2021/09/26 by Kelly Marchisio, Marchisio, Kelly, Youngser Park +11
Computer Science · #Advanced Graph Neural Networks #Computation and Language (cs.CL) #FOS: Computer and information sciences #Natural Language Processing Techniques #Topic Modeling
paper · pdf · doi:10.48550/arxiv.2109.12640
openalex publication_date 2021/09/26 · openalex created_date 2022/07/25 · openalex updated_date 2026/07/28
Much recent work in bilingual lexicon induction (BLI) views word embeddings\nas vectors in Euclidean space. As such, BLI is typically solved by finding a\nlinear transformation that maps embeddings to a common space. Alternatively,\nword embeddings may be understood as nodes in a weighted graph. This framing\nallows us to examine a node's graph neighborhood without assuming a linear\ntransform, and exploits new techniques from the graph matching optimization\nliterature. These contrasting approaches have not been compared in BLI so far.\nIn this work, we study the behavior of Euclidean versus graph-based approaches\nto BLI under differing data conditions and show that they complement each other\nwhen combined. We release our code at\nhttps://github.com/kellymarchisio/euc-v-graph-bli.\n