Natural Language Inference

Project type: Team

Team details

Mentor: Ehsaan Elhamifar

Project mate: Poorna Chandra Vemula

Introduction

Natural Language Inference (NLI) is the task of determining the relationship between two statements; More precisely truth value of “hypothesis” statement considering that “premise” statement is true/holding true.
There are only 3 possible relations, which are commonly referred to as entailment, contradiction and neutral.

• ‘Entailment’ stands for the situation when we can conclude that hypothesis is true based on the premise.
• ‘Contradiction’ is the situation when we can conclude that the hypothesis is false based on the premise.
• Finally, when we cannot definitely conclude whether hypothesis is true or false based on the premise as it does not provide enough information, it is considered as ‘neutral’.

NLI is a valuable testing ground for the development of semantic representations as it summarizes the hidden meaning between lines.

Overview

This is a team project focussed on exploring the archaic and modern ML algorithms including large language models (LLM) performance for the task of natural language inference (NLI). In this project, for the archaic ML exploration, we have considered the Logistic Regression, Naive Bayes and for the modern ML exploration we have considered LSTMs, Bi-LSTMs, and fine-tuned the transformer ‘BERT’. In the process we have engineered few features to find the overlapping of words between the statements as below, which we have improved the performance of archaic ML models.

The results are as follows:

Results