B Precisions on Eye-tracking Metrics and Preprocessing

Table B.1: Eye-tracking mappings from dataset-specific fields to the shared set of metrics.
Metrics Dundee GECO ZuCo 1 & 2
First fix. dur. (FFD) First_fix_dur FIRST_FIXATION_DURATION FFD
First pass dur. (FPD) First_pass_dur GAZE_DURATION GD
Fix. prob. (FXP) Fix_prob ¬ WORD_SKIP FXC > 0
Fix. count (FXC) nFix FIXATION_COUNT FXC
Tot. fix. Dur. (TFD) Tot_fix_dur TOT_READ_TIME TRT
Tot. Regres. Dur. (TRD) Tot_regres_from_dur GO_PAST - SEL._GO_PAST GPT - GD

Univocal gaze metrics conversion Table B.1 present the conversion scheme used to obtain a unified set of eye-tracking metrics from different corpora annotations. This method follows closely the approach adopted by Hollenstein and Zhang (2019). While the mapping is straightforward for shared metrics, the TRD metric needs to be computed for GECO and ZuCo. For GECO, the difference between go-past time (i.e. total time elapsed between the first access of a word boundary and the first access of subsequent words, including regressions) and its selective variant (i.e. go-past time only relative to the specific word, without accounting for regressions) gives an exact conversion to regression duration. Instead, in the ZuCo case, an approximate conversion using gaze duration (i.e. first pass duration) instead of selective go-past time is used since selective go-past time is not provided. ZuCo’s TRD estimate should be deemed an upper bound for regressions’ duration since gaze duration is always smaller than the selective go-past time when regressions are present and is precisely equal to it in the complete absence of regressions.

Averaging across participants Gaze metrics are averaged across participants for all experiments of this thesis work. Metrics missing for some participants due to skipping are replaced with the lowest recorded value across participants for that word before averaging. This procedure is preferred to zero-filling missing values since the latter produces significant drops in metrics associated with tokens skipped by multiple participants, making averaged values inconsistent with empirical observations.

References

Hollenstein, Nora, and Ce Zhang. 2019. “Entity Recognition at First Sight: Improving NER with Eye Movement Information.” In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), 1–10. Minneapolis, Minnesota: Association for Computational Linguistics. https://doi.org/10.18653/v1/N19-1001.