New Trends in MRC III

Conversational MRC (CMRC)

  • Most natural way people acquire knowledge is via a series of interrelated question-and-answer processes – multi-turn conversation

  • In CMRC, conversation history H acts as part of the context to help predict answers

  • Related datasets
    • CoQA
      • 8000 conversations about passages in seven domains. No limit to the answer forms, which requires more contextual reasoning

    • QuAC
      • The passage is given only to the answerer and the questioner asks questions based on the title of the passages

      • The answerer answers questions with spans of text of the original passage and determine whether the questioner can ask a follow-up question

  • Challenges in CMRC:
    • Conversational History
      • A follow-up question may be closely related to prior questions and answers

      • Dialogue pairs as conversational history are fed to CMRC systems as inputs

    • Coreference Resolution
      • Two types
        • Explicit
          • There are explicit markers, such as personal pronouns
            • For example, to answer the question, who had a birthday?
              • The model need to understand that “Today was her birthday”, the “her” is referring to “Jessica”

        • Implicit
          • Much harder to figure out. Short questions with certain intentions that implicitly refer to previous content
            • For example, question could be “Did she planned to have visitors?”, with a follow-up question of “How many visitors are there?”

      • Reddy et al. propose a hybrid model, DrQA + PGNet
        • Combines seq2seq and MRC models to extract and generate answers

        • They integrate information on conversational history by treating previous QA pairs as a sequence and append them to the context

      • Yatskar et al. use BiDAF++ with ELMo
        • Answer questions based on given context and conversational history

        • Rather than encoding previous dialogue information on context representation, they label answers to previous questions

      • Huang et al. introduced a flow mechanism to deeply understand conversational history
        • Encodes hidden context representations during the process of answering previous questions



Data Scientist

Leave a Reply