Complex Word Identification as a Sequence Labelling Task

Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 1148–1153 Florence, Italy, July 28 - August 2, 2019. c⃝2019 Association for Computational Linguistics 1148 Complex Word Identification as a Sequence Labelling Task Sian Gooding Dept of Computer Science and Technology University of Cambridge shg36@cam.ac.uk Ekaterina Kochmar ALTA Institute University of Cambridge ek358@cam.ac.uk Abstract Complex Word Identification (CWI) is con- cerned with detection of words in need of sim- plification and is a crucial first step in a sim- plification pipeline. It has been shown that re- liable CWI systems considerably improve text simplification. However, most CWI systems to date address the task on a word-by-word basis, not taking the context into account. In this pa- per, we present a novel approach to CWI based on sequence modelling. Our system is capa- ble of performing CWI in context, does not require extensive feature engineering and out- performs state-of-the-art systems on this task. 1 Introduction Lexical complexity is one of the main aspects con- tributing to overall text complexity (Dubay, 2004). It is typically addressed with lexical simplification (LS) systems that aim to paraphrase and substitute complex terms for simpler alternatives. Previous research has shown that Complex Word Identifi- cation (CWI) considerably improves lexical sim- plification (Shardlow, 2014; Paetzold and Specia, 2016a). This is achieved by identifying complex terms in text prior to word substitution. The per- formance of a CWI component is crucial, as low recall of this component might result in an overly difficult text with many missed complex words, while low precision might result in meaning dis- tortions with an LS system trying to unnecessarily simplify non-complex words (Shardlow, 2013). CWI has recently attracted attention as a stand- alone application, with at least two shared tasks focusing on it. Current approaches to CWI, in- cluding state-of-the-art systems, have a number of limitations. First of all, CWI systems typically ad- dress this task on a word-by-word basis, using a large number of features to capture the complexity of a word. For instance, the CWI system by Paet- zold and Specia (2016c) uses a total of 69 features, while the one by Gooding and Kochmar (2018) uses 27 features. Secondly, systems performing CWI in a static manner are unable to take the con- text into account, thus failing to predict word com- plexity for polysemous words as well as words in various metaphorical or novel contexts. For in- stance, consider the following two contexts of the word molar from the CWI 2018 shared task (Yi- mam et al., 2018). Molar has been annotated as complex in the first context (resulting in the bi- nary annotation of 1) by 17 out of 20 annotators (thus, the “probabilistic” label of 0.85), and as non-complex (label 0) in the second context: Contexts Bin Prob Elephants have four molars... 1 0.85 ... new molars emerge in the back of the mouth. 0 0.00 The annotators may have found the second con- text simpler on the whole, as molars is surrounded by …

Content truncated to 500 words