This is the last in a series of seven NLP articles. It presents the latest techniques and thought processes that are currently being researched for Question Answering Systems (QAS).
Abstract
QAS is based on the idea that the system would automate the process of answering a variety of questions, given the comprehension. The task involves an ensemble of ‘document processing’, ‘question processing’ and ‘answer processing’ techniques. The two biggest questions for researchers are ‘Can AI be used to answer a variety of questions? Can AI match (or even outperform) humans at this task?’
Developing QAS Models: The Challenges
- Diverse question sets require different approaches/strategies to answer.
- Whole-paragraph processing is necessary; some answers are spread across different sections of the paragraph.
- High-level (complete interpretation of the document) and low-level (interpretation of individual sentences) understanding is required.
- Handling different domains and different classes of questions (definition type, casual questions, confirmation question, etc.) is essential.
Where We Are Now: New Techniques vs. Improving Techniques
- After enormous efforts in this research area, we certainly don’t have a shortage of datasets (e.g. the SQuAD dataset has approximately 1 million question-answer pairs of around 500 articles).
- Modern deep learning architectures (like BERT, ALBERT (A Lite BERT) and attention models) have been able to achieve fairly high accuracy on some of the datasets.
- A model trained on the specific dataset still needs to have flexibility on different documents/paragraphs (e.g. the development of a QAS system to answer business questions from a news feed).
Maybe we can combine different sets of techniques to overcome the inefficiencies of each technique. But hybrid implementations are very computationally expensive and require some research time before deciding on any approach.
Food for Thought Experiment
Should we focus on improving the existing techniques? Or the development of new techniques/architectures for developing a generic QAS model? The computational complexity of NLP makes this question difficult to answer.
References
- https://stanford-qa.com
- https://thesai.org/Downloads/Volume8No9/Paper_31-Question_Answering_Systems_A_Review_on_Present_Developments.pdf
Technical articles are published from the Absolutdata Labs group, and hail from The Absolutdata Data Science Center of Excellence. These articles also appear in BrainWave, Absolutdata’s quarterly data science digest.
