11 мая 2022 г. в 11:00 в рамках семинара "Математические модели информационных технологий" департамента анализа данных и искусственного интеллекта и МЛ "Интеллектуальных систем и структурного анализа" под руководством С.О. Кузнецова состоится онлайн семинар по теме "Informative discourse feature selection for analysis of textual data".
Докладчик: Елизавета Гончарова, м.н.с., Международная лаборатория интеллектуальных систем и структурного анализа.
Аннотация: The presented research is dedicated to the analysis of the modern pre-trained language models (LMs) and their ability to inject linguistic features, such as discourse, during solving natural language processing tasks.Recent pre-trained LMs have shown state-of-the-art results on a bunch of NLP tasks, however, these models still suffer from the insufficient linguistic representation of a text that eventually leads to a low level of language understanding. In order to improve these models’ performance novel methods of discourse structure encoding have been proposed in the research. The introduced approaches allow us to incorporate discourse features into the LMs explicitly during pre-training or fine-tuning procedures without requiring significant modifications to the model's architecture. We provided the experimental evaluation of the discourse-aware models on various complex NLP tasks which are argumentation classification (AC), question answering (QA), and text summarization, and concluded that the modified models achieve results as good as or better than other discourse-free or more complex discourse-aware models on the well-known NLP benchmarks. Finally, the influence of discourse features on the models' explainability is considered. In the research, we introduced an independent explainability pipeline that is able to reveal relevant text spans based on the discourse relations assigned to them that can be used to explain deep learning models' decisions in supervised NLP tasks.
Источник: НИУ ВШЭ
