Synt (pronounced: "cent") is a python library for sentiment classification on social text.
The end-goal is to have a simple library that "just works". It should have an easy barrier to entry and be thoroughly documented.
- Can collect negative/positive tweets from twitter and store it to a local database (can also fetch a pre-existing samples database)
- Can train a classifier based on a samples database
- Can classifiy text and output a score between -1 and 1. (where -1 is negative, +1 is positive and anything close to 0 can be considered neutral)
- abilitiy to collect, train, guess, and test (accuracy) from cli
- A running Redis server
- virtualenv (recommended)
- python2.7 (no support for
- sqlite3 issue was fixed in 2.7 issue
- PyYAML (pip install pyyaml)
- unfortnatley nltk requires pyyaml before it can be installed bug
Usage / Installation
Note: Many of these commands have additional arguments you can pass, use the -h flag to get help on any particular command and see more options.
Grab the latest synt:
pip install -e git+https://github.com/Tawlk/synt/#egg=synt
Grab the sample database to train on (or build one (below)):
Note: On your first run of any cli command a config will be copied into ~/.synt/config.py that you should configure. It uses sane defaults. This will only happen on the first run of synt.
synt fetch --db_name "mysamples.db"
By default it will be stored as 'samples.db'.
If you'd prefer to build a fresh sample db and have the time, just run collect with the desired amount.
synt collect --max_collect 10000 --db_name 'awesome.db'
Note: You can also increment samples in a database by providing the same db name.
A basic example of training
synt train 'samples.db' 20000
Train takes two required arguments: a training database (name), and the amount of samples to train on.
At this point you might want to see the classifiers accuracy on the training data.
Accuracy takes a number of testing samples. By default 25% of your training sample count will be used as the testing set. You can over-ride this by providing the --test_samples argument.
The database used for these testing samples will be the same as the database used to train. The testing samples will be new samples and can be guaranteed to be samples the classifier hasn't already seen.
You should now have a trained classifier and its time to see some classification of text.
This will drop you into a synt prompt where you can write text and see the score between -1 and 1.
You can alternativley also just classify text without having to drop into a prompt:
synt guess --text "i like ponies and rainbows"
We have acheived best accuracy using stopwords filtering with tweets collected on negwords.txt and poswords.txt (see downloads).
In the future we will also add the MaxEnt and Decision tree classifiers and the functionality to do clasiffier voting.
Note that this is optimized for classification on social text as this is our primary usecase. However, with a little tweaking it should be possible to get good results on other corpus'.
This code is still in production; use at your own risk. You may be eaten by a grue.
Questions/Comments? Please check us out on IRC via irc://irc.freenode.net/#tawlk