UTU at SemEval-2016 Task 10: Binary Classification for Expression Detection (BCED)

Proceedings of SemEval-2016, pages 925–930, San Diego, California, June 16-17, 2016. c⃝2016 Association for Computational Linguistics UTU at SemEval-2016 Task 10: Binary Classification for Expression Detection (BCED) Jari Bj¨orne and Tapio Salakoski Department of Information Technology, University of Turku Turku Centre for Computer Science (TUCS) Faculty of Mathematics and Natural Sciences, FI-20014, Turku, Finland firstname.lastname@utu.fi Abstract The SemEval 2016 DiMSUM Shared Task concerns the detection of minimal semantic units from text and prediction of their coarse lexical categories known as supersenses. Our approach is to define this task as a binary clas- sification problem approachable by straight- forward machine learning methods. We start by detecting semantic units by match- ing text spans against several large dictionar- ies, including the English WordNet, expres- sions derived from the Yelp Academic Dataset and concepts from the English Wikipedia, generating a set of potential supersenses for each matched span. For each potential su- persense and text span pair a binary machine learning example is defined. We classify these examples using an ensemble method, taking as the final predicted supersense the one with the highest confidence score. Our system achieves good performance on the supersense classification task but has limited performance for detection of multi-word se- mantic units. We show that the task of super- sense prediction can be effectively defined as a binary classification task. 1 Introduction The SemEval 2016 DiMSUM Shared Task1 con- cerns the detection of minimal semantic units and their semantic classification using supersenses (Schneider et al., 2016). The minimal semantic units can consist of either single words or multiword expressions (MWE) in cases where multiple words 1http://dimsum16.github.io/ form a lexical item. The concept of MWE as used in the DiMSUM task is described in detail in Schneider et al. (2014). The supersenses are broad-coverage, coarse lexical categories, 26 for nouns and 15 for verbs. These supersenses mostly correspond to the WordNet lexicographer files. A detailed description of the supersenses is in Schneider and Smith (2015). Automated supersense tagging has previously been explored by e.g. Curran (2005), Ciaramita and Altun (2006) and Johannsen et al. (2014). In the DiMSUM task systems must both detect the minimal semantic units and assign correct super- senses for them. The final metric of the shared task is the averaged performance on these tasks, which are also evaluated independently. The systems can approach these tasks either independently (e.g. in a pipeline) or through a joint model. We approach the DiMSUM task as a machine learning classification task where examples are gen- erated for all semantic units found in a pre-defined dictionary. These semantic units are classified into one of the 41 supersenses by generating an example for each possible semantic unit + supersense pair, turning the task of assigning one of many super- senses into a binary classification problem. In this manner, the task becomes a generalized machine learning task for which almost any classifier can be used. The DiMSUM shared task defines three subtasks or data conditions in which the systems can partic- ipate, varying …

Content truncated to 500 words