Research

Problem. General-purpose GPT models generate fluent text but have no explicit mechanism for targeting a specific emotional tone, which limits their usefulness for applications like therapeutic writing aids, character dialogue, or emotionally-calibrated marketing copy.

Approach. Built an emotionally-aware generation pipeline (ECGText) that integrates GPT models with a sentiment and emotion classification stage, so generated text can be steered toward a target emotion rather than left to the base model's default tone.

Evaluation. Assessed output quality along four axes: cosine similarity (semantic fidelity to the prompt), winnowing (near-duplicate detection to check for degenerate repetition), part-of-speech distribution analysis (grammatical naturalness), and sentence structure statistics (syntactic diversity) — combining semantic, structural, and stylistic checks rather than relying on a single metric.

Problem. Extends the ECGText line of work by asking whether emotional control can be made more precise and consistent than a single classification stage allows.

Approach. Combines an LLM generator with RoBERTa, fine-tuned on the Go-Emotions dataset, to score candidate generations against a target emotion label. An iterative optimization pipeline regenerates and re-scores text until it converges on output that matches the intended emotion, rather than accepting the first generation.

Result. The iterative, RoBERTa-scored approach achieved higher emotional fidelity — how closely the output matches the intended emotion — and better linguistic coherence than single-pass generation baselines.