The UWNLP system at SemEval-2018 Task 7: Neural Relation Extraction Model with Selectively Incorporated Concept Embeddings

Proceedings of the 12th International Workshop on Semantic Evaluation (SemEval-2018), pages 788–792 New Orleans, Louisiana, June 5–6, 2018. ©2018 Association for Computational Linguistics The UWNLP system at SemEval-2018 Task 7: Neural Relation Extraction Model with Selectively Incorporated Concept Embeddings Yi Luan Mari Ostendorf Hannaneh Hajishirzi University of Washington {luanyi, ostendor, hannaneh}@uw.edu Abstract This paper describes our submission for the SemEval 2018 Task 7 shared task on seman- tic relation extraction and classification in sci- entific papers. We extend the end-to-end re- lation extraction model of (Miwa and Bansal, 2016) with enhancements such as a character- level encoding attention mechanism on select- ing pretrained concept candidate embeddings. Our official submission ranked the second in relation classification task (Subtask 1.1 and Subtask 2 Senerio 2), and the first in the re- lation extraction task (Subtask 2 Scenario 1). 1 Task Overview The SemEval 2018 Task 7 Shared Task (G´abor et al., 2018) focuses on the task of recognizing the semantic relation that holds between scientific con- cepts. The task involves semantic relation extrac- tion and classification into six categories specific to scientific literature: USAGE, RESULT, MODEL- FEATURE, PART WHOLE, TOPIC, COMPARE. Two types of tasks are proposed: 1) identifying pairs of entities that are instances of any of the six seman- tic relations (extraction task), and 2) classifying instances into one of the specific relation types (classification task). Consider the following input sentence: “[Unsu- pervised training] is first used to train a [phone n-gram model] for a particular domain.” Given the concept pair [Unsupervised training] and [phone n-gram model], the relation extraction task is to identify whether there is a relation between the concepts, while the the relation classification task is to identity the relation as USAGE. Relation direc- tionality is not taken into account for the evaluation of the extraction task. Directionality is taken into account when relevant for the classification task (5 out of the 6 semantic relations are asymmetrical). We will use this example throughout the paper to illustrate various parts of our system. The SemEval 2018 Task 7 dataset contains 350 abstracts from the ACL Anthology for training and validation, and 150 abstracts for testing each sub- task. Since the scale of the data is small for su- pervised training of neural systems, we introduce several strategies to leverage a large quantity of un- labeled scientific articles. In addition to initializing a neural system with pre-trained word embeddings, as in (Luan et al., 2017), we also try to incorporate embeddings of concepts that span multiple words. In neural models such as (Miwa and Bansal, 2016), phrases are often represented by an average (or weighted average) of the token’s sequential LSTM representation. The intuition behind explicit mod- eling of multi-word concept embeddings is that the concept use may be different from that of its indi- vidual words. Due to the size of the dataset and the nature of scientific literature, a large number of the scientific terms in the test set have never appeared in the training set, so …

Content truncated to 500 words