A Transition-based Method for Complex Question Understanding

Proceedings of the 29th International Conference on Computational Linguistics, pages 4203–4211 October 12–17, 2022. 4203 A Transition-based Method for Complex Question Understanding Yu Xia1, ∗, Wenbin Jiang2, Yajuan Lyu2, Sujian Li1, † 1MOE Key Lab of Computational Linguistics, Peking University, Beijing, China 2Baidu Inc., Beijing, China {yuxia, lisujian}@pku.edu.cn {jiangwenbin, lvyajuan}@baidu.com Abstract Complex Question Understanding (CQU) parses complex questions to Question De- composition Meaning Representation (QDMR) which is a sequence of atomic operators. Ex- isting works are based on end-to-end neural models which do not explicitly model the inter- mediate states and lack interpretability for the parsing process. Besides, they predict QDMR in a mismatched granularity and do not model the step-wise information which is an essential characteristic of QDMR. To alleviate the issues, we treat QDMR as a computational graph and propose a transition-based method where a de- cider predicts a sequence of actions to build the graph node-by-node. In this way, the partial graph at each step enables better representa- tion of the intermediate states and better inter- pretability. At each step, the decider encodes the intermediate state with specially designed encoders and predicts several candidates of the next action and its confidence. For inference, a searcher seeks the optimal graph based on the predictions of the decider to alleviate the error propagation. Experimental results demonstrate the parsing accuracy of our method against sev- eral strong baselines. Moreover, our method has transparent and human-readable intermedi- ate results, showing improved interpretability. 1 Introduction The task of complex question understanding (CQU) aims at converting complex questions which re- quire multi-hop reasoning into consecutive trivial sub-questions. An example of CQU is shown in Figure 1. To answer the question "return me the author in the University of Michigan whose papers have more than 5000 total citations", CQU models decompose it into several trivial sub-questions (e.g. "return authors"), and the final answer is obtained by consecutively answering the sub-questions. To ∗ This work was done during internship at Baidu Inc. † Corresponding author. Sentence return authors [SEP] return #1 in the University of Michigan [SEP] return papers of #2 [SEP] return citations of #3 [SEP] return the number of #4 for each #2 [SEP] return #2 where #5 is more than 5000 1 Dependency Graph 2 3 Computational Graph return authors return #1 in the University of Michigan return papers of #2 return citations of #3 return the number of #4 for each #2 return #2 where #5 is more than 5000 1 2 3 4 5 6 Question: Return me the author in the University of Michigan whose papers have more than 5000 total citations Figure 1: An example of CQU consists of a complex question and different modeling of QDMR. capture the meaning of questions over unstruc- tured sources such as text and images, Question Decomposition Meaning Representation (QDMR) (Wolfson et al., 2020) is proposed where questions are represented through a sequence of atomic ex- ecutable operators, and the final answer can be obtained by answering the operator sequences in or- …

Content truncated to 500 words