Abstract:
The dual-agent multi-hop reasoning method addresses the issue of target entity omission in single-agent systems caused by increased reasoning path length. However, this m...Show MoreMetadata
Abstract:
The dual-agent multi-hop reasoning method addresses the issue of target entity omission in single-agent systems caused by increased reasoning path length. However, this method still has the following challenges: First, even if the agent finds the correct answer within the knowledge graph, it may misclassify it as incorrect due to training set limitations, known as the false negative target problem. Second, agents may be misled by erroneous paths, leading to the creation of false positive paths. To tackle these issues, this paper proposes a dual-agent multi-hop reasoning method based on Reward Shaping and Action Dropout (RADAM). This approach improves the model in two ways: (1) by introducing an embedding-based single-hop reasoning model, TransE, to optimize the reward function and reduce false negatives; and (2) by incorporating a random masking mechanism to diminish the agent's sensitivity to spurious paths, thereby reducing false positive paths. Experimental results demonstrate that RADAM achieves more accurate and efficient answer retrieval across most benchmark datasets compared to baseline algorithms, with ablation experiments further confirming the effectiveness and synergy of reward shaping and action dropout.
Published in: 2024 5th International Conference on Machine Learning and Computer Application (ICMLCA)
Date of Conference: 18-20 October 2024
Date Added to IEEE Xplore: 21 November 2024
ISBN Information: