Unbiased Sentence Encoder For Large-Scale Multi-lingual Search Engines
This work addresses the challenge of biased and limited search data for multi-lingual search engines, offering an incremental improvement in encoder training methods.
The paper tackles the problem of training a multi-lingual sentence encoder for search engines by addressing biases in user search click data, which is skewed towards short queries and lacks coverage for unseen cases. The result is a universal encoder trained using a multi-task approach that combines public NLI datasets, translation data, and user search data to handle both short and long queries effectively.
In this paper, we present a multi-lingual sentence encoder that can be used in search engines as a query and document encoder. This embedding enables a semantic similarity score between queries and documents that can be an important feature in document ranking and relevancy. To train such a customized sentence encoder, it is beneficial to leverage users search data in the form of query-document clicked pairs however, we must avoid relying too much on search click data as it is biased and does not cover many unseen cases. The search data is heavily skewed towards short queries and for long queries is small and often noisy. The goal is to design a universal multi-lingual encoder that works for all cases and covers both short and long queries. We select a number of public NLI datasets in different languages and translation data and together with user search data we train a language model using a multi-task approach. A challenge is that these datasets are not homogeneous in terms of content, size and the balance ratio. While the public NLI datasets are usually two-sentence based with the same portion of positive and negative pairs, the user search data can contain multi-sentence documents and only positive pairs. We show how multi-task training enables us to leverage all these datasets and exploit knowledge sharing across these tasks.