Please enable JavaScript.

Coggle requires JavaScript to display documents.

Sarcasm Detection in Plain Text Using Deep Learning with TensorFlow…

- - - - Obtained from www.thesarcasmdetector.com which had tweets collected over a period from June-July 2014. The dataset consists of positive sentences(sarcastic)[3] and negative sentences[2]. It consists of about 25,000 clean sarcastic tweets and about 1,00,000 clean non-sarcastic tweets
      - Tweets were collected from San Francisco and New York to ensure the tweets are in English
    - - A crucial and time-consuming step when it comes to NLP projects
      - As the input is in text form, which cannot be directly given to the neural network as an input hence we need to extract features from the sentence
      - The base project has extracted multiple features from a sentence like n-grams, polarity using SentiNet and TextBlob, and the topic of the sentence using gensim
      - Features
        
        N-Grams
        
        Unigrams and bigrams were used
        
        Sentiment Analysis
        
        Base project extracted sentiment as follows
        
        Get polarity and subjectivity of the entire sentence
        
        Divide the sentence into two equal parts and then calculate polarities and subjectivity
        for those.
        
        Divide the sentence into three equal parts and then calculate polarities and subjectivity
        for those
        
        Two libraries were used to extract sentiments namely, TextBlob and sentiNet
        
        Topic extraction
        
        In topic extraction most relevant information of the sentence is extracted
        
        Implemented using python library gensim
        
        Polarity
        
        In this project, instead of using multiple features to train the model, we used only one feature and that is polarity of the sentence
        
        We used TextBlob for extracting polarity and subjectivity
        as follows
        
        List of features being extracted.
        
        Full sentence Polarity
        
        Full sentence Subjectivity
        
        Half sentence Polarity (1/2 and 2/2)
        
        Half sentence Subjectivity (1/2 and 2/2)
        
        Difference between polarities of two halves
        
        Third sentence Polarity (1/3, 2/3 and 3/3)
        
        Third sentence Subjectivity (1/3, 2/3 and 3/3)
        
        Difference between max and min polarity of the thirds
        
        Fourth sentence Polarity (1/4, 2/4, 3/4 and 4/4)
        
        Fourth sentence Subjectivity (1/4, 2/4, 3/4 and 4/4)
        
        Difference between max and min polarities of the fourths.
    - - Python library which is used for performing natural language processing tasks
      - Performs Natural Language processing functions like noun phrase extraction, part of speech tagging, sentimental analysis, classification (naive Bayes, decision tree), n-gram, etc
    - - The training set consists of 70% of the feature set.The testing set consists of the rest 30% of the feature set
      - After building the model, it is trained with the extracted features
      - Training is done using the Adam Optimizer to minimize the mean cost throughout the epoch.
      - There are two methods to train a neural network model i.e. online and batch processing.
    - - Project used Tensorboard which is provided by TensorFlow to
        create a visualization of the data flow graph that was built for TensorFlow
    - - TensorFlow provides auto-differentiation that helps a lot while implementing gradient based machine learning
      - TensorFlow, as the name suggests, allows you to build data flow graphs to perform mathematical computations at every node of the graph
      - The project uses a deep neural network, developed using
        TensorFlow
      - It was used because of its efficiency and ease of implementation when it comes to deep neural networks
      - A model with three hidden layers and one output layer was built for this project[8]. Each hidden layer has 100 neurons, and output layer has two neurons, one for each class, namely Sarcastic and Regular