imro832000/sentiment-analysis-of-tweets

Classifying the tweets into positive, negative or neutral

SENTIMENT ANALYSIS OF TWEETS

INTRODUCTION

In recent times people have started using social networking platforms like twitter..etc. This has generated huge amounts of structured and unstructured data. Processing these huge amounts of data can make us understand the intentions of the user which as a result can help in optimizing user experience which plays a key role. This data requires analysis due to the need to easily and accurately label sentiment classes on a large scale (huge data). Sentiment analysis is the automated process of analyzing text data and sorting it into sentiments positive, negative, or neutral. Using sentiment analysis tools to analyze opinions in Twitter data can help companies understand how people are talking about their brand. Appropriate natural language processing techniques are required to accomplish the goals. Methods that can analyze noisy data are required as the data is on a large scale we may not have it in a structured manner. In the marketing field, companies use it to develop their strategies, to understand customers' feelings, how people respond to their campaigns or product launches. In the political field, it is used to keep track of political views, to detect consistency and inconsistency between statements and actions at the government level. It can be used to predict election results as well. Sentiment Analysis also is used to monitor and analyze social phenomena, for the spotting of potentially dangerous situations and determining the general mood of the blogosphere. Sentiment analysis is also called opinion mining. Opinion mining involves analyzing opinions, sentiments, or mentality of the author from the transcription. Online opinions have a great and direct influence on the business of many e-commerce sites. Before buying any product on any online platform it is a general tendency for a user to look at the review of the product and then buy it according to the review given. This way sentiment analysis is very important in the industry.

APPROACH:

Sentiment analysis can be performed with different algorithms. They can be performed using machine learning and deep learning algorithms. Machine learning algorithms are good to a certain extent whereas deep learning techniques can dig deep inside the data and understand the patterns from the data. We read in the data set and analyse the data by preprocessing techniques. We only consider the “sentiment” and “tweet_text” column and drop the tweet_id column as it is not necessary. The first thing we do after this is removing the stopwords from the tweet_text column. Generally these words do not contribute and have no value in computing the sentiment of a tweet. After removing the stopwords we also remove the mentions in the text. In order for our deep learning model to understand the text data we convert into numbers. To use the text as input for our model, we first need to convert the words into integers that refer to an index in a dictionary. Only the most frequent words are kept. We clean the text by applying filters and putting the words to lowercase. Words are separated by spaces. Once the dictionary has been created we can convert the text to a list of integer indexes. This is done with the text_to_sequences method of the Tokenizer. These integers are now converted into one hot encoding features which are provided to the model. We also need to convert our target classes to numbers in order for the model to understand. We perform OneHotEncoding to the target labels with the to_categorical() method in keras. We use labelEncoder() function in sklearn to convert the ‘positive’, ‘negative’, ‘neutral’ to their numerical representation. After converting to numbers we apply the to_categorical() method in keras and convert it into onehot- encoded form. After our data is ready we use the validation set to evaluate the model performance when we tune the parameters of the model. We start with a model with 2 densely connected layers of 64 hidden elements. The input_shape for the first layer is equal to the number of words we allowed in the dictionary and for which we created one-hot-encoded features. As we need to predict 3 different sentiment classes, the last layer has 3 hidden elements. The softmax activation function makes sure the three probabilities sum up to 1. In the first layer we need to estimate 640064 weights. This is determined by (nb inputs * nb hidden elements) + nb bias terms, or (10000 x 64) + 64 = 640064 In the second layer we estimate (64 x 64) + 64 = 4160 weights In the last layer we estimate (64 x 3) + 3 = 195 weights. Because this project is a multi-class, single-label prediction, we use categorical_crossentropy as the loss function and softmax as the final activation function. We run for a predetermined number of epochs and will see when the model starts to overfit. Now, we can try to do something about the overfitting. There are different ways to do that.

1] Reduce the network's size by removing layers or reducing the number of hidden elements in the layers.

2] Add regularization, which comes down to adding a cost to the loss function for large weights.

3] Adding dropout layers, which will randomly remove certain features by setting them to zero.

1. Reducing the network's size: We reduce the network's size by removing one layer and lowering the number of hidden elements in the remaining layer to 32. We can see that it takes more epochs before the reduced model starts overfitting (around epoch 10). Moreover, the loss increases much slower after that epoch compared to the baseline model.

2. Adding Regularization: To address overfitting, we can also add regularization to our model. For the regularized model we notice that it starts overfitting earlier than the baseline model. However, the loss increases much slower afterwards.

3. Adding drop out layers: The model with dropout layers starts overfitting a bit later than the baseline model. The loss also increases slower than the baseline model. From the results obtained by training and testing our model, the model with the dropout layers performs the best on the test data.

MODEL ARCHITECTURE

Kaggle Score

Future Work

Working on different approches to improving the Accuracy...