TY - GEN

T1 - A Cross-lingual Comparison of Human and Model Relative Word Importance

AU - Morger, Felix

AU - Brandl, Stephanie

AU - Beinborn, Lisa

AU - Hollenstein, Nora

PY - 2022

Y1 - 2022

N2 - Relative word importance is a key metric for natural language processing. In this work, we compare human and model relative word importance to investigate if pretrained neural language models focus on the same words as humans cross-lingually. We perform an extensive study using several importance metrics (gradient-based saliency and attention-based) in monolingual and multilingual models, including eye-tracking corpora from four languages (German, Dutch, English, and Russian). We find that gradient-based saliency, first-layer attention, and attention flow correlate strongly with human eye-tracking data across all four languages. We further analyze the role of word length and word frequency in determining relative importance and find that it strongly correlates with length and frequency, however, the mechanisms behind these non-linear relations remain elusive. We obtain a cross-lingual approximation of the similarity between human and computational language processing and insights into the usability of several importance metrics.

AB - Relative word importance is a key metric for natural language processing. In this work, we compare human and model relative word importance to investigate if pretrained neural language models focus on the same words as humans cross-lingually. We perform an extensive study using several importance metrics (gradient-based saliency and attention-based) in monolingual and multilingual models, including eye-tracking corpora from four languages (German, Dutch, English, and Russian). We find that gradient-based saliency, first-layer attention, and attention flow correlate strongly with human eye-tracking data across all four languages. We further analyze the role of word length and word frequency in determining relative importance and find that it strongly correlates with length and frequency, however, the mechanisms behind these non-linear relations remain elusive. We obtain a cross-lingual approximation of the similarity between human and computational language processing and insights into the usability of several importance metrics.

M3 - Article in proceedings

SP - 11

EP - 23

BT - Proceedings of the 2022 CLASP Conference on (Dis)embodiment

PB - Association for Computational Linguistics (ACL)

T2 - 2022 CLASP Conference on (Dis)embodiment

Y2 - 15 September 2022 through 16 September 2022

ER -