R3: Reinforced Ranker-Reader for open-domain Question Answering

• Main Authors: WANG, Shuohang, YU, Mo, GUO, Xiaoxiao, WANG, Zhiguo, KLINGER, Tim, ZHANG, Wei, CHANG, Shiyu, TESAURO, Gerald, ZHOU, Bowen, JIANG, Jing
• Format: text
• Language: English
• Published: Institutional Knowledge at Singapore Management University 2018
• Subjects: Answer extraction; Ground truth; Neural network method; Open domain question answering; Question Answering; Reading comprehension; Databases and Information Systems; Theory and Algorithms
• Online Access: https://ink.library.smu.edu.sg/sis_research/4237; https://ink.library.smu.edu.sg/context/sis_research/article/5240/viewcontent/Reinforced_Reader_Ranker_2018_AAAI_afv.pdf
• Institution: Singapore Management University

In recent years researchers have achieved considerable success applyingneural network methods to question answering (QA). These approaches haveachieved state of the art results in simplified closed-domain settings such asthe SQuAD (Rajpurkar et al., 2016) dataset, which provides a pre-selectedpassage, from which the answer to a given question may be extracted. Morerecently, researchers have begun to tackle open-domain QA, in which the modelis given a question and access to a large corpus (e.g., wikipedia) instead of apre-selected passage (Chen et al., 2017a). This setting is more complex as itrequires large-scale search for relevant passages by an information retrievalcomponent, combined with a reading comprehension model that "reads" thepassages to generate an answer to the question. Performance in this settinglags considerably behind closed-domain performance. In this paper, we present anovel open-domain QA system called Reinforced Ranker-Reader (R3), based ontwo algorithmic innovations. First, we propose a new pipeline for open-domainQA with a Ranker component, which learns to rank retrieved passages in terms oflikelihood of generating the ground-truth answer to a given question. Second,we propose a novel method that jointly trains the Ranker along with ananswer-generation Reader model, based on reinforcement learning. We reportextensive experimental results showing that our method significantly improveson the state of the art for multiple open-domain QA datasets.