Linguistically Aware Model Routing for Financial Sentiment Analysis

Spring 2026 CSCI 5541 NLP: Class Project - University of Minnesota

TrustTheTokens

Small Model Results

1. Baseline DistilBERT (Zero-Shot) — PhraseBank

Baseline DistilBERT — PhraseBank confusion matrix Baseline DistilBERT — PhraseBank per-class report

2. Baseline FinBERT (Zero-Shot) — PhraseBank

Baseline FinBERT — PhraseBank confusion matrix Baseline FinBERT — PhraseBank per-class report

3. Fine-Tuned DistilBERT — PhraseBank (Test)

Fine-tuned DistilBERT — PhraseBank test confusion matrix Fine-tuned DistilBERT — PhraseBank test per-class report

4. Baseline TinyBERT (Zero-Shot) — Combined Dataset

Baseline TinyBert — Combined confusion matrix Baseline TinyBert — Combined per-class report

5. Baseline FinBERT (Zero-Shot) — Combined Dataset

Baseline FinBERT — Combined confusion matrix Baseline FinBERT — Combined per-class report

6. Fine-Tuned TinyBERT — Combined Dataset (Validation)

Fine-tuned TinyBERT — Combined validation confusion matrix Fine-tuned TinyBERT — Combined validation per-class report

7. Fine-Tuned TinyBERT — Combined Dataset (Test)

Fine-tuned TinyBERT — Combined test confusion matrix Fine-tuned TinyBERT — Combined test per-class report