processing, covering the tasks of discourse unit segmentation, connective identification, and relation classification. DISRPT includes 13 languages, with data from 24 corpora covering about 4 millions tokens and around 250,000 discourse relation instances from 4 discourse frameworks: RST, SDRT, PDTB, and Discourse Dependencies. We present an overview of the data, its development across three NLP shared tasks on discourse processing carried out in the past five years, and the latest modifications and added extensions. We also carry out an evaluation of state-of-the-art multilingual systems trained on the data for each task, showing plateau performance on segmentation, but important room for improvement for connective identification and relation classification. The DISRPT benchmark employs a unified format that we make available on GitHub and HuggingFace in order to encourage future work on discourse processing across languages, domains, and frameworks.

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.

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.

Question Generation (QG), the process of generating meaningful questions from a given context, has proven to be useful for several tasks such as question answering or FAQ generation. While most existing QG techniques generate simple, fact-based questions, this research aims to generate questions that can have complex answers (e.g. “why” questions). We propose a data augmentation method that uses discourse relations to create such questions, and experiment on existing English data. Our approach generates questions based solely on the context without answer supervision, in order to enhance question diversity and complexity. We use an encoder-decoder trained on the augmented dataset to generate either one question or multiple questions at a time, and show that the latter improves over the baseline model when doing a human quality evaluation, without degrading performance according to standard automated metrics.

Discourse analysis plays a crucial role in Natural Language Processing, with discourse relation prediction arguably being the most difficult task in discourse parsing. Previous studies have generally focused on explicit or implicit discourse relation classification in monologues, leaving dialogue an under-explored domain. Facing the data scarcity issue, we propose to leverage self-training strategies based on a Transformer backbone. Moreover, we design the first semi-supervised pipeline that sequentially predicts discourse structures and relations. Using 50 examples, our relation prediction module achieves 58.4 in accuracy on the STAC corpus, close to supervised state-of-the-art. Full parsing results show notable improvements compared to the supervised models both in-domain (gaming) and cross-domain (technical chat), with better stability.

Discourse processing suffers from data sparsity, especially for dialogues. As a result, we explore approaches to infer latent discourse structures for dialogues, based on attention matrices from Pre-trained Language Models (PLMs). We investigate multiple auxiliary tasks for fine-tuning and show that the dialogue-tailored Sentence Ordering task performs best. To locate and exploit discourse information in PLMs, we propose an unsupervised and a semi-supervised method. Our proposals thereby achieve encouraging results on the STAC corpus, with F1 scores of 57.2 and 59.3 for the unsupervised and semi-supervised methods, respectively. When restricted to projective trees, our scores improved to 63.3 and 68.1.

One crucial aspect of democracy is fair information sharing. While it is hard to prevent biases in news, they should be identified for better transparency. We propose an approach to automatically characterize biases that takes into account structural differences and that is efficient for long texts. This yields new ways to provide explanations for a textual classifier, going beyond mere lexical cues. We show that: (i) the use of discourse-based structure-aware document representations compare well to local, computationally heavy, or domain-specific models on classification tasks that deal with textual bias (ii) our approach based on different levels of granularity allows for the generation of better explanations of model decisions, both at the lexical and structural level, while addressing the challenge posed by long texts.

This paper describes our approach to Subtask 1 “News Genre Categorization” of SemEval-2023 Task 3 “Detecting the Category, the Framing, and the Persuasion Techniques in Online News in a Multi-lingual Setup”, which aims to determine whether a given news article is an opinion piece, an objective report, or satirical. We fine-tuned the domain-specific language model POLITICS, which was pre-trained on a large-scale dataset of more than 3.6M English political news articles following ideology-driven pre-training objectives. In order to use it in the multilingual setup of the task, we added as a pre-processing step the translation of all documents into English. Our system ranked among the top systems overall in most language, and ranked 1st on the English dataset.

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.

With the growing number of information sources, the problem of media bias becomes worrying for a democratic society. This paper explores the task of predicting the political orientation of news articles, with a goal of analyzing how bias is expressed. We demonstrate that integrating rhetorical dimensions via latent structures over sub-sentential discourse units allows for large improvements, with a +7.4 points difference between the base LSTM model and its discourse-based version, and +3 points improvement over the previous BERT-based state-of-the-art model. We also argue that this gives a new relevant handle for analyzing political bias in news articles.

Depression is a serious mental illness that impacts the way people communicate, especially through their emotions, and, allegedly, the way they interact with others. This work examines depression signals in dialogs, a less studied setting that suffers from data sparsity. We hypothesize that depression and emotion can inform each other, and we propose to explore the influence of dialog structure through topic and dialog act prediction. We investigate a Multi-Task Learning (MTL) approach, where all tasks mentioned above are learned jointly with dialog-tailored hierarchical modeling. We experiment on the DAIC and DailyDialog corpora – both contain dialogs in English – and show important improvements over state-of-the-art on depression detection (at best 70.6% F1), which demonstrates the correlation of depression with emotion and dialog organization and the power of MTL to leverage information from different sources.

In 2021, we organized the second iteration of a shared task dedicated to the underlying units used in discourse parsing across formalisms: the DISRPT Shared Task (Discourse Relation Parsing and Treebanking). Adding to the 2019 tasks on Elementary Discourse Unit Segmentation and Connective Detection, this iteration of the Shared Task included for the first time a track on discourse relation classification across three formalisms: RST, SDRT, and PDTB. In this paper we review the data included in the Shared Task, which covers nearly 3 million manually annotated tokens from 16 datasets in 11 languages, survey and compare submitted systems and report on system performance on each task for both annotated and plain-tokenized versions of the data.

We present an approach for discourse segmentation and discourse connective identification, both at the sentence and document level, within the Disrpt 2021 shared task, a multi-lingual and multi-formalism evaluation campaign. Building on the most successful architecture from the 2019 similar shared task, we leverage datasets in the same or similar languages to augment training data and improve on the best systems from the previous campaign on 3 out of 4 subtasks, with a mean improvement on all 16 datasets of 0.85%. Within the Disrpt 21 campaign the system ranks 3rd overall, very close to the 2nd system, but with a significant gap with respect to the best system, which uses a rich set of additional features. The system is nonetheless the best on languages that benefited from crosslingual training on sentence internal segmentation (German and Spanish).

We investigate linguistic markers associated with schizophrenia in clinical conversations by detecting predictive features among French-speaking patients. Dealing with human-human dialogues makes for a realistic situation, but it calls for strategies to represent the context and face data sparsity. We compare different approaches for data representation – from individual speech turns to entire conversations –, and data modeling, using lexical, morphological, syntactic, and discourse features, dimensions presumed to be tightly connected to the language of schizophrenia. Previous English models were mostly lexical and reached high performance, here replicated (93.7% acc.). However, our analysis reveals that these models are heavily biased, which probably concerns most datasets on this task. Our new delexicalized models are more general and robust, with the best accuracy score at 77.9%.

Various machine learning tasks can benefit from access to external information of different modalities, such as text and images. Recent work has focused on learning architectures with large memories capable of storing this knowledge. We propose augmenting generative Transformer neural networks with KNN-based Information Fetching (KIF) modules. Each KIF module learns a read operation to access fixed external knowledge. We apply these modules to generative dialog modeling, a challenging task where information must be flexibly retrieved and incorporated to maintain the topic and flow of conversation. We demonstrate the effectiveness of our approach by identifying relevant knowledge required for knowledgeable but engaging dialog from Wikipedia, images, and human-written dialog utterances, and show that leveraging this retrieved information improves model performance, measured by automatic and human evaluation.

Nous présentons des expériences visant à identifier automatiquement des patients présentant des symptômes de schizophrénie dans des conversations contrôlées entre patients et psychothérapeutes. Nous fusionnons l’ensemble des tours de parole de chaque interlocuteur et entraînons des modèles de classification utilisant des informations lexicales, morphologiques et syntaxiques. Cette étude est la première du genre sur le français et obtient des résultats comparables à celles sur l’anglais. Nos premières expériences tendent à montrer que la parole des personnes avec schizophrénie se distingue de celle des témoins : le meilleur modèle obtient une exactitude de 93,66%. Des informations plus riches seront cependant nécessaires pour parvenir à un modèle robuste.

Development of discourse parsers to annotate the relational discourse structure of a text is crucial for many downstream tasks. However, most of the existing work focuses on English, assuming a quite large dataset. Discourse data have been annotated for Basque, but training a system on these data is challenging since the corpus is very small. In this paper, we create the first demonstrator based on RST for Basque, and we investigate the use of data in another language to improve the performance of a Basque discourse parser. More precisely, we build a monolingual system using the small set of data available and investigate the use of multilingual word embeddings to train a system for Basque using data annotated for another language. We found that our approach to building a system limited to the small set of data available for Basque allowed us to get an improvement over previous approaches making use of many data annotated in other languages. At best, we get 34.78 in F1 for the full discourse structure. More data annotation is necessary in order to improve the results obtained with these techniques. We also describe which relations match with the gold standard, in order to understand these results.

Segmentation is the first step in building practical discourse parsers, and is often neglected in discourse parsing studies. The goal is to identify the minimal spans of text to be linked by discourse relations, or to isolate explicit marking of discourse relations. Existing systems on English report F1 scores as high as 95%, but they generally assume gold sentence boundaries and are restricted to English newswire texts annotated within the RST framework. This article presents a generic approach and a system, ToNy, a discourse segmenter developed for the DisRPT shared task where multiple discourse representation schemes, languages and domains are represented. In our experiments, we found that a straightforward sequence prediction architecture with pretrained contextual embeddings is sufficient to reach performance levels comparable to existing systems, when separately trained on each corpus. We report performance between 81% and 96% in F1 score. We also observed that discourse segmentation models only display a moderate generalization capability, even within the same language and discourse representation scheme.

In this paper, we investigate similarities between discourse and argumentation structures by aligning subtrees in a corpus containing both annotations. Contrary to previous works, we focus on comparing sub-structures and not only relations matches. Using data mining techniques, we show that discourse and argumentation most often align well, and the double annotation allows to derive a mapping between structures. Moreover, this approach enables the study of similarities between discourse structures and differences in their expressive power.

Discourse relation classification has proven to be a hard task, with rather low performance on several corpora that notably differ on the relation set they use. We propose to decompose the task into smaller, mostly binary tasks corresponding to various primitive concepts encoded into the discourse relation definitions. More precisely, we translate the discourse relations into a set of values for attributes based on distinctions used in the mappings between discourse frameworks proposed by Sanders et al. (2018). This arguably allows for a more robust representation of discourse relations, and enables us to address usually ignored aspects of discourse relation prediction, namely multiple labels and underspecified annotations. We show experimentally which of the conceptual primitives are harder to learn from the Penn Discourse Treebank English corpus, and propose a correspondence to predict the original labels, with preliminary empirical comparisons with a direct model.

Query-based open-domain NLP tasks require information synthesis from long and diverse web results. Current approaches extractively select portions of web text as input to Sequence-to-Sequence models using methods such as TF-IDF ranking. We propose constructing a local graph structured knowledge base for each query, which compresses the web search information and reduces redundancy. We show that by linearizing the graph into a structured input sequence, models can encode the graph representations within a standard Sequence-to-Sequence setting. For two generative tasks with very long text input, long-form question answering and multi-document summarization, feeding graph representations as input can achieve better performance than using retrieved text portions.

This work aims to contribute to our understanding of when multi-task learning through parameter sharing in deep neural networks leads to improvements over single-task learning. We focus on the setting of learning from loosely related tasks, for which no theoretical guarantees exist. We therefore approach the question empirically, studying which properties of datasets and single-task learning characteristics correlate with improvements from multi-task learning. We are the first to study this in a text classification setting and across more than 500 different task pairs.

The problem of detecting scientific fraud using machine learning was recently introduced, with initial, positive results from a model taking into account various general indicators. The results seem to suggest that writing style is predictive of scientific fraud. We revisit these initial experiments, and show that the leave-one-out testing procedure they used likely leads to a slight over-estimate of the predictability, but also that simple models can outperform their proposed model by some margin. We go on to explore more abstract linguistic features, such as linguistic complexity and discourse structure, only to obtain negative results. Upon analyzing our models, we do see some interesting patterns, though: Scientific fraud, for examples, contains less comparison, as well as different types of hedging and ways of presenting logical reasoning.

Discourse parsing is an integral part of understanding information flow and argumentative structure in documents. Most previous research has focused on inducing and evaluating models from the English RST Discourse Treebank. However, discourse treebanks for other languages exist, including Spanish, German, Basque, Dutch and Brazilian Portuguese. The treebanks share the same underlying linguistic theory, but differ slightly in the way documents are annotated. In this paper, we present (a) a new discourse parser which is simpler, yet competitive (significantly better on 2/3 metrics) to state of the art for English, (b) a harmonization of discourse treebanks across languages, enabling us to present (c) what to the best of our knowledge are the first experiments on cross-lingual discourse parsing.

Discourse segmentation is a crucial step in building end-to-end discourse parsers. However, discourse segmenters only exist for a few languages and domains. Typically they only detect intra-sentential segment boundaries, assuming gold standard sentence and token segmentation, and relying on high-quality syntactic parses and rich heuristics that are not generally available across languages and domains. In this paper, we propose statistical discourse segmenters for five languages and three domains that do not rely on gold pre-annotations. We also consider the problem of learning discourse segmenters when no labeled data is available for a language. Our fully supervised system obtains 89.5% F1 for English newswire, with slight drops in performance on other domains, and we report supervised and unsupervised (cross-lingual) results for five languages in total.

Discourse segmentation is the first step in building discourse parsers. Most work on discourse segmentation does not scale to real-world discourse parsing across languages, for two reasons: (i) models rely on constituent trees, and (ii) experiments have relied on gold standard identification of sentence and token boundaries. We therefore investigate to what extent constituents can be replaced with universal dependencies, or left out completely, as well as how state-of-the-art segmenters fare in the absence of sentence boundaries. Our results show that dependency information is less useful than expected, but we provide a fully scalable, robust model that only relies on part-of-speech information, and show that it performs well across languages in the absence of any gold-standard annotation.

We experiment with different ways of training LSTM networks to predict RST discourse trees. The main challenge for RST discourse parsing is the limited amounts of training data. We combat this by regularizing our models using task supervision from related tasks as well as alternative views on discourse structures. We show that a simple LSTM sequential discourse parser takes advantage of this multi-view and multi-task framework with 12-15% error reductions over our baseline (depending on the metric) and results that rival more complex state-of-the-art parsers.