An Innovative Similar Complaint Recommendation Model Integrating Semantic and Graph Embeddings

Artificial intelligence plays a vital role in driving the intelligentization of the complaint hotline 12345. To assist hotline operators in accurately understanding public demands and allocating them to municipal departments, we propose BERT-Graph that recommends historical complaint to pending complaint through a dual-tower recommendation model integrating semantic and graph embeddings. Since the complaint records are sparse, colloquial and similar complaint texts may transcribed semantically different by operators, we construct an event interact graph for the complaint text to avoid implying sole reliance on semantic information. We consider each complaint text as a node, the co-occurrence times of key words calculated by TF-IDF between texts as weighted edges, and we use the Graph Attention Network (GAT) to obtain graph embeddings. After obtaining the semantic embedding through pre-trained language model BERT, we concat it with the graph embedding and feed it to the dual-tower model. We add a self-attention module as the neural network in the dual-tower model and we use a fully connected layer to output the similarity score of the input texts. Experimental results on complaint hotline 12345 datasets of the Chinese city of Wuhan demonstrate that compared to other baseline models, BERT-Graph shows a significant improvement on the Mean Squared Error (MSE) metric, from 0.86 to 0.24.