Multilevel Heuristics for Rationale-Based Entity Relation Classification in Sentences

Proceedings of the 27th International Conference on Computational Linguistics, pages 1145–1155 Santa Fe, New Mexico, USA, August 20-26, 2018. 1145 Multilevel Heuristics for Rationale-Based Entity Relation Classification in Sentences Shiou Tian Hsu, Mandar Chaudhary and Nagiza F. Samatova North Carolina State University shsu3@ncsu.edu, mschaudh@ncsu.edu, samatova@csc.ncsu.edu Abstract Rationale-based models provide a unique way to provide justifiable results for relation classifi- cation models by identifying rationales (key words and phrases that a person can use to justify the relation in the sentence) during the process. However, existing generative networks used to extract rationales come with a trade-off between extracting diversified rationales and achieving good classification results. In this paper, we propose a multilevel heuristic approach to regulate rationale extraction to avoid extracting monotonous rationales without compromising classifica- tion performance. In our model, rationale selection is regularized by a semi-supervised process and features from different levels: word, syntax, sentence, and corpus. We evaluate our approach on the SemEval 2010 dataset that includes 19 relation classes and the quality of extracted ra- tionales with our manually-labeled rationales. Experiments show a significant improvement in classification performance and a 20% gain in rationale interpretability compared to state-of-the- art approaches. 1 Introduction The goal of sentence-level entity relation classification is to infer how two target entities are semantically associated in a sentence. It is a core NLP function that supports many high level tasks such as information extraction and knowledge graph population (Hendrickx et al., 2009; Niu et al., 2012). Recent advances in generative models facilitate improved classification performance along with justifiable results which improves model interpretability. A generative model will first extract the most representative fragments in the sentence that expresses the relation first, and augments the classifier with the rationales (Hsu et al., 2018) . These representative fragments are called rationales as per (Zhang et al., 2016; Lei et al., 2016; Hsu et al., 2018), and can be used to justify the results. We illustrate an example of rationale-based models using an example sentence which expresses a Instrument-Agency(e2,e1) relationship between the given target entities e1 = “attacker” and e2 = “instrument”: Rationale based models: She had struggled violently with her [attacker]e1, who killed her with a blunt [instrument]e2. (words marked in bold and the entities are the major sources to infer the relation) The most commonly used structure in generative models consists of two components: a Generator and a Discriminator. The Generator extracts rationales from an input sentence in an unsupervised fashion; the Discriminator, which is supervised, predicts the relation class utilizing the rationales. However, training a good Generator can be essentially hard due to the mode collapse problem (Chen et al., 2016; Che et al., 2017; Duhyeon and Hyunjung, 2018). Mode collapse is a commonly observed paradox where the Generator attempts to extract varied rationales, but converges to select monotonous rationales because the Discriminator failed to model diversified rationales. A collapsed Generator is not capable of providing This work is licensed under a Creative Commons Attribution 4.0 International License. License details: http:// creativecommons.org/licenses/by/4.0/ 1146 …

Content truncated to 500 words