I am working on sentiment analysis and I am using dataset given in this link: http://www.cs.jhu.edu/~mdredze/datasets/sentiment/index2.html and I have divided my dataset into 50:50 ratio. 50% are used as test samples and 50% are used as train samples and the features extracted from train samples and perform classification using Weka classifier, but my predication accuracy is about 70-75%.
Can anybody suggest some other datasets which can help me to increase the result - I have used unigram, bigram and POStags as my features.
There are many sources to get sentiment analysis dataset:
- huge ngrams dataset from google storage.googleapis.com/books/ngrams/books/datasetsv2.html
- http://www.sananalytics.com/lab/twitter-sentiment/
- http://inclass.kaggle.com/c/si650winter11/data
- http://nlp.stanford.edu/sentiment/treebank.html
- or you can look into this global ML dataset repository: https://archive.ics.uci.edu/ml
Anyway, it does not mean it will help you to get a better accuracy for your current dataset because the corpus might be very different from your dataset. Apart from reducing the testing percentage vs training, you could: test other classifiers or fine tune all hyperparameters using semi-automated wrapper like CVParameterSelection or GridSearch, or even auto-weka if it fits.
It is quite rare to use 50/50, 80/20 is quite a commonly occurring ratio. A better practice is to use: 60% for training, 20% for cross validation, 20% for testing.
I started to gather sentiment analysis tools/datasets/lexicons in one place, it could be useful for you too: https://github.com/laugustyniak/awesome-sentiment-analysis
Start PR if you want to add something more or just write to me. I worked a lot with Amazon data [millions of reviews].
Here is a list of datasets that give the sentiments for individual words.. http://positivewordsresearch.com/sentiment-analysis-resources/
来源：https://stackoverflow.com/questions/24605702/good-dataset-for-sentiment-analysis