Auto-generating Bilingual Dictionaries: Results of the TIAD-2017 Shared Task Baseline Algorithm

Inferring a bilingual dictionary L1>L3 given two bilingual dictionaries L1>L2 and L2>L3 is a non-trivial task, as seen in reports of large-scale, computationally-heavy experiments published in recent years (Soderland et al. (2009); Shezaf and Rappoport (2010)). Early works on this (cf. Tanaka and Umemura (1994)) have already noticed that the main obstacle in such inferences stems from the fact that polysemy is not isomorphic across languages, and often a monosemous lexical item in L1 can be polysemous in L2. In this paper, we present a new set of experiments on automatically generating bilingual dictionaries based on existing ones. The data used are a commercial set of bilingual dictionaries with a particular topology when viewed as a graph connecting source and target languages. We find that searching for cycles in this graph is an effective method for generating translation inferences, and reflect on the impact of the source data’s structure on these results and directions for future research.