Modal Dependency Parsing via Language Model Priming

Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2913 - 2919 July 10-15, 2022 ©2022 Association for Computational Linguistics Modal Dependency Parsing via Language Model Priming Jiarui Yao1, Nianwen Xue1, and Bonan Min2 1Brandeis University 2Raytheon BBN Technologies {jryao, xuen}@brandeis.edu, bonan.min@raytheon.com Abstract The task of modal dependency parsing aims to parse a text into its modal dependency struc- ture, which is a representation for the factual- ity of events in the text. We design a modal dependency parser that is based on priming pre-trained language models, and evaluate the parser on two data sets. Compared to baselines, we show an improvement of 2.6% in F-score for English and 4.6% for Chinese. To the best of our knowledge, this is also the first work on Chinese modal dependency parsing. 1 Introduction Modal dependency parsing (MDP) is the task of parsing a text into a modal dependency structure (MDS) (Vigus et al., 2019) in which each event in the text is linked to a conceiver, the informa- tion source of the event. An MDS is a graph in which the nodes are events and conceivers, and the edges represent the level of certainty that a conceiver holds with respect to the event. An ex- ample MDS is presented in Figure 1, where Jeroen Weimar is the conceiver of the event travelled, and is certain that the traveling event has happened, as indicated by the edge label Pos. Vigus et al. (2019) define 6 categories of modal strength, or levels of certainty, and they are full positive (Pos), partial positive (Prt), positive neutral (Neut), negative neu- tral (Neutneg), partial negative (Prtneg) and full negative (Neg). The root node of an MDS is always the author (AUTHOR) of a document, the ultimate source of all information sources mentioned in the text. Modal dependency parsing is thus the task of taking a text as input and parsing it into a modal dependency structure. MDP departs from previous approaches to event factuality prediction that cast it as an event classification (e.g. Saurí and Puste- jovsky (2012)) or regression (e.g. Lee et al. (2015)) problem aimed at just predicting the level of cer- tainty of an event. The level of certainty alone is insufficient in judging the factuality of an event, and knowing the information source (conceiver) is also crucially important. For example, in Figure 1, our judgment of whether the event travelled has happened also crucially depends on the credibility of the information source, Jeroen Weimar, in addi- tion to the level of certainty the information source holds towards the event. ROOT AUTHOR tested positive Jeroen Weimar confirmed travelled MODAL Pos Pos Pos Pos Figure 1: A modal dependency tree for “A person in Traralgon had tested positive to COVID-19 on Sunday. The Victorian government’s COVID-19 response com- mander Jeroen Weimar confirmed ‘this individual has travelled to Melbourne.’ ” Yao et al. (2021) develop the first modal depen- dency parser by first separately extracting …

Content truncated to 500 words