CycleKQR: Unsupervised Bidirectional Keyword-Question Rewriting

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 11875 - 11886 December 7-11, 2022 ©2022 Association for Computational Linguistics CycleKQR: Unsupervised Bidirectional Keyword-Question Rewriting Andrea Iovine1, Anjie Fang2, Besnik Fetahu2, Jie Zhao2, Oleg Rokhlenko2, Shervin Malmasi2 1University of Bari Aldo Moro, Italy 2Amazon.com, Inc., Seattle, WA, USA andrea.iovine@uniba.it {njfn,besnikf,olegro,malmasi}@amazon.com Abstract Users expect their queries to be answered by search systems, regardless of the query’s sur- face form, which include keyword queries and natural questions. Natural Language Under- standing (NLU) components of Search and QA systems may fail to correctly interpret seman- tically equivalent inputs if this deviates from how the system was trained, leading to subop- timal understanding capabilities. We propose the keyword-question rewriting task to improve query understanding capabilities of NLU sys- tems for all surface forms. To achieve this, we present CycleKQR, an unsupervised approach, enabling effective rewriting between keyword and question queries using non-parallel data. Empirically we show the impact on QA perfor- mance of unfamiliar query forms for open do- main and Knowledge Base QA systems (trained on either keywords or natural language ques- tions). We demonstrate how CycleKQR sig- nificantly improves QA performance by rewrit- ing queries into the appropriate form, while at the same time retaining the original semantic meaning of input queries, allowing CycleKQR to improve performance by up to 3% over su- pervised baselines. Finally, we release a dataset of 66k keyword-question pairs.1 1 Introduction Information Retrieval and Question Answering (QA) systems aim to fulfill user requests, from tra- ditional Web search to voice assistants. Users can issue queries in two forms (White et al., 2015): as keyword queries (“iphone 14 price”) or as fully- formed natural language questions (“how much is an iPhone 14?”). Natural Language Understand- ing (NLU) components are expected to correctly interpret semantically equivalent queries, regard- less of the form. Figure 1 (a) shows examples of queries with the same intent, but in different forms. 1https://github.com/amzn/kqr How much is iPhone 13? What is the price of iPhone 13? How much does iPhone 13 cost? iPhone 13 cost, cost of iPhone 13 iPhone 13price, price of iPhone 13 iPhone 13 offer, iPhone 13deal Questions Keywords The prices for iPhone 13 vary according to storage: $799 for 128GB. QA Answer Things to do Seattle Are you asking what tourist attractions are in Seattle ? Yes/No (a) (b) Figure 1: (a) Questions & keywords with the same intent. (b) Keyword-question rewriting for disambiguation. Work on question paraphrases has grown re- cently, with current approaches utilizing encoder- decoder architectures for generating questions (Hosking and Lapata, 2021). However, NLU sys- tems that handle different input forms must support two important functionalities beyond questions: • Understanding: Correctly interpret different query forms (cf. Figure 1 (a)), and create equiv- alent semantic representations, regardless of the query form, style, or other lexical variations. • Expression: Rewrite queries into different forms (e.g. keywords into questions). This increases transparency by relaying the query interpretation back to the users, or to ask clarification ques- …

Content truncated to 500 words