Arabic Lexical Substitution: AraLexSubD Dataset and AraLexSub Pipeline

Abstract

Lexical substitution aims to generate a list of equivalent substitutions (i.e., synonyms) to a sentence’s target word or phrase while preserving the sentence’s meaning to improve writing, enhance language understanding, improve natural language processing models, and handle ambiguity. This task has recently attracted much attention in many languages. Despite the richness of Arabic vocabulary, limited research has been performed on the lexical substitution task due to the lack of annotated data. To bridge this gap, we present the first Arabic lexical substitution benchmark dataset AraLexSubD for benchmarking lexical substitution pipelines. AraLexSubD is manually built by eight native Arabic speakers and linguists (six linguist annotators, a doctor, and an economist) who annotate the 630 sentences. AraLexSubD covers three domains: general, finance, and medical. It encompasses 2476 substitution candidates ranked according to their semantic relatedness. We also present the first Arabic lexical substitution pipeline, AraLexSub, which uses the AraBERT pre-trained language model. The pipeline consists of several modules: substitute generation, substitute filtering, and candidate ranking. The filtering step shows its effectiveness by achieving an increase of 1.6 in the F1 score on the entire AraLexSubD dataset. Additionally, an error analysis of the experiment is reported. To our knowledge, this is the first study on Arabic lexical substitution.

Keywords

• lexical substitution (LS)
• evaluation dataset
• pre-trained language model BERT