Multi-step Retriever-Reader Interaction for Scalable Open-domain Question Answering

TL;DR

The paper presents a scalable open-domain QA framework where a fast paragraph retriever and a neural reader iteratively interact via a multi-step-reasoner that reformulates the query based on the reader’s state. Paragraph representations are cached offline, enabling efficient MIPS-based retrieval over millions of paragraphs, while the reader aggregates evidence across multiple paragraphs. The approach is reader-architecture agnostic and shows consistent improvements for DrQA and BiDAF across TriviaQA-open, Triviaqa-unfiltered, SearchQA, Quasar, and SQuAD-Open. Training combines distant supervision, reinforcement learning, and a pretraining step to optimize query reformulation for improved evidence gathering. The results demonstrate scalability and robust gains from multi-step reasoning, with substantial performance improvements over baselines on several large-scale open-domain QA benchmarks.

Abstract

This paper introduces a new framework for open-domain question answering in which the retriever and the reader iteratively interact with each other. The framework is agnostic to the architecture of the machine reading model, only requiring access to the token-level hidden representations of the reader. The retriever uses fast nearest neighbor search to scale to corpora containing millions of paragraphs. A gated recurrent unit updates the query at each step conditioned on the state of the reader and the reformulated query is used to re-rank the paragraphs by the retriever. We conduct analysis and show that iterative interaction helps in retrieving informative paragraphs from the corpus. Finally, we show that our multi-step-reasoning framework brings consistent improvement when applied to two widely used reader architectures DrQA and BiDAF on various large open-domain datasets --- TriviaQA-unfiltered, QuasarT, SearchQA, and SQuAD-Open.

Figures (5)

• Figure 1: Our framework unrolled for two steps. The initial query is encoded and the retriever sends the top-$k$ paragraphs to the reader. The multi-step-reasoner component of our model takes in the internal state of the reader model and the previous query vector and does a gated update to produce a reformulated query. This new query vector is used by the retriever to re-rank the paragraphs and send different paragraphs to the reader. Thus the multi-step-reasoner facilitates iterative interaction between the retriever (search engine) and the reader (QA model)
• Figure 2: Scalability of retriever.
• Figure 3: Retrieval performance on Quasar-t. The match-LSTM based retriever of R$^3$ is a more powerful model than our intial retrieval model. However, after few steps of multi-step-reasoner, the performance increases suggesting that re-ranking via query-reformulation is retrieving relevant evidence from the corpus. We report the P@$k$ on the last step.
• Figure 3: F1 score w.r.t number of steps.
• Figure 4: Examples of how multi-step-reasoner iteratively modifies the query by reading context to find more relevant paragraphs. Figure (left) shows an example where the initial retrieved context did not have the answer but the context provided enough hint to get more relevant paragraph in the next step. In figure (right), both the retrieved paragraph have the answer string leading to a boost in the score of the answer span because of score aggregation of spans (§\ref{['sub:reader_model']}).