Sentiment Analysis and Text Generation using Many-to-One LSTMs on Airline Reviews

Overview

This project explores sentiment analysis and text generation utilizing Long Short-Term Memory (LSTM) neural networks.

---

Sentiment Analysis

We use LSTMs to analyze airline sentiments, aiming to predict sentiment labels (0 or 1) based on customer reviews. Converting text reviews into numerical data, we utilize many-to-one LSTMs for accurate predictions.

---

Text Generation

Text generation is also explored by training LSTMs on "Alice's Adventures in Wonderland." The model learns to predict the next word in a sequence, creating coherent and contextually relevant sentences. Challenges like language variability and context dependence are addressed through entropy scaling and softmax temperature techniques.

---

Aim

This project's objectives are twofold:

1. Develop a sentiment detection model using many-to-one LSTMs to predict sentiment labels (0 or 1) based on airline text reviews.
2. Utilize many-to-one LSTMs for text generation, training on "Alice's Adventures in Wonderland" and predicting the next word in a sequence.

---

Data Description

• airline_sentiment.csv: This dataset includes information on airline sentiments in CSV format, containing "airline_sentiment" (sentiment labels: 0 or 1) and "text" (customer reviews).

• alice.txt: Project Gutenberg's eBook of "Alice’s Adventures in Wonderland" by Lewis Carroll, provided as a text file. This dataset serves as training data for text generation using many-to-one LSTMs.

---

Tech Stack

• Language: Python
• Libraries: pandas, numpy, keras, tensorflow, collections, nltk

---

Approach

Sentiment Analysis

Dataset

• Obtain the airline sentiment dataset with sentiment labels and text reviews.

Preprocessing

• Perform data preprocessing, including text cleaning, tokenization, and stop word removal.
• Convert text reviews into a bag-of-words representation.

Many-to-One LSTM

• Utilize many-to-one LSTM architecture for sentiment detection.
• Feed the bag-of-words representation as input to the LSTM.

Training

• Split the dataset into training and testing sets.
• Train the LSTM model using the training set.
• Evaluate model performance on the testing set.

Text Generation

Dataset

• Obtain "Alice's Adventures in Wonderland" text dataset.

Preparing and Structuring

• Preprocess the text data by cleaning, tokenizing, and structuring sentences and phrases.
• Create sequences.

Many-to-One LSTM

• Implement many-to-one LSTM architecture for text generation.
• Train the LSTM model using the prepared dataset.

Challenges in Text Generation

• Recognize challenges such as language variability and context dependence in text generation.
• Understand that natural language uses diverse words with similar meanings, requiring careful consideration during generation.

Entropy Scaling

• Explore entropy scaling to introduce controlled randomness into text generation.

Softmax Temperature

• Introduce softmax temperature as a hyperparameter to control prediction randomness in LSTMs and neural networks.
• Predict using temperature.

---

Modular Code Overview

• data: Contains input data files.
• lib: Includes a Jupyter Notebook with code and documentation.
• output: Stores generated files and results.
• Readme.md: Provides project information.
• requirements.txt: Lists required packages.
• Engine.py: The main engine file.
• ML_Pipeline: Contains modules for different steps in the machine learning pipeline.

---