🧑‍🔬 Dianhai Yu

Dianhai Yu works on neural models for natural language processing, and their papers here span 2015 to 2018. Two problems recur: how to share learned representations across related tasks, and how to match pieces of text that relate to each other in indirect ways. The methods are attention-based neural architectures, first recurrent encoder-decoders for machine translation and later attention-only matching networks for conversational systems.

The most cited of these papers is Multi-Task Learning for Multiple Language Translation (2015). Instead of training a separate system for every language pair, it treats translation from English into several languages as one multi-task problem. A single bidirectional recurrent encoder is shared by all directions, and each target language has its own decoder and attention mechanism. The aim is to help language pairs that have little parallel text to learn from.

The later work moves to dialogue systems. Multi-Turn Response Selection for Chatbots with Deep Attention Matching Network (2018) considers a retrieval chatbot that must choose the best reply to a conversation of several turns. Its model drops recurrence and relies on attention alone, drawing on the Transformer. Stacked self-attention builds word representations at several levels of granularity, and cross-attention then aligns each earlier utterance with a candidate reply. This lets the model find replies tied to the context through links such as coreference rather than through shared words.

Yu's papers are mostly affiliated with Baidu in Beijing, where the 2015 translation work was carried out. Co-authors include Daxiang Dong, Hua Wu and Haifeng Wang.

Dianhai Yu has 2 papers in our library, in 2015–2018, most of them on Dialogue systems; often with Daxiang Dong, Hua Wu, Haifeng Wang.