Deep Hashing Retrieval of Dynamic Corpora

Knowledge-intensive language applications, like open-domain question answering, call for NLP systems to retrieve related evidence in a document collection. A efficient and scalable retrieval system is an ideal component in these applications. Up-to-data research developed sequence-to-sequence framework on this task, eschewing any format of term-frequency or query-document similarity calculation. This line of work greatly improved the efficient of retrieval but with the sacrifice of scalablity. This work proposes a new framework based on the seq2seq architecture, which can solve scalability issues that exist in such methods. Experimental results on the public dataset Natural Question demonstrate the effectiveness of this approach.