ChiSquareX at CMCL 2021 Shared task: Leveraging recent advances in Pre-Trained Language Models for Eye-Tracking Prediction
This is our attempt of the shared task on Eye-Tracking Data Prediction: Predicting human reading patterns at the CMCL 2021 workshop, part of the NAACL 2021 conference.
The workshop provides a venue for work in computational psycholinguistics, including computational and mathematical modeling of linguistic representations, development and processing. CMCL promotes the synergy between the field of Cognitive Modelling and Natural Language Processing applying methods from computational linguistics to problems in the cognitive modeling of any and all natural language abilities by the use of Cognitively inspired human-derived behavioral data, which reflect the semantic representations in the human brain to augment the neural nets to solve a range of tasks spanning syntax and semantics with the aim of teaching machines about language processing mechanisms.
Official information, data, and task details can be found here.
Our paper can be found here.
Results of different models on the dataset can be found here: The results have been in terms of the R^2 (coefficient of determination) scores.
Model | Feature Norm | Extra Features | nFix | FFD | GPT | TRT | fixProp |
---|---|---|---|---|---|---|---|
Roberta | Std Scl | STT | 0.8842 | 0.9246 | 0.7343 | 0.8823 | 0.9509 |
Roberta | Min Max | All | 0.8835 | 0.9132 | 0.7383 | 0.8542 | 0.9461 |
Roberta | Min Max | STT | 0.8417 | 0.9090 | 0.6456 | 0.8152 | 0.9492 |
Bert | Min Max | All | 0.7433 | 0.8528 | 0.4574 | 0.6925 | 0.9173 |
BiLSTM Minibatch | Min Max | All+GloVe | 0.6595 | 0.7275 | 0.6281 | 0.5988 | 0.7583 |
BiLSTM Minibatch | Min Max | All+Numberbatch | 0.6379 | 0.7203 | 0.5992 | 0.5388 | 0.7592 |
Roberta Token | Std Scl | All | 0.7038 | 0.6512 | 0.4697 | 0.6877 | 0.7146 |
Bert Token | Std Scl | All | 0.7231 | 0.7107 | 0.3440 | 0.6376 | 0.7635 |
Bert | Std Scl | All | 0.6497 | 0.6456 | 0.5409 | 0.6088 | 0.7171 |
BiLSTM Minibatch | Std Scl | All+GloVe | 0.6850 | 0.5976 | 0.5267 | 0.5940 | 0.7158 |
Our Final Leaderboard Test MAE(Mean Absolute Error): 4.6764
A huge thanks to the organizers Emmanuele Chersoni, Nora Hollenstein, Cassandra Jacobs, Yohei Oseki, Laurent Prévot, and Enrico Santus for creating this wonderful task.
Additionally, we would like to extend a big thanks to the makers and maintainers of the excellent HuggingFace repository, without which most of our research would have been impossible.