Sentiment Analysis

The idea here is to extract a high level "Positive" or "Negative" sentiment from a given piece of text. This can be useful in predicting customer churn for instance by analysing email interactions of ongoing support queries. The project is still an experiment.

At the moment, I am using gensim's doc2vec library based on Mikolov and Le's paper[1]. I am training a doc2vec model using a training dataset and then using simple logistic regression to predict a 2-way coarse-grained positive/negative probability. I combined the test and train dataset of phrases from imdb and added cornell's sentiment polarity v2.0 dataset to the mix. The data was already curated from my source, but I removed stopwords using nltk. Still experimenting with different d2v models but at the moment it seems like pv-dbow gives good results for this dataset (dm=0). Negative sampling is turned off (hs=1) and this seems to make some difference. Perhaps removing the stopwords compromises the context I am not sure yet (but pv-dbow should ignore context words already?).

Install and Run

• Install Anaconda python3 from here if you don't already have it
• Fork or clone the repo git clone https://github.com/kramamur/sentiment-analysis.git
• Dependencies - conda install gensim nltk numpy scikit-learn (numpy and scikit come with conda bjic)
• Download nltk data - nltk.download('punkt') and nltk.download('stopwords')
• Train the model - python train.py
• Insert your favorite text into infer.txt- Yelp reviews give fairly accurate results
• Predict - python sentiment.py

---

Citations
1. Distributed Representations of Sentences and Documents
   tmikolov@,qlv@google.com, (https://arxiv.org/pdf/1405.4053v2.pdf)
2. Large Movie Review Dataset, amaas@cs.stanford.edu (http://ai.stanford.edu/~amaas/papers/wvSent_acl2011.bib)
3. Cornell sentiment polarity dataset, v2.0 (http://www.cs.cornell.edu/people/pabo/movie-review-data/)