Improving cross-lingual text matching with dual-level collaborative coarse-to-fine filter alignment network

Abstract

Semantic alignment is a key component in Cross-Language Text Matching (CLTM) to facilitate matching (e.g., query-document matching) between two languages. The current solutions for semantic alignment mainly perform word-level translation directly, without considering the contextual information for the whole query and documents. To this end, we propose a Dual-Level Collaborative Rough-to-Fine Filter Alignment Network (DLCCFA) to achieve better cross-language semantic alignment and document matching. DLCCFA is devised with both a coarse-grained filter in word-level and a fine-grained filter in sentence-level. Concretely, for the query in word-level, we firstly extract top-k translation candidates for each token in the query through a probabilistic bilingual lexicon. Then, a Translation Probability Attention (TPA) mechanism is proposed to obtain coarse-grained word alignment, which generates the corresponding query auxiliary sentence. Afterwards, we further propose a Bilingual Cross Attention and utilize Self-Attention to achieve fine-grained sentence-level filtering, resulting in the cross-language representation of the query. The idea is that each token in the query works as an anchor to filter the semantic noise in the query auxiliary sentence and accurately align semantics of different languages. Extensive experiments on four real-world datasets of six languages demostrate that our method can outperform the mainstream alternatives of CLTM.