[3Q5-GS-9] Natural language processing, information retrieval: Semantic similarity

[3Q5-GS-9-02] Cross-lingual semantic textual similarity measures based on Universal Dependencies

〇Takaaki Tanaka¹, Yuki Arase², Masaaki Nagata¹, Onizuka Makoto² (1. NTT Communication Science Laboratories, 2. Graduate school of information science and technology, Osaka University)

Keywords:semantic textual similarity, semantic parsing, cross-lingual processing

We present a novel method for measuring cross-lingual semantic textual similarity (CL STS) based on semantic graphs. We employed UDepLambda as semantic a representation, which is a semantic interface
for Universal Dependencies (UD), to provide logical forms in an almost language-independent manner
and process dependency graphs.
Our proposed method aligns semantic graphs and enables us
to capture the similar and different points in the target sentence pairs.
We also combined conventional word similarity-based features with semantic graph-based features to increase the robustness of our method.
This feature combination outperforms other semantic graph-based methods
in our evaluation on STS benchmarks.