Causalqa: a causal framework for question answering
Neural networks have proven their success in various fundamental applications such as object detection, image segmentation, image and text generation and several NLP tasks. That said, neural networks are black-box function approximators with good approximation capability described by the universal a...
Saved in:
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Final Year Project |
| Language: | English |
| Published: |
Nanyang Technological University
2022
|
| Subjects: | |
| Online Access: | https://hdl.handle.net/10356/156616 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Institution: | Nanyang Technological University |
| Language: | English |
| Summary: | Neural networks have proven their success in various fundamental applications such as object detection, image segmentation, image and text generation and several NLP tasks. That said, neural networks are black-box function approximators with good approximation capability described by the universal approximation theorem. This black box nature prevents the utilisation of neural networks in high risk areas such as healthcare. This brings a need for explainable AI. The paradigm used to explore these possibilities is causality.
In this work, we introduce a novel question answering algorithm dubbed CausalQA which learns several subtasks such as causal structure learning. Furthermore, we introduce an interventional training paradigm based on previous theoretical works including recent theoretical works on linking graph neural networks to Structural Causal Models. We show proof of concept by evaluating it on a toy dataset and further evaluating it on question answering datasets. We achieve comparable performance to state-of-the-art models and empirically prove the ability. |
|---|