This repository contains the data and code for the paper "A Challenge Dataset and Effective Models for Aspect-Based Sentiment Analysis", EMNLP-IJCNLP 2019, [paper].

MAMS is a challenge dataset for aspect-based sentiment analysis (ABSA), in which each sentences contain at least two aspects with different sentiment polarities. MAMS dataset contains two versions: one for aspect-term sentiment analysis (ATSA) and one for aspect-category sentiment analysis (ACSA).

pytorch==1.1.0
spacy==2.1.8
pytorch-pretrained-bert==0.6.2
adabound==0.0.5
pyyaml==5.1.2
numpy==1.17.2
scikit-learn==0.21.3
scipy==1.3.1

Put the pretrained GloVe(https://nlp.stanford.edu/data/wordvecs/glove.840B.300d.zip) file glove.840B.300d.txt in folder ./data. Modify config.py to select task, model and hyper-parameters. When mode is set to term, base_path should point to an ATSA dataset. When mode is set to category, base_path should point to an ACSA dataset.

python preprocess.py

python train.py

python test.py

The BERT model pretrained by huggingface(https://github.com/huggingface/pytorch-transformers) is used in our experiments.