GCPG: A General Framework for Controllable Paraphrase Generation

Findings of the Association for Computational Linguistics: ACL 2022, pages 4035 - 4047 May 22-27, 2022 c⃝2022 Association for Computational Linguistics GCPG: A General Framework for Controllable Paraphrase Generation Kexin Yang♠ ∗ Dayiheng Liu♠ † Wenqiang Lei♢ Baosong Yang♠ Haibo Zhang♠ Xue Zhao♠ Wenqing Yao♠ Boxing Chen♠ ♠Alibaba Group ♢National University of Singapore {kexinyang0528, losinuris}@gmail.com Abstract Controllable paraphrase generation (CPG) in- corporates various external conditions to obtain desirable paraphrases. However, existing works only highlight a special condition under two in- dispensable aspects of CPG (i.e., lexically and syntactically CPG) individually, lacking a uni- fied circumstance to explore and analyze their effectiveness. In this paper, we propose a gen- eral controllable paraphrase generation frame- work (GCPG), which represents both lexical and syntactical conditions as text sequences and uniformly processes them in an encoder- decoder paradigm. Under GCPG, we recon- struct commonly adopted lexical condition (i.e., Keywords) and syntactical conditions (i.e., Part- Of-Speech sequence, Constituent Tree, Masked Template and Sentential Exemplar) and study the combination of the two types. In partic- ular, for Sentential Exemplar condition, we propose a novel exemplar construction method — Syntax-Similarity based Exemplar (SSE). SSE retrieves a syntactically similar but lex- ically different sentence as the exemplar for each target sentence, avoiding exemplar-side words copying problem. Extensive experiments demonstrate that GCPG with SSE achieves state-of-the-art performance on two popular benchmarks. In addition, the combination of lexical and syntactical conditions shows the sig- nificant controllable ability of paraphrase gener- ation, and these empirical results could provide novel insight to user-oriented paraphrasing. 1 Introduction Paraphrase generation (Madnani and Dorr, 2010) refers to restating a given sentence into an alter- native surface form while keeping the semantics unchanged.1 It is of long-standing interest (McKe- own, 1983), with various applications such as ques- ∗ Work is done during internship at DAMO Academy † Corresponding author. 1In this paper, we make our first attempt with English corpora. There was a picture of the revolving earth that have emerged. There was a picture of the rotating earth that showed up. showed up rotating Lexically Controlling [Keywords] A picture of the revolving earth emerged. Syntactically Controlling [Exemplar] The job at school went well. A picture of the rotating earth showed up. Combination Figure 1: A toy example to explain what effect lexically controlling and syntactically controlling have on para- phrasing. tion answering (Gan and Ng, 2019), machine trans- lation (Mallinson et al., 2017), and sentence simpli- fication (Martin et al., 2020). However, a sentence can be re-expressed in various surface forms. Lack- ing control might result in undesirable results (Gu et al., 2019). For example, a sentence that contains an intricate syntactic structure may cause difficul- ties for aphasic patients (Shewan, 1985). In that case, we could attempt to paraphrase it based on syntactical control. To obtain desirable surface forms, most recent works focus on controllable paraphrase generation (CPG) by incorporating external conditions. Ex- isting efforts to CPG can be roughly divided into two types: lexically and syntactically CPG. Lex- ically CPG is concerned with …

Content truncated to 500 words