IW3C2 2018 [arXiv]
Whats Unique This paper present an emperical system towards question answering by arranging a pipeline of tasks, and evaluating re-usable components for each tasks, and thus solving QA with high performance.
How It Works
- Following figure gives a really good illustration.
- As shown, there are three tasks needed to solve, and each task have different components implemented.
- Different questions can be answered by invoking different components at each task.
Source: Author
- Architecture is illustrated in the following figure
- For each questions, features like question length, question word, answer type, POS tags etc have been extracted, and each components performance was predicted based on supervised learning framework.
- That gives performance estimation for each components.
Source: Author
- System is evaluated on LC-QuAD and QALD-5 datasets, and improvements and relative performance of about 15% was improved.
