LG-DimABSA: A Language-Grounded Approach to Dimensional Aspect-Based Sentiment Analysis

Setup Instructions

1. Install pixi from https://pixi.prefix.dev/latest/
2. Get english data from https://github.com/DimABSA/DimABSA2026/tree/main/task-dataset/track_a/subtask_3/eng and put it in /data/raw/eng

Model Weights

Push/pull models to/from HuggingFace. Set HF_TOKEN in .env for pushing (needs write token). Pulling is public.

pixi run hf-push
pixi run hf-pull

Laptop vs Restaurant

All training and evaluation scripts support --data {laptop,restaurant} to choose the dataset. Model checkpoints are saved to checkpoints/ with the data suffix (e.g., biaffine_laptop.pt, categorization_restaurant.pt).

Default is laptop. Switch to restaurant:

pixi run train-biaffine --data restaurant
pixi run eval-biaffine --data restaurant
pixi run train-categorization --data restaurant
pixi run eval-categorization --data restaurant
pixi run train-va-regression --data restaurant
pixi run eval-va-regression --data restaurant
pixi run eval-pipeline --data restaurant

Visualizing Data

pixi run marimo edit src/data/visualize.py --watch

and then open the localhost link it gives. this uses marimo as an ipynb alternative

Dataset Testing

pixi run test-dataset

visualizes one batch of size 4 from the simple dataloader.

Running Aspect Opinion Extraction

Candidate Extraction (spaCy rule-based)

pixi run eval-candidates

Pair matching

Biaffine Attention Matching

paper

pixi run train-biaffine
pixi run eval-biaffine

Categorisation

RoBERTa Classification

pixi run train-categorization
pixi run eval-categorization

VA Regression

VA Embeddings

pixi run train-va-regression
pixi run eval-va-regression

Full Pipeline

pixi run eval-pipeline