EMNLP@CPH: Is frequency all there is to simplicity?

First Joint Conference on Lexical and Computational Semantics (*SEM), pages 408–412, Montr´eal, Canada, June 7-8, 2012. c⃝2012 Association for Computational Linguistics EMNLP@CPH: Is frequency all there is to simplicity? Anders Johannsen, Héctor Martínez, Sigrid Klerke†, Anders Søgaard Centre for Language Technology University of Copenhagen {ajohannsen|alonso|soegaard}@hum.ku.dk sigridklerke@gmail.com† Abstract Our system breaks down the problem of rank- ing a list of lexical substitutions according to how simple they are in a given context into a series of pairwise comparisons between can- didates. For this we learn a binary classifier. As only very little training data is provided, we describe a procedure for generating artifi- cial unlabeled data from Wordnet and a corpus and approach the classification task as a semi- supervised machine learning problem. We use a co-training procedure that lets each classi- fier increase the other classifier’s training set with selected instances from an unlabeled data set. Our features include n-gram probabilities of candidate and context in a web corpus, dis- tributional differences of candidate in a cor- pus of “easy” sentences and a corpus of normal sentences, syntactic complexity of documents that are similar to the given context, candidate length, and letter-wise recognizability of can- didate as measured by a trigram character lan- guage model. 1 Introduction This paper describes a system for the SemEval 2012 English Lexical Simplification shared task. The task description uses a loose definition of simplic- ity, defining “simple words” as “words that can be understood by a wide variety of people, including for example people with low literacy levels or some cog- nitive disability, children, and non-native speakers of English” (Specia et al., 2012). Feature r Nsf 0.33 Nsf+1 0.27 Nsf−1 0.27 Lsf -0.26 Lmax -0.26 RIproto(l) -0.18 Scn -0.17 Sw -0.17 Scp -0.17 Feature r RIproto(f) -0.15 Cmax -0.14 RIorig(l) -0.11 Ltokens -0.10 Cmin 0.10 SWfreq 0.08 SWLLR 0.07 Cavg -0.04 Table 1: Pearson’s r correlations. The table shows the three highest correlated features per group, all of which are significant at the p < 0.01 level 2 Features We model simplicity with a range of features divided into six groups. Five of these groups make use of the distributional hypothesis and rely on external cor- pora. We measure a candidate’s distribution in terms of its lexical associations (RI), participation in syn- tactic structures (S), or corpus presence in order to assess its simplicity (N, SW, C). A single group, L, measures intrinsic aspects of the substi- tution candidate, such as its length. The substitution candidate is either an adjective, an adverb, a noun, or a verb, and all candidates within a list share the same part of speech. Because word class might influence simplicity, we allow our model to fit parameters specific to the candidate’s part of speech by making a copy of the features for each part of speech which is active only when the candidate is in the given part of speech. 408 Simple Wikipedia (SW) These two features con- tain relative frequency counts of the substitution form in Simple English …

Content truncated to 500 words