IGA: An Intent-Guided Authoring Assistant

Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 5972–5985 November 7–11, 2021. c⃝2021 Association for Computational Linguistics 5972 IGA: An Intent-Guided Authoring Assistant Simeng Sun1∗ Wenlong Zhao1 Varun Manjunatha2 Rajiv Jain2 Vlad Morariu2 Franck Dernoncourt2 Balaji Vasan Srinivasan2 Mohit Iyyer1 University of Massachusetts Amherst1 Adobe Research2 {simengsun,wenlongzhao,miyyer}@cs.umass.edu {vmanjuna, rajijain, morariu, dernonco, balsrini}@adobe.com Abstract While large-scale pretrained language mod- els have significantly improved writing as- sistance functionalities such as autocomplete, more complex and controllable writing assis- tants have yet to be explored. We leverage ad- vances in language modeling to build an in- teractive writing assistant that generates and rephrases text according to fine-grained au- thor specifications. Users provide input to our Intent-Guided Assistant (IGA) in the form of text interspersed with tags that correspond to specific rhetorical directives (e.g., adding de- scription or contrast, or rephrasing a particu- lar sentence). We fine-tune a language model on a dataset heuristically-labeled with author intent, which allows IGA to fill in these tags with generated text that users can subsequently edit to their liking. A series of automatic and crowdsourced evaluations confirm the quality of IGA’s generated outputs, while a small- scale user study demonstrates author prefer- ence for IGA over baseline methods in a cre- ative writing task. We release our dataset, code, and demo to spur further research into AI-assisted writing. 1 Introduction Writing can be both an exhilarating, creative expe- rience as well as a frustrating slog. Can recent ad- vances in neural language modeling help improve the human writing experience? Pretrained Trans- former language models (Radford et al., 2019) have improved writing aids such as email “auto- complete” (Chen et al., 2019), while commercial tools such as Grammarly and Microsoft Editor can rewrite full sentences to increase clarity.1 Few existing writing assistants provide sup- port for the underlying cognitive process of writ- ing (Greer et al., 2016). In this paper, we ex- plore more advanced writing assistance functions: ∗Most of the work done during an internship at Adobe. 1As these systems are not open-sourced, it is unclear how exactly they are implemented. CONTEXT: Sarah�didn't�come�to�school�today, CONTRAST EFFECT CAUSE DESCRIPTION BIOGRAPHY IDIOM REPHRASE homework news exam <NA> <NA> <NA> <NA> because�she�had�to�get�her�homework�done. therefore�there�is�no�more�big�news! but�still�took�the�exam�and�had�it�done. it�was�an�unplanned�break. David,�the�principal,�believes�she�will�miss tomorrow's�class. She�was�busy�making�ends�meet. Sarah�did�not�arrive�at�school�today. Figure 1: General concept overview of our Intent- Guided Authoring assistant IGA. Given context, by specifying different writing intents, the system gener- ates output satisfying the intent. In addition to well- formed sentence fragments, keywords can also be part of user input, serving as arguments for the intents, and are preserved in the output. specifically, we build an authoring assistant capa- ble of following fine-grained user directives (e.g., add descriptive text, use idiomatic language, or paraphrase a clunky bit of wording). Our system, the Intent-Guided Assistant (IGA), combines con- trollable text generation with text infilling (Zhu et al., 2019; Keskar et al., 2019a; Lewis et al., 2020; Donahue et al., 2020); more specifically, we adapt the tag-based control of Keskar et al. (2019b) to include a set of …

Content truncated to 500 words