Update(21 Sept. 2018): I don't actively maintain this repository. This work was done for a course project and the dataset cannot be released because I don't own the copyright. However, everything in this repository can be easily modified to work with other datasets. I recommend reading the sloppily written project report for this project which can be found in docs/.
We use and compare various different methods for sentiment analysis on tweets (a binary classification problem). The training dataset is expected to be a csv file of type tweet_id,sentiment,tweet where the tweet_id is a unique integer identifying the tweet, sentiment is either 1 (positive) or 0 (negative), and tweet is the tweet enclosed in "". Similarly, the test dataset is a csv file of type tweet_id,tweet. Please note that csv headers are not expected and should be removed from the training and test datasets.
There are some general library requirements for the project and some which are specific to individual methods. The general requirements are as follows.
numpyscikit-learnscipynltk
The library requirements specific to some methods are:
keraswithTensorFlowbackend for Logistic Regression, MLP, RNN (LSTM), and CNN.xgboostfor XGBoost.
Note: It is recommended to use Anaconda distribution of Python.
- Run
preprocess.py <raw-csv-path>on both train and test data. This will generate a preprocessed version of the dataset. - Run
stats.py <preprocessed-csv-path>where<preprocessed-csv-path>is the path of csv generated frompreprocess.py. This gives general statistical information about the dataset and will two pickle files which are the frequency distribution of unigrams and bigrams in the training dataset.
After the above steps, you should have four files in total: <preprocessed-train-csv>, <preprocessed-test-csv>, <freqdist>, and <freqdist-bi> which are preprocessed train dataset, preprocessed test dataset, frequency distribution of unigrams and frequency distribution of bigrams respectively.
For all the methods that follow, change the values of TRAIN_PROCESSED_FILE, TEST_PROCESSED_FILE, FREQ_DIST_FILE, and BI_FREQ_DIST_FILE to your own paths in the respective files. Wherever applicable, values of USE_BIGRAMS and FEAT_TYPE can be changed to obtain results using different types of features as described in report.
- Run
baseline.py. WithTRAIN = Trueit will show the accuracy results on training dataset.
- Run
naivebayes.py. WithTRAIN = Trueit will show the accuracy results on 10% validation dataset.
- Run
logistic.pyto run logistic regression model OR runmaxent-nltk.py <>to run MaxEnt model of NLTK. WithTRAIN = Trueit will show the accuracy results on 10% validation dataset.
- Run
decisiontree.py. WithTRAIN = Trueit will show the accuracy results on 10% validation dataset.
- Run
randomforest.py. WithTRAIN = Trueit will show the accuracy results on 10% validation dataset.
- Run
xgboost.py. WithTRAIN = Trueit will show the accuracy results on 10% validation dataset.
- Run
svm.py. WithTRAIN = Trueit will show the accuracy results on 10% validation dataset.
- Run
neuralnet.py. Will validate using 10% data and save the best model tobest_mlp_model.h5.
- Run
lstm.py. Will validate using 10% data and save models for each epock in./models/. (Please make sure this directory exists before runninglstm.py).
- Run
cnn.py. This will run the 4-Conv-NN (4 conv layers neural network) model as described in the report. To run other versions of CNN, just comment or remove the lines where Conv layers are added. Will validate using 10% data and save models for each epoch in./models/. (Please make sure this directory exists before runningcnn.py).
- To extract penultimate layer features for the training dataset, run
extract-cnn-feats.py <saved-model>. This will generate 3 files,train-feats.npy,train-labels.txtandtest-feats.npy. - Run
cnn-feats-svm.pywhich uses files from the previous step to perform SVM classification on features extracted from CNN model. - Place all prediction CSV files for which you want to take majority vote in
./results/and runmajority-voting.py. This will generatemajority-voting.csv.
dataset/positive-words.txt: List of positive words.dataset/negative-words.txt: List of negative words.dataset/glove-seeds.txt: GloVe words vectors from StanfordNLP which match our dataset for seeding word embeddings.Plots.ipynb: IPython notebook used to generate plots present in report.