TIER: Temporal Convolutional Network Information Extractor With Conditional Random Field

As data scales expand, effectively modeling long-distance dependencies has become crucial for information extraction. We propose TIER, which combines the Temporal Convolutional Network (TCN) and the Conditional Random Field (CRF). The dilated convolution structure of TCN is used to expand the Receptive Fields (RFs) to capture long-distance context features, and the global dependency between labels is modeled through the CRF layer to improve the information extraction performance further. We conducted experimental evaluations on three information extraction tasks, and the results show that our model outperforms SOTA models on four datasets. Specifically, on the MSRA and DUIE 2.0 datasets, F1-scores reached 94.97% and 93.64%, improving by 4.82% and 4.34% compared to the average performance of SOTA models. On the CLUENER dataset, an improvement of 4.12% is achieved compared to SOTA models, demonstrating the superiority of our approach. Additionally, we conducted low-resource experiments, and under 30%-70% of the MSRA and CLUENER training data, F1-scores remained at relatively high levels, verifying that our method maintains good adaptability and generality in low-resource scenarios.