Lexical Substitution is Practical for Rare Word Simplification

Lexical Substitution is Practical for Rare Word Simplification Takumi Maruyama Nagaoka University of Technology maruyama@jnlp.org Kazuhide Yamamoto Nagaoka University of Technology yamamoto@jnlp.org Abstract We attempted to generate a sentence by using the concept of core vocabulary whilst preserv- ing the original meaning of the sentence in the text simplification subtask. The correlation between word simplicity and frequency shows that many complex words are less frequently used. Hence, accurately simplifying sentences containing rare words is important in the sim- plification task. We explored a simplification model that works robustly. The machine trans- lation approach and the lexical substitution were evaluated in the dataset that includes rare words (referred to as RARE) and the dataset that does not include rare words (referred to as NORMAL). The machine translation ap- proach exhibits the best performance in the NORMAL dataset, but it exhibits very poor performance in the RARE dataset. Mean- while, the lexical substitution model works ro- bustly in both the datasets, and it displays flu- ent as well as adequate outputs. These results imply that for practical text simplification sys- tems, accumulating human knowledge is im- portant even though its construction is costly. 1 Introduction The number of foreigners who are visiting Japan has been increasing over the years. Japan hosts ap- proximately 28 million visitors in a year1. In addi- tion, approximately 2.32 million foreigners are liv- ing in the country2, and this number is increasing by 1https://www.jnto.go.jp/jpn/statistics/ visitor_trends 2https://www.e-stat.go.jp time. According to a survey conducted by the Na- tional Institute for Japanese Language and Linguis- tics, the number of people who can understand basic Japanese exceeds the number of people who can un- derstand English (Iwata, 2010). Hence, a simplified text is one of the important methods of providing in- formation to foreigners. The level of simplicity of a certain text is de- termined by various factors, such as the sentence length, the vocabulary used and the discourse struc- ture. Thus, objectively defining the level of text sim- plicity is difficult. Therefore, we focus on the extent of the vocabulary used in a text. According to the studies emphasising on the requirement of large vo- cabulary to understand a text, Laufer (1992) and Na- tion (2006) showed that reading comprehension is correlated with the ration of vocabulary (referred to as text covering ratio) known by a reader with regard to a certain text. That is, if a text is represented only by a certain known vocabulary, then understanding that text becomes easy. We challenge to simplify the original sentences with a vocabulary of 2,000 words (referred to as the core vocabulary) whilst preserv- ing its original meaning. In practical text simplification systems, simplify- ing rare words is very important. The correlation between word simplicity and frequency shows that many complex words are used less frequently. How- ever, no knowledge is available on the most robust model that works for rare words. Therefore, to de- termine a robust system in this task, we compare the machine translation approach and lexical substitu- …

Content truncated to 500 words