Sentiment Analysis on Movie Reviews
Sentiment analysis is a well-known task in the realm of natural language processing. Given a set of texts, the objective is to determine the polarity of that text. [9] provides a comprehensive survey of various methods, benchmarks, and resources of sentiment analysis and opinion mining. The sentiments can consist of different classes. In this study, we consider two cases: 1) A movie review is positive (+) or negative (-). This is similar to [2], where they also employ a novel similarity measure. In [10], authors perform sentiment analysis after summarizing the text. 2) A movie review is very negative (- -), somewhat negative (-), neutral (o), somewhat positive (+), or very positive (+ +). For the first case, we picked a Kaggle1 competition called “Bag of Words Meets Bags of Popcorn”. The challenge consists of two main parts. In the first part, we try a variety of basic sentiment analysis techniques. This provides a reasonable baseline to asses further complex methods. In the second part, we try different variants of the basic models. The objective of this part is to train a binary classifier for movie reviews (i.e., output classes are positive/negative). As in many natural language tasks, the first task here is to clean up, and convert the input texts (movie reviews) into numbers. This can be done using a variety of methods such as bag of words, word to vector, etc. Afterwards, we train the classifier. For the second case we used the data set used in [6]. Each example in this dataset (both training and test) is decomposed using a recursive tree. We use different variations of the recursive neural networks (RNN) (see [3, 4]) for the classification. Unlike the first case, here we have a multi-class (5 classes) problem.
Research Paper Link: Download Paper