TTI-COIN at SemEval-2017 Task 10: Investigating Embeddings for End-to-End Relation Extraction from Scientific Papers

Proceedings of the 11th International Workshop on Semantic Evaluations (SemEval-2017), pages 985–989, Vancouver, Canada, August 3 - 4, 2017. c⃝2017 Association for Computational Linguistics TTI-COIN at SemEval-2017 Task 10: Investigating Embeddings for End-to-End Relation Extraction from Scientific Papers Tomoki Tsujimura, Makoto Miwa, and Yutaka Sasaki COmputational INtelligence Laboratory Toyota Technological Institute 2-12-1 Hisakata, Tempaku-ku, Nagoya, Aichi, 468-8511, Japan {sd16420, makoto-miwa, yutaka.sasaki}@toyota-ti.ac.jp Abstract This paper describes our TTI-COIN sys- tem that participated in SemEval-2017 Task 10. We investigated appropriate em- beddings to adapt a neural end-to-end en- tity and relation extraction system LSTM- ER to this task. We participated in the full task setting of the entity segmentation, entity classification and relation classifica- tion (scenario 1) and the setting of relation classification only (scenario 3). The sys- tem was directly applied to the scenario 1 without modifying the codes thanks to its generality and flexibility. Our evaluation results show that the choice of appropriate pre-trained embeddings affected the per- formance significantly. With the best em- beddings, our system was ranked third in the scenario 1 with the micro F1 score of 0.38. We also confirm that our system can produce the micro F1 score of 0.48 for the scenario 3 on the test data, and this score is close to the score of the 3rd ranked system in the task. 1 Introduction Semantic relationships between entities are use- ful for building knowledge bases and semantic search engines. Their automatic extraction has been widely studied in the natural language pro- cessing (NLP) field (Li and Ji, 2014; Miwa and Sasaki, 2014; Miwa and Bansal, 2016). SemEval- 2017 Task 10 (Augenstein et al., 2017) deals with relation extraction from scientific papers. While entity detection and relation extraction have often been treated as separate tasks, several studies show that joint treatment of these tasks can improve extraction performance on both tasks (Li and Ji, 2014; Miwa and Sasaki, 2014). We em- ployed the state-of-the-art neural network-based end-to-end entity and relation extraction system LSTM-ER1 (Miwa and Bansal, 2016). The model of the system is built on pre-trained word em- beddings, and it has a tree-structured bidirectional long short-term memory-based recurrent neural network (LSTM-RNN) layer stacked on a sequen- tial bidirectional LSTM-RNN layer. It predicts en- tities and relations in an end-to-end manner with shared parameters, and the parameters in word embeddings and both LSTM-RNNs are updated simultaneously during training. We first checked the applicability of the sys- tem in our evaluation. The system was originally developed for ACE (automatic content extraction) corpora, but it does not depend on specific tasks and has high configurability (Miwa and Anani- adou, 2015) since it prepares a separate configu- ration file, where task specific settings like hyper- parameters can be specified. The system was suc- cessfully applied to the end-to-end relation extrac- tion setting (scenario 1 in the task) without mod- ifying the original codes in our experiments, al- though small modifications to the inputs were re- quired. This shows the generality of the system. Using this system, we …

Content truncated to 500 words