Learning to Generate Equitable Text in Dialogue from Biased Training Data

• Long-tail concepts are underrepresented in training, and inequitable generation can isolate users, reduce satisfaction, and hinder common ground and dialogue task success.Source: §1 · p. 1
• The paper defines precise constraints for diverse notions of dialogue equity and compares them with traditional classification notions of equity.Source: §1 · p. 2
• Learning-theoretic algorithms can learn equitable dialogue through data augmentation, and under stated conditions can do so from biased training data without modifying it.Source: §1 · p. 2
• The theory predicts which algorithms may be more equitable on unseen data and identifies cases where naïve bias-removal augmentation can hurt equity.The authors report that experiments confirm the counter-intuitive augmentation finding.Source: §1 · p. 2

2.1 Learning Theory for Dialogue

• The goal distribution pairs dialogue contexts with goal dialogues, while the learned dialogue distribution conditions generated dialogues on model parameters and context.Source: §2.1 · p. 2
• A test maps dialogue to a score and can represent BLEU, sentiment, or human evaluation; test divergence compares scores on goal and predicted dialogue, with larger values indicating poorer replication along the tested dimension.Source: §2.1 · p. 3

2.2 Related Works on Equity

• Prior work addresses inequity through representation debiasing and evaluation or generation-scoring methods; this paper takes an extrinsic, dialogue-focused approach with formal theoretical analysis.Source: §2.2 · p. 3
• Relative to Liu et al. (2020)'s formal dialogue-fairness definition, this paper's equity definition covers arbitrary dialogue contexts and more general dataset biases, linking the analysis to learning theory.Source: §2.2 · pp. 3–4

3.1 Formal Definitions for Equity

• The protected attribute is a sensitive variable that may depend on dialogue context, and the framework extends to multiple categories by assessing each group’s parity gap.For example, race categories can be checked separately; the dataset meets the stated criterion when all measured gaps are zero.Source: §3.1 · pp. 4–5
• Score parity requires equal expected scores across protected-attribute groups, with the chosen score specifying which aspect of language must be equitable.The parity gap measures the difference between group-conditional expected scores, allowing degrees of inequity to be measured.Source: §3.1 · pp. 4–5
• With an attribute-independent score, score parity reflects demographic parity; a correctness-based score can instead express equal error rates across groups.The latter is relevant because language can be predictive of protected attributes, creating a trade-off between accuracy and invariance.Source: §3.1 · p. 5

3.2 Evaluating Equity with Learning Theory

• If test divergence to an equitable goal is at most ϵ/2, the learned dialogue’s parity gap is at most ϵ.The result uses the scoring function as the test; minimizing divergence with an equitable goal therefore supports equitable learned dialogue.Source: §3.2 · p. 6
• LEATHER-based algorithms for human-likeness can be adapted to equitable generation by using an equitable goal distribution and augmenting training data.The framework treats the goal distribution and test as key choices; human dialogue cannot simply be assumed equitable.Source: §3.2 · p. 6

Context and Its Role (Assumptions)

• The paper frames the central question as when human-like learning from possibly biased human dialogue can also learn equitable dialogue.Its assumptions characterize bias as inequity in the contextualized dialogue distribution, distinct from non-systemic noise quantified through expectations.Source: §3.3 · p. 6, §3.2 · p. 6, Context and Its Role (Assumptions) · p. 7
• Learning from biased human dialogue can also support equity when the data is context-aware and preserves context.The theorem connects human-like learning on such data to reducing an upper bound on divergence from an equitable goal; context-awareness assumes bias arises from aggregate attribute disproportions, while context-preservation keeps attribute-indicating context features consistent.Source: Context and Its Role (Assumptions) · p. 7
• When training samples are limited, augmentation can reduce data efficiency by introducing dependencies among examples and enlarging a learning-bound term.The passage warns that this can ultimately affect model performance, with the impact depending on the augmentation strategy.Source: Context and Its Role (Assumptions) · p. 7

4.2 Results

• LEATHER improves on CL across all reported human-likeness and equity metrics after 100 epochs of pre-training and 100 epochs of self-play.Table 1 compares the first four metrics for human-likeness and the last five for equity; no metric values are supplied here.Source: Table 1 · p. 8, §4.2 · p. 9

4.1 Dataset, Algorithms, and Evaluation

• The experiments use the GuessWhat?! corpus, model its question-player, and use an automated answer-player trained on human data.Gender is the protected attribute; the labeling scheme treats it as present when human dialogues use a gendered word.Source: §4.1 · p. 8
• CL combines supervised learning on human dialogue with a later self-play phase, while LEATHER re-incorporates human-human data during self-play.DS modifies LEATHER by downsampling to balance protected-attribute occurrence.Source: §4.1 · p. 8
• Evaluation measures human-likeness on full test data and equity through accuracy parity and human judgments of correct gendered-word use.The paper focuses on accuracy parity because it expects systemic dataset biases to affect accuracy parity more than toxicity or sentiment parity.Source: §4.1 · p. 8

4.2 Results

• DS improves on CL on all reported metrics except TD F, but performs worse than LEATHER on all metrics except ∆M.The authors relate DS’s weaker performance to the data inefficiency of its substantial data augmentation.Source: §4.2 · p. 9
• Theoretically, 2TD always bounds the parity gap ∆, and empirical results agree with this bound in every reported case.Test divergence also predicts all reported improvements from CL to LEATHER and may help rank algorithms by equity.Source: §4.2 · p. 9

• The paper formalizes mathematical notions of dialogue equity and uses computational learning theory to study how algorithm design can achieve it.Source: §5 · p. 9
• The study reports that algorithms can be designed for both equity and human-likeness, while theory can predict unexpected consequences of data augmentation.Source: §5 · p. 9
• The paper identifies proxy statistics that are useful for ranking algorithm equity in a task-oriented dialogue setting.Source: §5 · p. 9

• The empirical study primarily tests gender bias on the GuessWhat?! task, so broader empirical support across attributes and tasks remains needed.Source: Limitations · p. 9
• Applying insights from Theorem 3.2 to new data requires checking context-awareness and context-preservation assumptions, which may not hold as bias types change.Source: Limitations · pp. 9–10
• Only a smaller subset of gender annotations came from annotators primed to judge correctness and equity; the authors say more in-depth human evaluation could better support their results.Source: Limitations · p. 10

• Some proposed algorithms rely on pretrained word or image embeddings, whose associated harms or biases may remain after mitigation efforts.Source: Ethics Statement · p. 10
• The authors recommend rigorous human evaluation for bias before deploying models trained with these techniques.Source: Ethics Statement · p. 10
• The human-subject protocol was board-approved; participants volunteered and were compensated under board-approved regulations.Source: Ethics Statement · p. 10

A.1 Proof of Thm. 3.1

• Δ(Ĝθ) ≤ ε whenever TDG(θ) ≤ ε/2, as stated by the claim for an equitable goal and scoring function.The proof concludes that the equity gap is bounded by 2ε, the desired result.Source: §A.1 · p. 13
• Equity means the expected score is a shared value x for every protected attribute, which the proof substitutes into the expanded sum.The proof then splits the absolute-value expression into cases using the minimum and maximum expected scores.Source: §A.1 · p. 13
• In the two cases where expected scores straddle x, the x terms cancel; in the other cases, the expressions are bounded below by M − μ.Together these cases establish the stated bound on the equity gap.Source: §A.1 · p. 13

A.2.1 Proof

• With context-aware, context-preserving human data, the theorem bounds 2β × TDG(θ) with probability at least 1 − δ for all θ.The supplied statement does not include the expression that bounds it.Source: §A.2.1 · pp. 13–14
• The argument assumes context awareness, context preservation, conditional independence, and matching model and answer-player sampling distributions.It also assumes pairwise independence among the generated and human dialogue variables conditional on context.Source: §A.2.1 · pp. 13–14
• The proof transforms the model-data joint density into a human-data joint density using conditional independence and matching sampling distributions.The resulting density can compute TDH, whereas the original density computes TDG; both expectations use non-negative variables.Source: §A.2.1 · pp. 14–15
• A structural risk minimization PAC bound for a finite hypothesis space supplies the final generalization step without a test set.The proof applies the bound to the true test divergence and the observed divergence.Source: §A.2.1 · p. 15

A.2.2 Remarks on Data Efficiency

• For one protected attribute, ideal balancing has α = 2β and reduces data efficiency by a factor of 1/√2β; estimation and multiple attributes can add further costs.The formal comparison depends on the augmentation strategy and estimation error, and is left for future work.Source: §A.2.2 · p. 15

A.2.3 Assumption of Pairwise Independence

• Pairwise independence means that, conditional on context, the dialogue and attribute variables do not provide additional information about one another.The text gives gender and question-answer pairs conditional on the image as a GuessWhat?! example.Source: §A.2.3 · pp. 15–16

A.3 Labeling Scheme

• The labeling scheme is used to satisfy context awareness and preservation, supporting an equitable goal under score parity.The appendix establishes that the constructed distribution exists and that its test divergence and parity gap are zero.Source: §A.3 · pp. 16–17

A.4 Downsampling

• Downsampling balances occurrences of the protected-attribute labels, but the cooperative-learning dataset retains less than one-sixth of the original data.A separate balanced dataset is created for each protected attribute when training is no longer constrained by model-training inefficiency.Source: §A.4 · p. 17

A.5 GuessWhat?! Game Rules and Statistics

• GuessWhat?! has a question-player identify an image’s goal object by asking yes/no questions, while the answer-player knows both the image and object.After labeling and downsampling, the dataset has about 3,200 examples for F and 6,400 for M, each half labeled a = 1; the original M-to-F ratio is about 2:1.Source: §A.5 · pp. 17–18

A.6 Cooperative Learning

• Cooperative learning updates the question generator, object guesser, and shared encoder through self-play after pretraining on human-human dialogue.The encoder captures features shared by the other models, and LEATHER adds human data during training.Source: §A.6 · p. 18