Active Learning for Interactive Relation Extraction in a French Newspaper’s Articles

Proceedings of Recent Advances in Natural Language Processing, pages 886–894 Sep 1–3, 2021. https://doi.org/10.26615/978-954-452-072-4_101 886 Active Learning for Interactive Relation Extraction in a French Newspaper’s Articles Cyrielle Mallart1, 2 Michel Le Nouy1 Guillaume Gravier3 Pascale Sébillot2 cyrielle.mallart@irisa.fr, michel.lenouy@ouest-france.fr, guig@irisa.fr, pascale.sebillot@irisa.fr (1) SIPA Ouest-France, 10 rue du Breil, 35000 Rennes, France (2) INSA Rennes, IRISA, Campus de Beaulieu, 35042 Rennes, France (3) CNRS, IRISA, Campus de Beaulieu, 35042 Rennes, France Abstract Relation extraction is a subtask of natural lan- guage processing that has seen many improve- ments in recent years, with the advent of com- plex pre-trained architectures. Many of these state-of-the-art approaches are tested against benchmarks with labelled sentences contain- ing tagged entities, and require important pre- training and fine-tuning on task-specific data. However, in a real use-case scenario such as in a newspaper company mostly dedicated to lo- cal information, relations are of varied, highly specific type, with virtually no annotated data for such relations, and many entities co-occur in a sentence without being related. We ques- tion the use of supervised state-of-the-art mod- els in such a context, where resources such as time, computing power and human annotators are limited. To adapt to these constraints, we experiment with an active-learning based rela- tion extraction pipeline, consisting of a binary LSTM-based lightweight model for detecting the relations that do exist, and a state-of-the-art model for relation classification. We compare several choices for classification models in this scenario, from basic word embedding averag- ing, to graph neural networks and Bert-based ones, as well as several active learning acqui- sition strategies, in order to find the most cost- efficient yet accurate approach in our French largest daily newspaper company’s use case. 1 Motivation Relation extraction is a mature field of natural lan- guage processing that aims at finding the relations between identified entities in texts. Recent research has focused on the use of trainable models to au- tomatically extract and classify relations between entities. The state-of-the-art research on relation ex- traction focuses on large, complex models, that re- quire either a long time to train or some pre-training with a fine-tuning phase. Task-specific labelled data is needed to train the final classification model, with research relying on benchmarks such as Zhang et al. (2017) or Hendrickx et al. (2010), which are often made of single sentences with clearly tagged entities and definitive, non-ambiguous relations. In a real-life scenario, however, relation extrac- tion is used in a language processing pipeline, e.g., in order to confront different reports of a same event, or to grasp a general picture of a situation. In the specific case of a regional newspaper com- pany such as ours, the intent is to create a knowl- edge graph from the content of the newspaper’s own articles, so as to facilitate journalists’ inves- tigation with an easy access to information and to possible relations that are otherwise drowned in pages of text. In this realistic scenario, data is of a very different nature from standard corpora, exhibiting certain features …

Content truncated to 500 words