SCA-CLS: A New Semantic-Context-Aware Framework for Community-Oriented Lexical Simplification.

Community-oriented lexical simplification aims to transform complex words within a sentence into semantically consistent but simple substitute words from a community-specific vocabulary. Most state-of-the-art contextual word embedding models generate substitutes by extracting contextual information of complex words. Although these models take context into account, they fail to capture rich semantics of complex words with polysemy, resulting in many spurious and semantically non-equivalent candidates. Thus, this paper proposes a novel Semantic-Context-Aware framework for Community-oriented Lexical Simplification (SCA-CLS), which integrates gloss (sense definition) into BERT to identify the actual sense of the complex word (especially for polysemy) in current context and ranks substitutes by proposed gloss similarity. In addition, a new complexity feature is proposed to enhance substitute ranking. Experiment results on Wikipedia dataset show that SCA-CLS outperforms the state-of-the-art Merge-Sort model on both substitute generation and ranking tasks, indicating its effectiveness for community-oriented lexical simplification.