Comparing Statistical Models for Retrieval based Question-answering Dialogue: BERT vs Relevance Models

In this paper, we compare the performance of four models in a retrieval based question answering dialogue task on two moderately sized corpora (~ 10,000 utterances). One model is statistical and uses cross language relevance while others are deep neural networks utilizing BERT architecture along with different methods. The has previously outperformed LSTM similar whereas been proven to perform well variety NLP tasks, achieving state-of-the-art results many them. Results show that outperforms architectures learning question-answer mappings. achieves better by mapping new questions existing questions.