Multi-lingual text clustering method using bilingual semantic correspondence analysis

Among most existing methods of multi-lingual text clustering, the dictionary-based approaches did not effectively deal with ambiguity and polysemy problem, the approaches based on machine translation inevitably introduced noise, Cross-lingual latent semantic indexing (CL-LSI) does not fully take into account bilingual semantic relationship. The paper proposes a new method based on Bilingual Semantic Correspondence Analysis (BiSCAN). In the new method, a bilingual document-pair are looked as two views for the same semantic content and a single topic space is created for each language by using partial least squares (PLS) method in order to capture statistical dependencies of bilingual semantic correspondence. The topic spaces are merged for document clustering task. Experimental results on Chinese-English stories collected from the Hong Kong government news website show that performance of the presented method is over or near to mono-lingual clustering in the original feature spaces and significantly better than CL-LSI.