Applied neural classification of eye-tracking features for unknown word recognition.
This repository contains an implementation of a feedforward neural network (MLP) for binary classification of eye-tracking data, targeting the detection of unknown word recognition. The work builds on and extends the study by Ryzhova, Škrjanec et al. (2023).
- Task: binary classification of unknown vs. known word recognition
- Input: engineered eye-tracking features
- Models explored: logistic regression, MLP, recurrent architectures (LSTM/CNN)
- Framework: PyTorch
A PyTorch-based MLP classifier outperformed the logistic regression baseline reported in the original study. While recurrent architectures were initially explored due to their ability to model sequential patterns, the simpler feedforward model achieved stronger and more stable performance on this dataset.
- The MLP exceeded the logistic regression baseline from the original paper.
- Recurrent models did not consistently outperform the feedforward baseline in this setting.
- This suggests that, for the given feature representation, model simplicity was advantageous.
Details, evaluation, and analysis are documented in the final report:
- Training and evaluation code for the MLP classifier
- Supporting scripts for data loading from CSV
- Jupyter notebooks for exploratory analysis and prototyping
- Reproducible full training runs via standalone Python scripts
Install dependencies:
pip install -r requirements.txt
Run a full training experiment:
python MLP.py
(See the report for details on tasks, features, and evaluation setup.)
Notes This work was developed as part of an academic seminar on eye tracking at Saarland University (Winter Semester 2023–2024). The repository reflects an applied modeling study rather than a production system.
Acknowledgements Thanks to Margarita Ryzhova and Iza Škrjanec for their original research, assistance with the dataset, and guidance on modeling approaches.
License MIT