UnsupervisedABSA

We propose a BERT based three-step mixed semi-supervised model, which jointly detects aspect and sentiment in a given review sentence. The first step takes a small set of seed words for each aspect and each sentiment class to construct class vocabulary for each class using a context-aware BERT masked language model. The second step extracts aspect/opinion term(s) using POS tags and constructed vocabularies in step one. In the last step, extracted aspect and opinion words are used as label data to train a BERT based joint deep neural network for aspect and sentiment classification.

CASC

In this work, we leverage power of post-trained, domain knowledge BERT (DK-BERT) and present a simple and highly efficient semi-supervised hybrid CASC approach for Context aware Aspect category and Sentiment Classification. Our model is built in a simple three-step process:

We take a small set of seed words for each aspect and each sentiment class and then construct class vocabulary for each class that contains semantically coherent words with the seed words using BERT masked language model (MLM).
We take unlabeled training corpus and extract potential aspects and opinion terms using POS tags and class vocabularies constructed in the previous step.
In the last step, We make use of extracted aspect and opinion term as label data and jointly train BERT based neural model for aspect and sentiment classification.

Requirements

To install all dependencies:

pip install -r requirements.txt

Quick start

Run command:

python main.py

Datasets

Prepared datasets for both laptop and restaurant domain are available under datasets/ directory, acquired from Huang et al..

Configuration

All configuration and model hyperparameters can be found at config.py.

Configuring Domain

config = {
    'domain': 'laptop',
    'device': 'cpu'
}

The domain attribute determines which domain is used for training the model, which can be set to laptop or restaurant. Moreover, device can be set to cuda for training model on GPU.

Configuring data path

path_mapper = {
    'laptop': './datasets/laptop',
    'restaurant': './datasets/restaurant'
}

The path_mapper is responsible for providing root directory paths for each domain. The root directory should contain 2 files namely, train.txt (consisting of line separated reviews from the corpus) and test.txt (with serial number, aspect category, sentiment category and review sentence separated by tabs for each test example).

Providing seed words aspect_seed_mapper and sentiment_seed_mapper are used for providing seed words for aspect and sentiment classes for each domain.

Aspect seeds

aspect_seed_mapper = {
    'laptop': {
        'support': {"support", "service", "warranty", "coverage", "replace"},
        'os': {"os", "windows", "ios", "mac", "system", "linux"},
        'display': {"display", "screen", "led", "monitor", "resolution"},
        'battery': {"battery", "life", "charge", "last", "power"},
        'company': {"company", "product", "hp", "toshiba", "dell", "apple", "lenovo"},
        'mouse': {"mouse", "touch", "track", "button", "pad"},
        'software': {"software", "programs", "applications", "itunes", "photo"},
        'keyboard': {"keyboard", "key", "space", "type", "keys"}
    },
    'restaurant': {
        'food': {"food", "spicy", "sushi", "pizza", "taste", "delicious", "bland", "drinks", "flavourful"},
        'place': {"ambience", "atmosphere", "seating", "surroundings", "environment", "location", "decoration", "spacious", "comfortable", "place"},
        'service': {"tips", "manager", "waitress", "rude", "forgetful", "host", "server", "service", "quick", "staff"}
    }
}

Sentiment seeds