Learning to Generate Equitable Text in Dialogue from Biased Training Data

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 2898–2917 July 9-14, 2023 ©2023 Association for Computational Linguistics Learning to Generate Equitable Text in Dialogue from Biased Training Data Anthony Sicilia Intelligent Systems Program University of Pittsburgh anthonysicilia@pitt.edu Malihe Alikhani School of Computing and Information University of Pittsburgh malihe@pitt.edu Abstract The ingrained principles of fairness in a dia- logue system’s decision-making process and generated responses are crucial for user engage- ment, satisfaction, and task achievement. Ab- sence of equitable and inclusive principles can hinder the formation of common ground, which in turn negatively impacts the overall perfor- mance of the system. For example, misusing pronouns in a user interaction may cause ambi- guity about the intended subject. Yet, there is no comprehensive study of equitable text gen- eration in dialogue. Aptly, in this work, we use theories of computational learning to study this problem. We provide formal definitions of equity in text generation, and further, prove formal connections between learning human- likeness and learning equity: algorithms for im- proving equity ultimately reduce to algorithms for improving human-likeness (on augmented data). With this insight, we also formulate rea- sonable conditions under which text generation algorithms can learn to generate equitable text without any modifications to the biased train- ing data on which they learn. To exemplify our theory in practice, we look at a group of algo- rithms for the GuessWhat?! visual dialogue game and, using this example, test our the- ory empirically. Our theory accurately predicts relative-performance of multiple algorithms in generating equitable text as measured by both human and automated evaluation. 1 Introduction Machine learning models for text-generation in dia- logue have trouble learning the “long tail” of a data distribution; i.e., the data concepts not frequently observed during training. For example, dataset bi- ases like gender imbalance can induce a long tail in training data whereby important data relation- ships involving gender are underrepresented, like women in sports (Hendricks et al., 2018). When training, generative models often fail to learn these concepts in the long tail, and ultimately, learn in- Figure 1: Examples from GuessWhat?! dataset, which consists of game dialogues where a question-player queries an answer-player to identify a secret goal object known only to the answer-player. Images containing men outnumber images with women 2 to 1, forming a "long tail" in the data distribution. In examples above, human annotations use visual (contextual) cues to agree on gender or balk when appropriate. Traditional algo- rithms can be incorrect and overconfident, inheriting dataset bias towards male pronouns. Algorithms mo- tivated by learning theory (LEATHER) are more robust, utilizing context in a human-like way. equitable, stereotyping behaviors instead (see Fig- ure 1). These non-inclusive behaviors not only de- crease user-satisfaction by isolating users (Mehrabi et al., 2021), but also impede common ground, hin- dering the task-success of the dialogue system. Despite the multi-faceted impact of inequitable text generation in dialogue, we do not have a com- prehensive and theoretically grounded framework …

Content truncated to 500 words