LightRel at SemEval-2018 Task 7: Lightweight and Fast Relation Classification

Proceedings of the 12th International Workshop on Semantic Evaluation (SemEval-2018), pages 778–782 New Orleans, Louisiana, June 5–6, 2018. ©2018 Association for Computational Linguistics LIGHTREL at SemEval-2018 Task 7: Lightweight and Fast Relation Classification Tyler Renslow DFKI, Saarbr¨ucken, Germany tdrenslow@gmail.com G¨unter Neumann DFKI, Saarbr¨ucken, Germany neumann@dfki.de Abstract We present LIGHTREL, a lightweight and fast relation classifier. Our goal is to develop a high baseline for different relation extraction tasks. By defining only very few data-internal, word-level features and external knowledge sources in the form of word clusters and word embeddings, we train a fast and simple linear classifier. 1 Introduction The main motivation for our participation at SemEval-2018 (G´abor et al., 2018) was the ideal opportunity to test and improve our relation ex- traction system LIGHTREL. The system design and development was inspired by the work de- scribed in (Nguyen and Grishman, 2015). Their goal was to depart from traditional relation ex- traction approaches with complicated feature en- gineering by exploring a deep neural network that would minimize its dependence on external toolk- its and resources, e.g. external word embeddings. That allowed them to design a rather lightweight relation extraction approach that would basically only require supervised training data, external word embeddings and a few hyperparameters. Their ”end-to-end” relation extraction approach produced competitive results. Since tuning hyperparameters for neural net- works can be a intricate process, we considered whether it would be possible to define an even sim- pler system and use it as a baseline for our future research. Thus, we adopted some of the design decisions made by (Nguyen and Grishman, 2015) and combined them with a well-known, fast linear classifier, viz. LibLinear (Fan et al., 2008). Following Nguyen and Grishman (2015), we represent a relation mention as a sequence of to- kens. The core idea of our approach consists of transforming this sequence into a a vector of fixed length, such that each token (or word) is repre- sented only by: 1) the word itself, 2) its shape (a small, fixed amount of character-based features), 3) the word’s cluster id, and 4) the word’s embed- ding of fixed size. For this competition, we introduce a new relation-level feature, namely the ID of the word directly following and preceding entities one and two, respectively. Furthermore, we ignore all to- kens to the left of the first entity and to the right of second entity. The size of the whole vector there- fore hinges on the maximum number of elements between the two entities found in the training set. These representations are then used to train a LibLinear model. Note that this reduces man- ual feature engineering to defining the shape fea- tures, finding an appropriate number of clusters and word embedding dimensions, and hyperpa- rameters for LibLinear. All other information is automatically computed from the training data. In this sense, we consider our system lightweight. We initially developed and tested our approach on the previous and widely-used SemEval-2010 Task 8 data set (Hendrickx et al., 2010), and obtained as …

Content truncated to 500 words