nustm / dalm

Cross-Domain Data Augmentation with Domain-Adaptive Language Modeling for Aspect-Based Sentiment Analysis

The code for our ACL2023 paper (https://aclanthology.org/2023.acl-long.81/)

Jianfei Yu, Qiankun Zhao, Rui Xia. "Cross-Domain Data Augmentation with Domain-Adaptive Language Modeling for Aspect-Based Sentiment Analysis"

Datasets

The training data comes from four domains: Restaurant(R) 、 Laptop(L) 、 Service(S) 、 Devices(D).

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.)

Usage

1. Domain-Adaptive Pseudo Labeling

To assign pseudo labels to unlabeled data in the target domain, run below code:

bash pseudo_label.sh

2. Domain-Adaptive Language Modeling

Train a domain-adaptive language model, generate target-domain labeled data, and finally use the generated data for the main tasks. We use LSTM and GPT2 as decoder in language modeling respectively.

2.1 To train the GPT2-based DALM for data generation and evaluation, run below code:

bash GPT2.sh

2.2 To train the LSTM-based DALM for data generation and evaluation, run below code:

bash LSTM.sh

Acknowledgements

• Some code in LSTM-based language modeling are based on the codes of DAGA, many thanks!

dalm's Issues

Questions about the quantity of generating data

Thanks for your outstanding Contribution, your data augmentation method is very novel! After carefully reading your papaer, there's one quesiton that would like to consult:

Is the quantity of unlabeled dats used for generation pseudo label always one-to-one ? Is it feasible that generating multiple pseudo data given one sample?

请补充一下requirements.txt

可以写一个requirements吗？想复现，但完全不知道需要什么版本的什么库。