Two-way deep RNN for sentiment classification.
Switch branches/tags
Nothing to show
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
assets
cmd/sentiment
doc
.gitignore
.travis.yml
CODE_OF_CONDUCT.md
CONTRIBUTING.md
LICENSE
README.md
english.json
inference.go
model.pb
model_512_256_0.8696.pb
train_model.py

README.md

BiDiSentiment GoDoc Build Status

Text sentiment analyser for Go and Python.

Runs on top of Tensorflow. The architecture is a bidirectional two-level character LSTM.

Happy and sad gophers

Installation of the inference module

go get gopkg.in/vmarkovtsev/BiDiSentiment.v1/...

Usage of the inference module

For every line in stdin, the command line tool prints the probability of the negative sentiment (the probability of the positive sentiment is 1 - negative).

echo "This is the worst movie I have ever seen, it sucks balls!" | sentiment 
0.9961139
echo "This is the best movie I have ever seen, just love it!" | sentiment
0.000769752

API expects string batches for performance reasons:

import "gopkg.in/vmarkovtsev/BiDiSentiment.v1"

func main() {
  session, _ := sentiment.OpenSession()
  defer session.Close()
  result, _ := sentiment.Evaluate(
    []string{"This is the best movie I have ever seen, simply love it!"},
    session)
  println(result[0])
}

Science

arch

We scan through the text by byte in both directions, and the length is constrained to 180 bytes. The training is written in Python and is based on Keras and Tensorflow. The achieved accuracy on 20% validation is 87%. The train dataset was 1.5 million tweets. The default parameters were used. The model easily overfits, so only the first 5 epochs were used.

validation

Update the model in the Go inference application by executing

go-bindata  -nomemcopy -nometadata -pkg assets -o assets/bindata.go  model.pb english.json

Contributins

...are welcome! See CONTRIBUTING.md and CODE_OF_CONDUCT.md.

License

MIT, see LICENSE.