In our research we performed Data Augmentation on the HAABSA++ model of M. Trusca. After forking her work as a basis (https://github.com/mtrusca/HAABSA_PLUS_PLUS), we used the EDA from T. Liesting (https://github.com/tomasLiesting/HAABSADA) and added our own BERT based augmentation methods. In order to perform the research yourself, you can follow the instructions as given by M. Trusca and run the augmentation on the data files. For our research, we use the fourth method of the LCR-ROT model and within this LCR-ROT model the BERT embeddings are being used.

HAABSA++

The code for A Hybrid Approach for Aspect-Based Sentiment Analysis Using Contextual Word Emmbeddings and Hierarchical Attention

The hybrid approach for aspect-based sentiment analysis (HAABSA) is a two-step method that classifies target sentiments using a domain sentiment ontology and a Multi-Hop LCR-Rot model as backup.

• HAABSA paper: https://personal.eur.nl/frasincar/papers/ESWC2019/eswc2019.pdf

Keeping the ontology, we optimise the embedding layer of the backup neural network with context-dependent word embeddings and integrate hierarchical attention in the model's architecture (HAABSA++).

Software

The HAABSA source code: https://github.com/ofwallaart/HAABSA needs to be installed. Then the following changes need to be done:

• Update files: config.py, att_layer.py, main.py, main_cross.py and main_hyper.py.
• Add files:
  • Context-dependent word embeddings:
    • getBERTusingColab.py (extract the BERT word embeddings);
    • prepareBERT.py (prepare the final BERT emebdding matrix, training and tesing datasets);
    • prepareELMo.py (extract the ELMo word emebddings and prepare the final ELMo embedding matrix, training and testing datasets);
    • raw_data2015.txt, raw_data2016.txt (Data folder).
  • Hierarchical Attention:
    • lcrModelAlt_hierarchical_v1 (first method);
    • lcrModelAlt_hierarchical_v2 (second method);
    • lcrModelAlt_hierarchical_v3 (third method);
    • lcrModelAlt_hierarchical_v4 (fourth method).

The training and testing datasets are in the Data folder for SemEval 2015 and SemEval 2016. The files are available for Glove, ELMo and BERT word emebddings.

*Even if the model is trained with contextul word emebddings, the ontology has to run on a dataset special designed for the non-contextual case.

Word embeddings

• GloVe word embeddings (SemEval 2015): https://drive.google.com/file/d/14Gn-gkZDuTVSOFRPNqJeQABQxu-bZ5Tu/view?usp=sharing
• GloVe word embeddings (SemEval 2016): https://drive.google.com/file/d/1UUUrlF_RuzQYIw_Jk_T40IyIs-fy7W92/view?usp=sharing
• ELMo word embeddings (SemEval 2015): https://drive.google.com/file/d/1GfHKLmbiBEkATkeNmJq7CyXGo61aoY2l/view?usp=sharing
• ELMo word embeddings (SemEval 2016): https://drive.google.com/file/d/1OT_1p55LNc4vxc0IZksSj2PmFraUIlRD/view?usp=sharing
• BERT word embeddings (SemEval 2015): https://drive.google.com/file/d/1-P1LjDfwPhlt3UZhFIcdLQyEHFuorokx/view?usp=sharing
• BERT word embeddings (SemEval 2016): https://drive.google.com/file/d/1eOc0pgbjGA-JVIx4jdA3m1xeYaf0xsx2/view?usp=sharing

Download pre-trained word emebddings:

• GloVe: https://nlp.stanford.edu/projects/glove/
• Word2vec: https://code.google.com/archive/p/word2vec/
• FastText: https://fasttext.cc/docs/en/english-vectors.html