Learn NLP with Me
I am kickstarting “Learn NLP with Me” blog series where I document my learning of new NLP concepts. The blog posts will either be my notes or me trying to explain what I have learned. Through this series, I am hoping to track all the new NLP concepts I have learned throughout the year as well as having digital copies of all my NLP notes, not to mention it will help me substantially in learning these new concepts. I will also list out all the sources I have used to learn new concepts.
Machine Reading Comprehension (MRC) Introduction & Brief History
A sub-area within NLP that assess machine’s understanding of natural languages by measuring its ability to answer questions based on a given context. Research work on MRC systems date back to 1970s but it didn’t really attract attentions until 2015. In the 1990s, most methods for solving MRC tasks are either rule-based or machine-learning-based which has few main limitations. These limitations include substantial human effort in crafting rules and features, incapable of generalisation when dealing with datasets with many different types of articles, and failure in extracting contextual information and fixing the long-term dependencies issue.
These limitations made early MRC systems less practical and therefore they weren’t widely adopted. However, things change in 2015 for two main reasons:
- Deep learning techniques has shown good results in capturing contextual information and has outperformed traditional systems
- The development of standard large-scale datasets means that a) deep learning is made possible, and b) there’s a standardised benchmark dataset to evaluate and improve MRC systems
The attention on MRC has grown exponentially since 2015 and the types of MRC tasks has become more diverse.