Question answer system for numerical reasoning in finance
Natural Language Processing (NLP) has seen rapid progress in the past few years resulting in different applications of it like machine translation, sentiment analysis, text summarization, question answering systems and so on. At the same time there has been a technological revolution in the finance...
Saved in:
Main Author: | |
---|---|
Other Authors: | |
Format: | Final Year Project |
Language: | English |
Published: |
Nanyang Technological University
2023
|
Subjects: | |
Online Access: | https://hdl.handle.net/10356/166057 |
Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
Institution: | Nanyang Technological University |
Language: | English |
Summary: | Natural Language Processing (NLP) has seen rapid progress in the past few years resulting in different applications of it like machine translation, sentiment analysis, text summarization, question answering systems and so on. At the same time there has been a technological revolution in the finance industry resulting in widespread use of different branches of AI, especially NLP. One of the applications used are the Question Answer Systems. These systems perform analysis on the passage or context provided based on the question asked and return the best possible answer to the user. They have been able to match human-like accuracy on reading comprehensions of multiple datasets. However, there are limits to this system which get exposed when numerical analysis and inference is needed like in financial documents.
In this paper, we first introduce and explain the idea of a Question Answer System. We then study and perform thorough analysis of existing models for this purpose. These include FinQA, Numnet, NAQAnet, TATQA and a few other models trained primarily on the Discrete Reasoning Over Paragraphs (DROP) dataset. We then reimplement these methodologies to understand the use of different hyperparameters. Finally, we choose to use an existing transformer called T5ForConditionalGeneration that is pre trained and will be finetuned for our purpose by training it on numerical analysis datasets like DROP. Finally, I conclude off by comparing my model with other models and performing experiments and offering insights for future development. |
---|