3.2 KiB
ANLP_WS24_CA2
This repository contains the necessary scripts, data, and notebooks for analyzing and modeling the linguistic and structural features of humor in jokes. The project focuses on leveraging NLP techniques to analyze humor in text data, and aims to predict the humor score numerically using regression models.
Objektive
Leverage advanced NLP techniques (CNN, BERT, and Transformer) to analyze text data and build an application that predicts humor ratings.
Research Question
- Can Deep neural networks predict humor ratings with an RMSE greater than or equal to the baseline of 0.8609 ?
Data Source
The data is sourced from the SemEval-2021 Task 7: It contains a dataset of humor and offense ratings for jokes. The jokes are annotated with a humor rating on a scale from 0 to 4.
- Traindata: HaHackathon.https://homepages.inf.ed.ac.uk/s1573290/data.html -> associated paper: https://aclanthology.org/2021.semeval-1.9.pdf#:~:text=HaHackathon%20is%20the%20first%20shared%20task%20to%20combine,its%20average%20ratings%20for%20both%20humor%20and%20offense
- Testdata: Since no test data was available, the traindata also was used as test data and divided into test, train and validation data
Data embeddings
- gloVe 6B tokens: https://nlp.stanford.edu/projects/glove/
Preprocessing Steps
1. Load and clean data: The data set is loaded and all rows with missing humor_rating values are removed. In addition, the target variable for the humor rating is extracted.
2. text embeddings: Pre-trained GloVe embeddings are loaded and converted into a matrix that can be used for modeling.
3. data splitting: The data set is split into training, test and validation data to train and evaluate the models later.
4. ensemble data indices: Various methods for creating data indices are provided to prepare the training data for ensemble methods.
Repository Structure
data/: Contains the datasethack.csv, which includes raw joke data.- Notebooks:
- Used for data analysis and visualization.
- represent the models
Getting Started
Install Requirements
Run the following command to install the required dependencies:
pip install -r requirements.txt
Preprocess Data
This is carried out automatically when the models are executed
Workflow
1. Preprocessing
The text data is cleaned and transformed into formats suitable for analysis. The preprocessing steps include tokenization, stopword removal, and lemmatization.
2. Model Training
Various machine learning models, including Convolutional Neural Networks (CNNs), Long Short-Term Memory Networks (LSTMs), BERT, and Transformers, are trained to predict the humor rating of jokes based on their linguistic features.
3. Model Evaluation
The trained models are evaluated to determine their performance in predicting humor ratings. Metrics such as RNSE (Root Mean Squared Error) and R² scores are used to assess the models.
4. Classification and Regression
While the primary goal of the project is to predict the numerical humor rating (regression task), we also experiment with classification models for humor detection (e.g., humor vs. non-humor)