Automatically learning human rationales for machine learning

Previous research has shown that enriching labels for classification tasks such as sentiment analysis with human annotators’ “rationales” can produce substantial improvements in classification performance. Moreover, in low-resource settings these rationales can be used to learn attention [1]. Human rationales can be annotated manually or can be learned automatically [2]. We want to combine these two methods by learning the rationales directly from human brain activity data [3].

This MSc thesis project includes developing a method to extract passive human rationales from eye-tracking and electroencephalography (EEG) data and building a machine learning system for tasks such as sentiment analysis which evaluates the quality of the rationales.

Supervised by Nora Hollenstein


  1. [1] Bao, Y. et al. "Deriving Machine Attention from Human Rationales." Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing. 2018.
  2. [2] Yessenalina, A., Y. Choi, and C. Cardie. "Automatically generating annotator rationales to improve sentiment classification." Proceedings of the ACL 2010 Conference Short Papers. Association for Computational Linguistics, 2010.
  3. [3] Hollenstein N., Rotsztejn J., Troendle M., Pedroni A., Zhang C. and Langer N. “ZuCo, a simultaneous EEG and eye-tracking resource for natural sentence reading.” Scientific Data. 2018.