Denoising Neural Relation Extraction for Spatio-Temporal Recommendation System
Ye Wang, Lihong Guo, Yang Yu, Yuan Gao · IEEE Transactions on Big Data · 2024
The Point-of-Interest (POI) recommendation system in location-based social networks is pivotal, offering versatile applications. Personalized recommendations hinge on pre-established user relationships, emphasizing geographic proximity and character similarities. To attain spatial and temporal information, natural language text mining, particularly relation extraction tasks, emerges as crucial, drawing from the abundant textual data on the Internet. However, challenges arise with traditional supervised methods relying on human-annotated data and distant supervision introducing noises, which prickly and seriously degrade the effectiveness of the sentence encoding for valid relation mentions. To address this issue, we aim to rectify the sentence encoder by acquiring external knowledge from an existing fully supervised dataset. More specifically, we craft transfer learning, together with adversarial training, to discover the semantic consistency in the valid relation mentions. During training, the sentence encoder receives data from both domains successively and generates a consistent output for them. While the discriminator distinguishes, as far as possible, which domain these sentences come from. After the above training procedure, the encoder can be made insensitive to the noise in distantly supervised data for denoising purposes. The experimental results on the widely used dataset NYT10 demonstrate that our model outperforms current state-of-the-art methods for relation extraction.