eugenesiow / sa-transformers

Sentiment Analysis with Transformer Models.

Sentiment Analysis with Transformers

Classification Results

~4000 comments

Model Type	Model Name	Accuracy
XLNet	xlnet-base-cased	0.784
RoBERTa	roberta-base	0.802
RoBERTa	distilroberta-base	0.763
BERT	bert-base-cased	0.780
BERT	bert-base-uncased	0.680
BERT	bert-large-cased	0.772
FLAIR	LSTM	0.701
FLAIR	GRU	0.706
ALBERT	albert-base-v1	0.668
DistilBERT	distilbert-base-cased	0.742

~6000 comments

Model Type	Model Name	Accuracy
XLNet	xlnet-base-cased	0.822
RoBERTa	roberta-base	0.844

Hence, a pre-trained RoBERTa (base) transformer language model was used to train this classifier. Final model accuracy was 84.4%.

Docker

1. Build the image: docker build -t dso/sa-transformers:v1.0 ..
2. Run the container: docker run -d -p 5128:80 --name sa-transformers-container dso/sa-transformers:v1.0.
3. Export the image for distribution: docker image save -o sa-transformers-v1.0.tar dso/sa-transformers:v1.0.
   • For archiving, a gzip image is recommended: docker save dso/sa-transformers:v1.0 | gzip -c > sa-transformers-v1.0.tar.gz.
4. To load an image from a gzip archive: docker load < sa-transformers-v1.0.tar.gz.

RESTful API

• URL /sentiment

• Method POST

• URL Params None

• Data Params

  • Format: application/json

  Example Input

  A JSON array list of input text to analyse:

  ["Text to analyse for sentiment #1", 
  "Text to analyse for sentiment #1"]
  

• Success Response

  • Code: 200

  Example Results

  A JSON array list of sentiment outputs, order is that of the input array order. Negative: -1, Positive: 1, Neutral: 0.

  [0,
  -1]
  

• Invalid JSON Input

  • Code: 400

  Example Results

  {
      "ERROR": "Please provide a valid JSON. "
  }