Abstract
When people do the reading comprehension, they often try to find the words from the passages which are similar to the question words first. Then people deduce the answer based on the context around these similar words. Therefore, the position information may be helpful in finding the answer rapidly and is useful for reading comprehension. However, previous attention-based machine reading comprehension models typically focus on the interaction between the question and the context representation without considering the position information. In this paper, we introduce the position information to machine reading comprehension and investigate the performance of the position information. The position information is experimented in three different ways: 1) position encoder; 2) attention mechanism; and 3) position mapping embedding. By experimenting on TriviaQA dataset, we have demonstrated the effectiveness of position information.
Highlights
As a task of automatic question answering, machine reading comprehension is usually defined as a task to answer a corresponding question based on the some natural language documents
We experiment on TriviaQA dataset, and the results prove that the method of using prior hypothesis and attention mechanism is an appropriate application of position information in machine reading comprehension tasks, which is more effective t han other direct characterization methods
As we describe in previous, the gathered method of TriviaQA makes the dataset closer to the machine reading comprehension task which people usually deal with, we mainly evaluate the model on TriviaQA
Summary
As a task of automatic question answering, machine reading comprehension is usually defined as a task to answer a corresponding question based on the some natural language documents. In this excerpt, the answer ‘‘Cliff Thorburn’’ is found surrounded the question words ‘‘snooker player’’ and ‘‘The Grinder’’ (‘‘nickname’’ is a key word if we associate it with question word ‘‘known as’’).
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have