User-Centric Gender Rewriting

Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 618 - 631 July 10-15, 2022 ©2022 Association for Computational Linguistics User-Centric Gender Rewriting Bashar Alhafni, Nizar Habash, Houda Bouamor† Computational Approaches to Modeling Language Lab New York University Abu Dhabi †Carnegie Mellon University in Qatar {alhafni,nizar.habash}@nyu.edu, hbouamor@qatar.cmu.edu Abstract In this paper, we define the task of gen- der rewriting in contexts involving two users (I and/or You) – first and second grammatical persons with independent grammatical gender preferences. We focus on Arabic, a gender- marking morphologically rich language. We develop a multi-step system that combines the positive aspects of both rule-based and neu- ral rewriting models. Our results successfully demonstrate the viability of this approach on a recently created corpus for Arabic gender rewriting, achieving 88.42 M2 F0.5 on a blind test set. Our proposed system improves over previous work on the first-person-only version of this task, by 3.05 absolute increase in M2 F0.5. We demonstrate a use case of our gender rewriting system by using it to post-edit the output of a commercial MT system to provide personalized outputs based on the users’ gram- matical gender preferences. We make our code, data, and pretrained models publicly available.1 1 Introduction Gender bias is a fundamental problem in natural language processing (NLP) and it has been re- ceiving an increasing attention across a variety of core tasks such as machine translation (MT), co-reference resolution, and dialogue systems. Re- search has shown that NLP systems have the ability to embed and amplify gender bias (Sun et al., 2019), which not only degrades users’ experiences but also creates representational harm (Blodgett et al., 2020). The embedded bias within NLP systems is usually attributed to training models on biased data that reflects the social inequalities of the world we live in. However, even the most balanced of models can still exhibit and amplify bias if they are designed to produce a single text output without taking their users’ gender preferences into consider- ation. Therefore, to provide the correct user-aware 1https://github.com/CAMeL-Lab/ gender-rewriting/ output, NLP systems should be designed to pro- duce outputs that are as gender-specific as the users information they have access to. Users information could be either embedded as part of the input or provided externally by the users themselves. In cases where this information is unavailable to the system, generating all gender-specific forms or a gender-neutral form is more appropriate. Producing user-aware outputs becomes more challenging for systems targeting multi-user con- texts (first and second persons, with indepen- dent grammatical gender preferences), particularly when dealing with gender-marking morphologi- cally rich languages. In this paper, we define the task of gender rewriting in contexts involving two users (I and/or You) – first and second grammatical persons with independent grammatical gender pref- erences and we focus on Arabic, a gender-marking morphologically rich language. The main contribu- tions of our work are as follows: 1. We introduce a multi-step gender rewriting system that combines …

Content truncated to 500 words