Distant Supervision for Sentiment Attitude Extraction

Proceedings of Recent Advances in Natural Language Processing, pages 1022–1030, Varna, Bulgaria, Sep 2–4, 2019. https://doi.org/10.26615/978-954-452-056-4_118 1022 Distant Supervision for Sentiment Attitude Extraction Nicolay Rusnachenko1, Natalia Loukachevitch1,2, Elena Tutubalina3,4 1 Bauman Moscow State Technical University, Moscow, Russia 2 Lomonosov Moscow State University, Moscow, Russia 3 Kazan Federal University, Kazan, Russia 4 Samsung-PDMI Joint AI Center, PDMI RAS, St. Petersburg, Russia kolyarus@yandex.ru, louk_nat@mail.ru, elvtutubalina@kpfu.ru Abstract News articles often convey attitudes be- tween the mentioned subjects, which is es- sential for understanding the described sit- uation. In this paper, we describe a new approach to distant supervision for extract- ing sentiment attitudes between named en- tities mentioned in texts. Two factors (pair-based and frame-based) were used to automatically label an extensive news col- lection, dubbed as RuAttitudes. The latter became a basis for adaptation and train- ing convolutional architectures, including piecewise max pooling and full use of in- formation across different sentences. The results show that models, trained with RuAttitudes, outperform ones that were trained with only supervised learning ap- proach and achieve 13.4% increase in F1- score on RuSentRel collection.1 1 Introduction Relation extraction nowadays remains one of the popular tasks in the natural language processing domain. The relation types to be extracted from texts may vary and result in different tasks: se- mantic classification of relations between a pair of common nominals (Hendrickx et al., 2009), source-target sentiment relation extraction (Ellis et al., 2014), opinion expression towards entities and events (Deng and Wiebe, 2015a), attitude ex- traction between mentioned named entities (Rus- nachenko and Loukachevitch, 2018), etc. Dealing with one of these tasks, the greatest dif- ficulty one encounters is the complexity of the sen- tence structure. As for analytical articles, the idea expressed by the author could be conveyed in dif- ferent variants, which is a feature of natural lan- guages. 1The code is available on https://github.com/ nicolay-r/attitudes-extraction-ds When relation extraction is performed automat- ically using machine learning approaches, this complexity results in a lack of training exam- ples. One technique that helps to accomplish this task is distant supervision (DS), initially proposed by (Mintz et al., 2009). It assumes to extract and label data by relying on assumptions based on a prepared knowledge base. Although many meth- ods have been proposed in such domains as sen- timent analysis and relation extraction (Turney, 2002; Zeng et al., 2015), the domain of sentiment attitude extraction remains understudied. This paper describes a new approach to distant supervision for extracting sentiment attitudes be- tween named entities mentioned in texts. It is worth noting that DS faces the problem of wrong labels, which becomes a reason of noisy labeled data. To address the shortcomings of noisy label- ing, in this paper we exploit two primary sources of automatic annotation: • Prior knowledge about current attitudes be- tween political entities (figures); • Sentiment frames that define attitudes be- tween participants of a situation. The obtained corpus annotated with attitudes was used to train convolutional neural networks (CNNs), adapted for relation extraction and full use …

Content truncated to 500 words