Where to Submit? Helping Researchers to Choose the Right Venue

Findings of the Association for Computational Linguistics: EMNLP 2020, pages 65–83 November 16 - 20, 2020. c⃝2020 Association for Computational Linguistics 65 Reducing Sentiment Bias in Language Models via Counterfactual Evaluation Po-Sen Huang♠♦ Huan Zhang♥♥♦ Ray Jiang♠ Robert Stanforth♠ Johannes Welbl♠♣♥ Jack W. Rae♠♣ Vishal Maini♠ Dani Yogatama♠ Pushmeet Kohli♠ ♠DeepMind ♥University of California, Los Angeles ♣University College London Abstract Advances in language modeling architectures and the availability of large text corpora have driven progress in automatic text generation. While this results in models capable of gener- ating coherent texts, it also prompts models to internalize social biases present in the training corpus. This paper aims to quantify and reduce a particular type of bias exhibited by language models: bias in the sentiment of generated text. Given a conditioning context (e.g., a writing prompt) and a language model, we analyze if (and how) the sentiment of the generated text is affected by changes in values of sensitive attributes (e.g., country names, occupations, genders) in the conditioning context using a form of counterfactual evaluation. We quan- tify sentiment bias by adopting individual and group fairness metrics from the fair machine learning literature, and demonstrate that large- scale models trained on two different corpora (news articles, and Wikipedia) exhibit consid- erable levels of bias. We then propose embed- ding and sentiment prediction-derived regular- ization on the language model’s latent repre- sentations. The regularizations improve fair- ness metrics while retaining comparable levels of perplexity and semantic similarity. 1 Introduction Language modeling has advanced rapidly due to efficient model architectures (Vaswani et al., 2017; Dai et al., 2019) and the availability of large-scale datasets (Radford et al., 2019; Zellers et al., 2019). Large-scale language models have been applied not only for representation extraction to support downstream tasks (Peters et al., 2018; Devlin et al., 2019), but are also used for many natural language generation applications (Radford et al., 2019; So- laiman et al., 2019; Zellers et al., 2019; Zhang ♦Denotes equal contribution. ♥Work done during an internship at DeepMind. ♠Corresponding author: posenhuang@google.com. had a grand time organising...(0.97) hear from her all the time all the problems...(0.17) 're working on a prototype for her banana bread recipe...(0.51) Conditioning Text with Attribute Generated Continuations Sentiment Distribution My friend is a/ an _, and we... ... ... Figure 1: Conditioning text “My friend is a/an <occupation>, and we...”, alongside various text con- tinuations generated by a GPT-2 language model. On the right, the empirical sentiment distribution of the generated texts is shown: they reveal a system- atic difference in sentiment depending on occupation (“baker’’ or “accountant”) in the conditioning context. et al., 2019). While the generation of coherent text is becoming increasingly practical, it also prompts models to internalize social biases present in the training corpus. Investigating the social impact and fairness of the text generated from language models has thus received considerable research in- terest (Solaiman et al., 2019; Wallace et al., 2019; Sheng et al., 2019). In this paper, we aim to both quantify and …

Content truncated to 500 words