Bayesian Q-learning in multi-objective reward model for homophobic and transphobic text classification in low-resource languages: A hypothesis testing framework in multi-objective setting
Bibliographic record
Abstract
Most Reinforcement Learning (RL) algorithms optimize a single-objective function, whereas real-world decision-making involves multiple aspects. For hate comment classification, an agent must balance maximizing the F1-score while minimizing False Positives (FP) to enhance precision and reduce misclassifications. However, such multi-objective optimization introduces uncertainties in decision-making. To address this, we propose a Bayesian Q-Learning framework with a convolutional neural network policy. The policy outputs action logits, integrated with Q-value estimates sampled via Thompson Sampling from a Gaussian posterior. Our reward function combines F1-score (objective 1) and a penalty for misclassification (objective 2) to optimize learning. To validate our framework, firstly we show that our framework classifies the hate-comments comparatively better than other baselines by scoring an F1-score of 83%, 93%, 77% and 71% in English-Tamil, English, Kannada and Malayalam datasets for detecting homophobic and transphobic comments respectively. Secondly, we demonstrate that the variance of Q-value estimates in our Bayesian posterior decreases significantly over time, indicating that the agent has learned an optimal policy that effectively balances the competing objectives. This finding is further supported by statistical t-tests conducted across all datasets, which confirm the significance of the observed variance reduction. Additionally, we observe our agent’s multi-objective optimization path in 3D space, which shows its ability to balance reward (F1-score) and regret. Furthermore, we compare the action selection between our Bayesian approach and non-Bayesian action clustering using K-Means algorithms, where our analysis highlights coherent clustering which indicates structure exploration, while non-Bayesian approach shows premature convergence to suboptimal policies.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.006 | 0.016 |
| Meta-epidemiology (narrow) | 0.001 | 0.001 |
| Meta-epidemiology (broad) | 0.003 | 0.001 |
| Bibliometrics | 0.001 | 0.001 |
| Science and technology studies | 0.001 | 0.002 |
| Scholarly communication | 0.002 | 0.002 |
| Open science | 0.003 | 0.002 |
| Research integrity | 0.003 | 0.003 |
| Insufficient payload (model declined to judge) | 0.004 | 0.001 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".