Conjunct Resolution in the Face of Verbal Omissions

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 13623–13640 July 9-14, 2023 ©2023 Association for Computational Linguistics Conjunct Resolution in the Face of Verbal Omissions Royi Rassin1 Yoav Goldberg1,2 Reut Tsarfaty1 1Bar-Ilan University 2Allen Institute for Artificial Intelligence {rassinroyi, yoav.goldberg, reut.tsarfaty} @gmail.com Abstract Verbal omissions are complex syntactic phe- nomena in VP coordination structures. They oc- cur when verbs and (some of) their arguments are omitted from subsequent clauses after being explicitly stated in an initial clause. Recovering these omitted elements is necessary for accu- rate interpretation of the sentence, and while humans easily and intuitively fill in the missing information, state-of-the-art models continue to struggle with this task. Previous work is limited to small-scale datasets, synthetic data creation methods, and to resolution methods in the dependency-graph level. In this work we propose a conjunct resolution task that operates directly on the text and makes use of a split-and- rephrase paradigm in order to recover the miss- ing elements in the coordination structure. To this end, we first formulate a pragmatic frame- work of verbal omissions which describes the different types of omissions, and develop an au- tomatic scalable collection method. Based on this method, we curate a large dataset, contain- ing over 10K examples of naturally-occurring verbal omissions with crowd-sourced annota- tions of the resolved conjuncts. We train vari- ous neural baselines for this task, and show that while our best method obtains decent perfor- mance, it leaves ample space for improvement. We propose our dataset, metrics and models as a starting point for future research on this topic. 1 Introduction Natural language is economic, and many elements in the message are not spelled out but omitted by speakers, left for the receiver of the message to complete. The hearer, either a human or an algo- rithm, should then complete the missing informa- tion and recover the intended meaning. This kind of omission and recovery occurs at all levels of conversation, from syntax to pragmatics, and is per- formed naturally and intuitively by humans, to the extent that they often don’t even realize that some- thing was missing in the message they received. Of that amount, about 75% was based on natural gas, 18% on oil, and 7% on hydroelectric power. Of that amount, about 75% was based on natural gas. Of that amount, about 18% was based on oil. Of that amount, about 7% was based on hydroelectric power. Figure 1: A demonstration of the conjunct resolution task. At the top, the input sentence with the omitted words in bold and the coordinating element is under- lined. Below, the expected rewrite, consisting of a single fully-specified standalone statement for each of the con- juncts. The rewrite must not include the coordinating element. An important class of omission phenomena — and the focus of this work — involves verbs and their arguments, especially around coordination structures. A verb and some of its arguments that appear in an initial clause may …

Content truncated to 500 words