Acr2Vec: Learning Acronym Representations in Twitter

Acronyms are common in Twitter and bring in new challenges to social media analysis. Distributed representations have achieved successful applications in natural language processing. An acronym is different from a single word and is generally defined by several words. To this end, we present Acr2Vec, an algorithmic framework for learning continuous representations for acronyms in Twitter. First, a Twitter ACRonym (TACR) dataset is automatically constructed, in which an acronym is expressed by one or more definitions. Then, three acronym embedding models have been proposed: MPDE (Max Pooling Definition Embedding), APDE (Average Pooling Definition Embedding), and PLAE (Paragraph-Like Acronym Embedding). The qualitative experimental results (i.e., similarity measure) and quantitative experimental results (i.e., acronym polarity classification) both show that MPDE and APDE are superior to PLAE.