Research

Investigating Entity Resolution in Text-to-SQL Models

April 2024
PDF available here.

This project explores using SQL views generated by object-relational mapping (ORM) as input to a Text-to-SQL model. This novel technique is intended to ease entity resolution, the key to a model’s understanding of how database objects can be described in natural language. I constructed a software pipeline to automatically generate views to represent the entities in a given database and generate corresponding training data to be fed into an encoder-decoder model. My undergraduate thesis provides an in-depth explanation of leveraging object-relational mapping as the cornerstone of creating the views. While experimental results were inconclusive, I discuss factors that may have influenced the model’s performance and future directions for the technique.

Leveraging Object-Relational Mapping to Reduce Entity Impedance In Natural Language Interfaces for Relational Databases

April 2022
PDF available here.

This thesis demonstrates the use of an object-relational mapping (ORM) framework to reduce entity impedance in a natural language interface for relational databases (NLI-DB). Specifically, we present a database-agnostic SQL view construction algorithm that partitions an arbitrary database into entities that maximize performance of a natural language understanding (NLU) model. This model represents an important contribution to automating the creation of a natural language interface for a single database in two ways. First, it eliminates ambiguity when mapping a user’s information need to a specific database table. Second, it allows for the use of simple natural language models for generating SQL queries to service a given information need. Practically, our approach aims to increase database access for non-technical users by minimizing their need for SQL knowledge. We accomplish this by mapping a user’s information need to a specific set of database columns exposed by a specific entity view.