Analysis and control of online interactions through neural natural language processing. (Analyse et contrôle des interactions en ligne avec des réseaux neuronaux artificiels de traitement automatique du langage naturel)

Leo Laugier

Abstract: Natural Language Processing is motivated by applications where computers should gain a semantic and syntactic understanding of human language. Recently, the field has been impacted by a paradigm shift. Deep learning architectures coupled with self-supervised training have become the core of state-of-the-art models used in Natural Language Understanding and Natural Language Generation. Sometimes considered as foundation models, these systems pave the way for novel use cases. Driven by an academic-industrial partnership between the Institut Polytechnique de Paris and Google Ai Research, the present research has focused on investigating how pretrained neural Natural Language Processing models could be leveraged to improve online interactions.This thesis first explored how self-supervised style transfer could be applied to the toxic-to-civil rephrasing of offensive comments found in online conversations. In the context of toxic content moderation online, we proposed to fine-tune a pretrained text-to-text model (T5) with a denoising and cyclic auto-encoder loss. The system, called CAE-T5, was trained on the largest toxicity detection dataset to date (Civil Comments) and generates sentences that are more fluent and better at preserving the initial content compared to earlier text style transfer systems, according to several scoring systems and human evaluation. Plus the approach showed it could be generalized to additional style transfer tasks, such as sentiment transfer.Then, a subsequent work investigated the human labeling and automatic detection of toxic spans in online conversations. Contrary to toxicity detection datasets and models which classify whole posts as toxic or not, toxic spans detection aims at highlighting toxic spans, that is to say the spans that make a text toxic, when detecting such spans is possible. We released a new labeled dataset to train and evaluate systems, which led to a shared task at the 15th International Workshop on Semantic Evaluation. Systems proposed to address the task include strongly supervised models trained using annotations at the span level as well as weakly supervised approaches, known as rationale extraction, using classifiers trained on potentially larger external datasets of posts manually annotated as toxic or not, without toxic span annotations. Furthermore, the ToxicSpans dataset and systems proved useful to analyze the performances of humans and automatic systems on toxic-to-civil rephrasing.Finally, we developed a recommender system based on online reviews of items, taking part in the topic of explaining users' tastes considered by the predicted recommendations. The method uses textual semantic similarity models to represent a user's preferences as a graph of textual snippets, where the edges are defined by semantic similarity. This textual, memory-based approach to rating prediction holds out the possibility of improved explanations for recommendations. The method is evaluated quantitatively, highlighting that leveraging text in this way can outperform both memory-based and model-based collaborative filtering baselines.