Factoid Question Answering with Distant Supervision.

Area of Science:

• Artificial Intelligence
• Natural Language Processing

Background:

• Deep learning has advanced automatic question answering (QA).
• Training deep neural networks for QA requires extensive annotated data, which is laborious for new domains or languages.

Purpose of the Study:

• To propose a distantly supervised method for automatically generating question-answering pairs.
• To improve the quality and relevance of generated questions by incorporating user interests and expression styles from community QA data.
• To address the data annotation bottleneck in training QA models.

Main Methods:

• A distantly supervised method was developed to automatically generate question-answering pairs.
• Community QA data was explored to mine diverse question paraphrases and estimate question asking probabilities.
• Generated data was used to train QA models, with and without limited annotated data.

Main Results:

• A QA model trained solely on generated data achieved 49.34% accuracy on real-world questions.
• Incorporating generated data significantly improved performance when limited annotated training data was available.
• An improvement of 1.35 absolute points was observed on the WebQA dataset.

Conclusions:

• Distantly supervised methods can effectively generate QA data for training deep learning models.
• Leveraging community QA data enhances the relevance and diversity of synthetic training data.
• The proposed approach offers a viable solution for data-scarce domains in question answering.