An Improved Query Similarity Model for Online Health Community Forum Using Cross-Attention Mechanism on Siamese Network

Nowadays, many users join online health community forums to seek information, acquire knowledge, and hear how patients handle similar illnesses. However, finding similar patients is time consuming and laborious. In this circumstance, this study provides an automated query similarity method to find similar queries already discussed. Siamese neural network is a proven technique to assess the semantic similarity between sentences. However, Siamese neural network cannot alone provide satisfactory results due to long-term dependencies among words between sentences. Hence, this paper is an improved system of a previous study that proposed a Siamese neural network with a transfer learning technique for medical query similarity tasks. To examine the semantic similarity of the queries, a cross-attention method is constructed using a Siamese neural network (CASNN). The cross-attention mechanism manages the interactive semantic relationship between words. To handle the presence of numerous medical terms in the query, the model is also fine-tuned on a large medical question–answer dataset using a transfer learning approach. The best model (CASNN), when compared to other variations of the model, could obtain an F1-score of 89%. The experimental results show that the proposed method effectively obtains a better generalization ability among medical queries than the previous method.