Discovering Biased News Articles Leveraging Multiple Human Annotations

Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pages 120–125 Marseille, 11–16 May 2020 c⃝ European Language Resources Association (ELRA), licensed under CC-BY-NC 120 The Alice Datasets: fMRI & EEG Observations of Natural Language Comprehension Shohini Bhattasalip, Jonathan R. Brennanm, Wen-Ming Luh△, Berta Franzluebbers†, John T. Hale† pUniversity of Maryland, College Park, MD, USA mUniversity of Michigan, Ann Arbor, MI, USA △National Institutes of Health, Bethesda, MD, USA †University of Georgia, Athens, GA shohini@umd.edu, jobrenn@umich.edu, wenming.luh@nih.gov, {berta, jthale}@uga.edu Abstract The Alice Datasets combine observations from magnetic resonance imaging as well as electrophysiology while human participants listened to the same literary narrative in English. Along with these neural signals and the text of the story, we also provide a variety of word-by-word predictors motivated by research in computational linguistics and cognitive science. These predictors range from prosody to morphology to syntax. These annotated, naturalistic datasets can be used to replicate prior work and test new hypotheses about natural language comprehension in the brain. Keywords: fMRI, EEG, neurolinguistics, naturalistic, English, cognitive methods, corpus 1. Introduction While there has been a tradition of using naturalistic lan- guage corpora in computational linguistics, this practice has only recently become popular within the cognitive neu- roscience of language. In this neighboring field “naturalis- tic” stimuli are now offering a new way to study language comprehension in the brain, in synergy with natural lan- guage processing (NLP) tools (Armeni et al., 2017; Bren- nan, 2016). From a neurolinguistic perspective, contextually-situated naturalistic stimuli present an opportunity to investigate multiple linguistic levels, including phonemes, syllables, words, phrases, sentences and discourse, all simultaneously within the same dataset. Such an opportunity creates new avenues for linking hypotheses between various linguis- tic representations and neurobiological architectures in the brain (Kandylaki and Bornkessel-Schlesewsky, 2019). This paper introduces a new language resource, the Al- ice Datasets, which is naturalistic in this sense (Maguire, 2012; Willems, 2015; Hamilton and Huth, 2018). It pack- ages together neuroimaging and electrophysiological mea- sures from many speakers listening to the same text – the first chapter of Lewis Carroll’s Alice’s Adventure in Won- derland (1865). It shares an underlying book with the cross-linguistic project, Alice in the Language Localizer Wonderland (Fedorenko, 2016) and an emphasis on nat- uralistic text with the Narrative Brain Dataset (Lopopolo et al., 2018) (hereafter NBD). However the current project is based on a longer continuous narrative than in either of these projects (12 mins 24 secs vs 6 mins 30 secs on an av- erage for NBD). This larger corpus facilitates the study of linguistic phenomena that may not be attested sufficiently often in smaller samples. The Alice Datasets include both high spatial resolution (fMRI) and high temporal resolution (EEG) neural signals. And since this resource is in English (rather than Dutch, the language of the NBD), it offers the chance to apply a very wide array of speech and natural language technologies. Together, these properties make the Alice Datasets a first-line choice for those seeking to com- pare AI …

Content truncated to 500 words