Don’t Change Me! User-Controllable Selective Paraphrase Generation

Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics, pages 3522–3527 April 19 - 23, 2021. ©2021 Association for Computational Linguistics 3522 Don’t Change Me! User-Controllable Selective Paraphrase Generation Mohan Zhang1,2, Luchen Tan1, Zhengkai Tu1, Zihang Fu1, Kun Xiong1, Ming Li1,3, Jimmy Lin1,3 1RSVP.ai, 2 University of Toronto, 3David R. Cheriton School of Computer Science, University of Waterloo zhangmo4@cs.toronto.edu, lctan@rsvp.ai, siriuself1103@gmail.com, {zhfu,kun}@rsvp.ai, {mli,jimmylin}@uwaterloo.ca Abstract In the paraphrase generation task, source sen- tences often contain phrases that should not be altered. Which phrases, however, can be context dependent and can vary by applica- tion. Our solution to this challenge is to pro- vide the user with explicit tags that can be placed around any arbitrary segment of text to mean “don’t change me!” when generating a paraphrase; the model learns to explicitly copy these phrases to the output. The con- tribution of this work is a novel data genera- tion technique using distant supervision that allows us to start with a pretrained sequence- to-sequence model and fine-tune a paraphrase generator that exhibits this behavior, allowing user-controllable paraphrase generation. Addi- tionally, we modify the loss during fine-tuning to explicitly encourage diversity in model out- put. Our technique is language agnostic, and we report experiments in English and Chinese. 1 Introduction Notions of semantic similarity and paraphrase are highly context dependent. Consider “I’m looking for cheap hotels in New York.” vs. “What are cheap lodging options in Beijing?”: from the perspec- tive of intent classification, both express similar intents, but from the perspective of paraphrasing in a community QA application, a user looking for the answer to one question would not find the other response helpful. This is because the location “New York” anchors the information need, and any changes to it would be unacceptable to the user. It is not always the case that named entities are “immutable” in this respect: consider a user looking for vacation destinations in the South of France. From the perspective of an advertiser, there might be good reason to tempt the user with alternative locations such as the Italian Riviera; the user may even welcome these suggestions. Also, it is not always the case that these immutable anchors are named entities: For example, some metrics assign high similarity to antonyms, and so “cheap hotels” and “expensive hotels” might be considered seman- tically close, but obviously not from the perspective of an end user looking for inexpensive lodging. Although whether or not certain words can be changed without affecting the meaning of a sen- tence is highly dependent on context, the user of a paraphrase generation system usually would know. Consider the application of paraphrase generation in a community QA application, where a developer wishes to automatically generate question variants to increase the chances of a semantic match: A na¨ıve system will indeed generate “What are cheap lodging options in Beijing?” as a paraphrase to “I’m looking for cheap hotels in New York.” What if we provide the user with a …

Content truncated to 500 words