Activate 2018

Information Retrieval (IR) based question answering systems have many applications in the real world. Recent advances in DL give us a huge possibility to improve IR apps and engines, and allow us to incorporate systems like chatbots. In this talk we will show our study on comparing traditional ML models vs DL models (both supervised and unsupervised) for different QA tasks such as answer paragraph selection, question-question similarity (FAQ matching) and answer span selection, and discuss the pros and cons of each method. For instance, using modern state-of-the-art DL models is quite expensive and cannot be easily scaled, thus we will present how to leverage Solr's payloads and indexes to improve runtime performance of DL and other ML models.