Date of Award

Spring 2024

Document Type

Thesis

Degree Name

Master of Science (MS)

Department

Computer Science

Committee Chairperson

Richard Burns Ph.D.

Committee Member

Si Chen Ph.D.

Committee Member

Ashik Ahmed Bhuiyan Ph.D.

Abstract

Since its introduction in 2017, the transformer architecture has revolutionized natural language processing, leading to the development of large language models (LLMs). Encoder-decoder transformers excel at tasks requiring deep understanding, such as summarization and question-answering. Decoder-only variants have been optimized for generating coherent, extended texts.

Despite their success in deterministic tasks like translation, LLMs have been less explored in ambiguous tasks such as interpersonal conflict resolution. This study addresses this gap by evaluating LLMs on four new datasets derived from the "Am I the A**hole" (AITA) subreddit, featuring discussions of interpersonal conflicts. These datasets challenge models with real-world data including ambiguous judgments and toxic language.

This research utilizes Google's Flan-T5 and Meta's Llama-2-Chat to represent both architectures. Finetuned on the AITA datasets, these models were evaluated on their ability to classify and justify conflicts and their tendency to generate toxic language. Findings suggest that the most effective strategy involves finetuning an encoder-decoder LLM on a dataset cleaned of toxicity, followed by iterative refinement using Reinforcement Learning with Human Feedback (RLHF) to align with ethical standards.

To our knowledge, this is the first work examining the use of transformer-based LLMs for real-world interpersonal conflict resolution, offering insights for applications in social and therapeutic contexts where sensitive advice is crucial. It also contributes to discussions on the ethical implications of deploying AI in sensitive areas, suggesting ways it can complement human judgment.

Download

Share

COinS