Cross-Domain Review Generation for Aspect-Based Sentiment Analysis

This repository contains code for our ACL2021 Findings paper:

Cross-Domain Review Generation for Aspect-Based Sentiment Analysis

Datasets

The training data comes from four domains: Restaurant(R) 、 Laptop(L) 、 Service(S) 、 Devices(D).
For each domain transfer pairs, the unlabeled data come from a combination of training data from the two domains (ratio: 1:1).

The in-domain corpus(used for training BERT-E) come from yelp and amazon reviews.

Click here to get BERT-E (BERT-Extented) , and the extraction code is by0i. (Please specify the directory where BERT is stored in modelconfig.py.)

Folder

• aspect_output: aspects and opinions extracted by double propagation.
• ds-bert：training language models for target domains and using them for generating pseudo samples.
• pseudo_output: generated pseudo samples.
• ABSA: using pseudo samples for absa task.
• raw_data: traning data and testing data for four domains.

Usage

1. Using double propagation to extract aspects and opinions from target unlabeled data

2. Generating pseudo samples

2.1 Training DS-BERT

• To get target language model in ds-bert (bert_lm_models), run below code ：

bert-b-based:
bash ./ds-bert/run_mlm_bert_b.sh

bert-e-based:
bash ./ds-bert/run_mlm_bert_e.sh

2.2 Generation

• To generate pseudo samples in pseudo_output, run below code ：

bert-b-based:
bash ./ds-bert/generate_bert_b.sh

bert-e-based:
bash ./ds-bert/generate_bert_e.sh

3.Using pseudo samples for ABSA

• To get the results for ABSA by using pseudo samples, run below code ：

bert-b-based:
bash ./absa/run_base_bert_b.sh

bert-e-based:
bash ./absa/run_base_bert_e.sh