AI-Sentiment-Analysis with Tensorflow and Keras

1. Setup Project
2. Project Describtion
3. Code Describtion

# 1. Setup the project and try if your sentence is sarcastic or not
1. Clone the repository and run it with visual studio code.
2. In the code file change variable "json_datei" to link to the json file "sarcasm.json".
json_datei = r'YOUR_LINK\sarcasm.json' . This is the dataset of 26.000 headlines.
3. Run the script.
4. When the plots are poping up you have to close them. Then the code will go on and show outputs in the console.
5. Scroll down to the last lines in the code. There is the sentence variable. Type in a sentence to find out if the trained model thinks it is sarcastic or not. The output values in the console are between 0.1, 0.2... to 1. The number one says it is sarcastic and the number zero says it is not sarcastic.

2. Project Description

This project utilizes TensorFlow and Keras to build a neural network for detecting sarcasm in headlines. The dataset, containing headlines and corresponding sarcasm labels, is loaded and preprocessed using tokenization and padding techniques. A neural network model is constructed with an embedding layer followed by global average pooling and dense layers. The model is trained on the dataset to achieve accurate sarcasm detection and its performance is visualized through accuracy and loss graphs. Additionally, word embeddings are extracted and visualized for further analysis.

3. Short description of the code. The code file is commented as well:

The code begins by importing necessary libraries such as JSON for handling JSON data and tensorflow for building and training neural networks. The script defines parameters such as vocabulary size, embedding dimension and maximum sequence length. It then loads a dataset of 26.000 headlines and their sarcasm labels from a JSON file, splits the data into training and testing sets and tokenizes the text data while converting it into padded sequences for input into the neural network. The neural network model is built using Keras's Sequential API. It consists of an embedding layer to convert words into dense vectors, followed by a global average pooling layer to reduce dimensionality. The model is compiled with binary cross entropy loss and the Adam optimizer. The model is trained on the training data for 30 epochs with validation performed on the testing data. The training history, including accuracy and loss metrics, is stored for visualization.

The image is showing the accuracy for the training and the test network. The accuracy of the test network is 0.9622 with 20.000 training sentences. The model has 6500 sentences for testing and the accuracy for that is 0.807. The model isn’t trained with the sentences for testing! The script includes a function to plot accuracy and loss graphs based on the training history. Here you can see the output of the graphs:

The trained model's embedding layer is used to extract word embeddings. Then they are saved to files for visualization in a separate tool: TensorFlow Embedding Projector. Finally, the model is used to predict sarcasm in new sentences provided in the script. The sentences are tokenized, padded and fed into the model for prediction.