Our Contributions

1. We are the first to study cross-lingual question retrieval in programming cQA, using a combination of machine translation and a monolingual model. We observe that the translation quality strongly influences performance.
2. We leverage back-translation to adapt an NMT model to our domains' idiosyncrasies and reveal that this leads to considerable improvements.
3. We narrow the gap to a setup with access to an external state-of-the-art commercial MT service by up to 85%.

Improving Cross-Lingual Question Retrieval with Back-Translation

We leverage a standard approach to cross-lingual information retrieval, depicted below (figure 1 of our paper), which is to machine translate the query to our target language (NMT) and then continue with a monolingual question retrieval model for re-ranking (QR).

We investigate two adaptations if this approach, which both rely on back-translation (depicted below; table 1 of our paper):

1. We adapt a state-of-the-art neural machine translation model to our domains' idiosyncrasies by extending the training corpus with synthetic parallel in-domain sentences from cQA forums. We obtain such sentences by first training an NMT model on out-of-domain data and translate in-domain sentences to our target language with it. We then extend the training corpus with these synthetic parallel sentences and train a new NMT model with this data. This method we use is known as back-translation in machine translation literature.
2. We enhance the robustness of a neural question similarity model to common translation errors by adding back-translated texts during training. This time, we translate questions from our source language to the target language and back. This technique is sometimes used in NLP for automatic paraphrasing.

Both adaptations yield considerable improvements and narrow the gap to a setting with an external state-of-the-art commercial MT service by up to 85%.