sentiment-analysis

Sentiment Analysis of Tweets using RNN and LSTM.

This project implements and compares two deep learning models, a Simple Recurrent Neural Network (SimpleRNN) and a Long Short-Term Memory (LSTM) network, to classify the sentiment of tweets into three categories: positive, negative, or neutral.

The notebook is structured into two main parts:

• Model Training: Preprocesses the text data, builds, trains, and evaluates the RNN and LSTM models. The best-performing model (LSTM) is saved.
• Sentiment Prediction: Loads the saved LSTM model to perform sentiment prediction on new, user-provided text.

📊 Dataset

The model is trained on the Tweets_Data.csv dataset. This dataset contains tweets along with their corresponding sentiment.

• text: The full text of the tweet.
• sentiment: The target label (positive, negative, neutral).

⚙️ Methodology

The project follows a standard natural language processing (NLP) workflow:

Data Loading & Cleaning

The Tweets_Data.csv file is loaded into a pandas DataFrame.

Text Preprocessing

• Cleaning: Non-alphabetic characters are removed, and the text is converted to lowercase.
• Stop Word Removal: Common English stop words (e.g., "the", "a", "is") are removed using NLTK's stopwords corpus.
• Lemmatization: Words are reduced to their base or dictionary form (e.g., "going" -> "go") using NLTK's WordNetLemmatizer.

Feature Engineering

• Tokenization: The cleaned text is converted into sequences of integers using tensorflow.keras.preprocessing.text.Tokenizer, with a vocabulary size of 5000 words.
• Padding: All sequences are padded to a uniform length (max_len) to ensure consistent input size for the models.
• Label Encoding: The categorical sentiment labels ('negative', 'neutral', 'positive') are one-hot encoded into numerical format.

Model Building

Two sequential models are constructed using Keras:

• SimpleRNN Model: Embedding -> SimpleRNN -> Dropout -> Dense (output)
• LSTM Model: Embedding -> LSTM -> Dropout -> Dense (output)

Both models use categorical_crossentropy as the loss function and the adam optimizer.

Training & Evaluation

• The data is split into 80% for training and 20% for validation.
• The models are trained for 5 epochs with a batch size of 64. EarlyStopping is used to prevent overfitting.
• The performance is evaluated using a classification report and a confusion matrix.

📈 Results

The project compares the performance of the SimpleRNN and LSTM models. The LSTM model demonstrated superior performance and was chosen as the final model.

Model	Accuracy	Key F1-Scores
SimpleRNN	~65%	0.61 (negative), 0.62 (neutral), 0.71 (positive)
LSTM	~69%	0.66 (negative), 0.67 (neutral), 0.75 (positive)