Queens are Powerful too: Mitigating Gender Bias in Dialogue Generation (1911.03842v2)

Abstract: Models often easily learn biases present in the training data, and their predictions directly reflect this bias. We analyze gender bias in dialogue data, and examine how this bias is actually amplified in subsequent generative chit-chat dialogue models. We measure gender bias in six existing dialogue datasets, and focus on the most biased one, the multi-player text-based fantasy adventure dataset LIGHT, as a testbed for our bias mitigation techniques. The LIGHT dataset is highly imbalanced with respect to gender, containing predominantly male characters, likely because it is entirely collected by crowdworkers and reflects common biases that exist in fantasy or medieval settings. We consider three techniques to mitigate gender bias: counterfactual data augmentation, targeted data collection, and bias controlled training. We show that our proposed techniques mitigate gender bias in LIGHT by balancing the genderedness of generated dialogue utterances and are particularly effective in combination. We quantify performance using various evaluation methods---such as quantity of gendered words, a dialogue safety classifier, and human studies---all of which show that our models generate less gendered, but equally engaging chit-chat responses.

Mitigating Gender Bias in Dialogue Generation

In this paper, the authors investigate the presence and amplification of gender bias in dialogue generation models, with a focus on data-driven methodologies and contemporary language processing models. The paper systematically addresses both the identification and mitigation of gender bias using various techniques in a predominantly male-biased dataset, referred to as LIGHT.

Measurement of Gender Bias

The authors carry out an initial evaluation to identify gender bias in several dialogue datasets, including LIGHT, ConvAI2, Reddit, and others. They measure gender bias by calculating the percentage of gendered words and the proportion of male-biased words in these datasets. LIGHT emerges as markedly male-biased, making it a focal point for addressing gender bias.

1. Character Gender Imbalance: LIGHT demonstrates a notable imbalance, with 1.6 times as many male characters as female ones. This stark asymmetry is quantitatively detailed in Table \ref{table:gender_balance}, contrasting LIGHT with relatively balanced datasets like ConvAI2.
2. Persona Bias: Personas within LIGHT often augment biases through their descriptions, emphasizing gender roles rooted in historical contexts. A qualitative analysis illustrated this, with examples like female personas portraying stereotypical roles, such as household chores and childcare, as highlighted in Table \ref{table:personaexamples}.
3. Dialogue Bias: Bias propagation from personas to dialogues is evident, with dialogues reflecting skewed gender roles (see Table \ref{table:dialogueexample} for representative dialogues). The authors employ textual analysis to identify and quantify gendered language uses, outlining profound male bias amplified during the dialogue generation process.

Techniques for Bias Mitigation

The paper explores three strategies to mitigate gender biases, particularly in LIGHT:

• Counterfactual Data Augmentation (CDA): Adapting techniques usually applied to word embeddings, the authors extend CDA to dialogue data—systematically swapping gendered words to produce diversified train sets. This approach, however, risks generating ungrammatical or nonsensical sentences due to cross-gender swaps.
• Positive-Bias Data Collection: The authors curate new dataset personas by manually editing existing ones and generating novel profiles that promote gender balance. They couple this with targeted dialogue collection efforts that seek diversified dialogue utterances, effectively reducing bias (quantitatively shown in Table \ref{table:gender_balance}).
• Bias Controlled Training: Utilizing conditional training, models are trained using special tokens indicating genderedness to modulate the number and nature of gendered words generated. Models can be adjusted to reduce bias during inference by specifying desired gender configurations.

Results and Evaluation

The evaluation of these techniques is conducted using both quantitative metrics and qualitative assessments:

• Quantitative Analysis: In Figure \ref{table:sensitive_f1_results}, performance metrics such as percentage of gendered words and male bias reveal advances in bias reduction and general F1-score improvement through combined techniques. The ALL model—merging CDA, positive data collection, and bias control—demonstrated optimal gender-balanced output control.
• Safety and Quality Assessment: The paper presents safety and quality measures through classifier-driven offensive language detection (Table \ref{table:offensive}), coupled with human evaluations of dialogue engagingness (Figure \ref{fig:human_eval}), revealing the thoughtful balance achieved between gender-bias mitigation and dialogue quality.

Implications and Future Directions

This paper contributes foundational insights into managing biases during dialogue model training and generation. Practically, methods like bias-controlled training offer flexibility in real-world applications, where customizable dialogue outputs can be pivotal. Theoretically, the paper underscores the intricacies of gendered language dynamics and the importance of early bias intervention during dataset creation.

Future research directions may involve refining gender classifiers to encompass a wider spectrum of gendered terms, thereby enhancing robustness. Additionally, exploring bias control for other socially relevant aspects, beyond gender, could further broaden the applicability of bias mitigation techniques.

Overall, the paper lays significant groundwork for developing equitable dialogue systems, catalyzing advances in AI applications that navigate dynamic linguistic, cultural, and social landscapes.