D-pFoodReQ: Answer-Frequency-Aware Debiasing for Constrained Knowledge-Base Question Answering in Food Recommendation

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Food recommendation systems must satisfy dietary preferences and nutritional, health, and ingredient constraints. In constrained food knowledge-base question answering, unequal positive training-answer frequency may influence ranking without necessarily reflecting query–recipe fit. We develop D-pFoodReQ by replacing the pFoodReQ answer ranker with BAMnet-D while retaining its food-knowledge and constraint-processing pipeline. BAMnet-D introduces a separate answer-frequency branch alongside semantic matching, applies candidate-set-relative loss reweighting, and uses a counterfactual-inspired intervention that sets the explicit frequency input to zero during validation and testing. Across three runs using the same protocol on the pFoodReQ benchmark, D-pFoodReQ achieved an F1 score of 62.96 ± 2.06%, compared with 61.33 ± 2.27% for pFoodReQ. Its mean average precision and mean average recall were 66.50 ± 0.54% and 65.38 ± 0.66%, respectively. On 1355 questions for which no gold answer appeared as a positive training label, D-pFoodReQ attained a mean answer-set F1 nearly identical to that of pFoodReQ while achieving higher Recall@5 and NDCG@5. Ablation and output-composition analyses indicated that loss reweighting alone did not account for the full improvement, while retaining the explicit frequency term was associated with a larger share of answers observed as positive training labels. These results support answer-frequency-aware ranking for constrained food knowledge-base question answering.