A Unified Model Using Distantly Supervised Data and Cross-Domain Data in NER.

Area of Science:

• Natural Language Processing
• Machine Learning
• Artificial Intelligence

Background:

• Supervised Named Entity Recognition (NER) models require extensive hand-annotated data, which is often scarce.
• Distantly supervised (DS) and cross-domain (CD) data are typically used separately to augment limited annotated datasets.
• DS data offers in-domain dictionary information, while CD data provides cross-domain insights, yet both present challenges like noise and distribution mismatch.

Purpose of the Study:

• To propose a unified model, PARE (PArtial learning and REinforcement learning), for simultaneously leveraging DS and CD data in NER.
• To address the noise inherent in DS data using partial learning with a novel label strategy.
• To mitigate the performance degradation caused by distribution mismatching in CD data via reinforcement learning.

Main Methods:

• The PARE model integrates partial learning with a new label strategy to effectively manage noisy DS data.
• Reinforcement learning is employed within PARE to address the distribution mismatching issues associated with CD data.
• The model is designed to utilize both DS and CD data concurrently as external resources.

Main Results:

• Experiments conducted on three datasets demonstrate that the PARE model significantly outperforms existing baseline models.
• The proposed PARE model shows robust performance improvements in NER tasks.
• The model's effectiveness was validated across diverse datasets, highlighting its generalizability.

Conclusions:

• The PARE model offers a unified approach to effectively utilize both distantly supervised and cross-domain data for Named Entity Recognition.
• PARE successfully handles noisy DS data and distribution mismatches in CD data, leading to superior performance.
• The model is particularly valuable in scenarios where limited or no hand-annotated in-domain data is available for NER system development.