Zero-Shot Language Transfer for Cross-Lingual Sentence Retrieval Using Bidirectional Attention Model

We present a neural architecture for cross-lingual mate sentence retrieval which encodes sentences in a joint multilingual space and learns to distinguish true translation pairs from semantically related sentences across languages. The proposed model combines a recurrent sequence encoder with a bidirectional attention layer and an intra-sentence attention mechanism. This way the final fixed-size sentence representations in each training sentence pair depend on the selection of contextualized token representations from the other sentence. The representations of both sentences are then combined using the bilinear product function to predict the relevance score. We show that, coupled with a shared multilingual word embedding space, the proposed model strongly outperforms unsupervised cross-lingual ranking functions, and that further boosts can be achieved by combining the two approaches. Most importantly, we demonstrate the model’s effectiveness in zero-shot language transfer settings: our multilingual framework boosts cross-lingual sentence retrieval performance for unseen language pairs without any training examples. This enables robust cross-lingual sentence retrieval also for pairs of resource-lean languages, without any parallel data.