Canonicalizing Knowledge Bases for Recruitment Domain

Online recruitment industry holds large amount of user-generated content in the form of job postings, resumes etc. This content finds its way in the knowledge bases (KB) causing duplicate and non-standard representations of entities (like company names, institute names, designations, skills etc.) These non-standard entity representations impact various applications such as search, recommendations and information retrieval. Therefore, KB canonicalization i.e, mapping multiple references of same entities into unique clusters is imperative for online recruitment platforms. Research suggests various approaches that use enriched semantic context or external context (from sources like Freebase) to perform KB Canonicalization. In fields where such external sources of context do not exist the problem remains challenging. To address these challenges, we propose a novel deep Siamese architecture with character-based attention and word embeddings that (a) estimates pairwise similarity between all entity mentions, and (b) then uses these similarity (scores) to create canonical clusters representing unique entity in the KB. Our experiments on recruitment domain dataset comprising of 62,288 unique entities of various types such as companies, institutes, skills, and designations demonstrate the effectiveness of our approach. We also provide insights on different network architectures, each of which encapsulate a different set of variation while performing canonicalization.