Yiyao Li, Lu Wang, Jung Jae Kim, Chor Seng Tan, & Ye Luo. (2024). On the Selection of Positive and Negative Samples for Contrastive Math Word Problem Neural Solver. Proceedings of the 17th International Conference on Educational Data Mining, 96-106. https://doi.org/10.5281/zenodo.12729780
Abstract:
Math Word Problem (MWP) solving needs to discover the quantitative relationships over natural language narratives. Recent studies demonstrate that the contrastive learning framework can assist models in identifying semantically similar examples while distinguishing between different mathematical logics. This alleviates models’ dependence on shallow heuristics for predicting problem solutions. However, we have observed significant disparities in the positive and negative sample selections among different contrastive learning frameworks, sometimes even exhibiting complete contradictions. This discrepancy is attributed to the lack of an effective criterion for evaluating the distance between the mathematical logics of word problems. To address this issue, we introduce a novel formula for evaluating mathematical equation similarity in the context of word problem association. Our formula enables flexible focus on either global or local differences in the mathematical logic, thereby implementing distinct criteria for similarity calculation. We investigate the impact of various positive and negative sample selection strategies on contrastive learning models using the proposed formula, aiming to identify the optimal strategy. Experimental results highlight the effectiveness of our approach, revealing substantial performance improvements compared to existing baselines
License type:
Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0)
Funding Info:
This research / project is supported by the MOE - Science of Learning Grant
Grant Reference no. : MOE-MOESOL2021-0006