Quick overview

ASC Tagset

Learned correspondences between form and function vary in terms of complexity and abstraction (e.g., Fillmore, Kay, & O’Connor, 1998; Goldberg, 2013), and so do the examples of the ASCs. There is currently no definitive set of ASCs that should be included in an ASC tagset. The current study drew on a range of previous literature (e.g., Biber et al., 1999; Goldberg, 1995; 2006; Hwang & Kim, 2022) and selected nine ASC types as a tagset. These are briefly summarized below with frequently occuring syntactic (Kyle, 2016; Kyle & Crossley, 2017; Kyle et al., 2021) and semantic (Fillmore, 1968; Palmer et al., 2005; Palmer, Gildea, & Xue, 2010) frames.

ASC tag	Syntactic frame (example)	Semantic frame (example)
ATTR	Subj-Vcopular-NP/AdjP/PP	Xtheme is Yattribute
INTRAN_S	Subj-V	Xagent acts
INTRAN_MOT	Subj-V-Adverbial Particle/PP	Xtheme moves Ypath
INTRAN_RES	Subj-V-NP/AdjP	Xpatient becomes Yresult
TRAN_S	Subj-V-Obj	Xagent acts on Ypatient
DITRAN	Subj-V-Objindirect-Objdirect	Xagent causes Yrecipient to receive Ztheme
CAUS_MOT	Subj-V-Obj-PP	Xagent causes Ytheme to move Zpath/goal
TRAN_RES	Subj-V-Obj-NP/AdjP	Xagent causes Ytheme to become Zstate
PASSIVE	Subj-auxVpast participle (-by-PP)	Xtheme undergo V (by Yagent)

Dataset

We developed an ASC treebank by building on the English portion of the Universal Propositions (UP) project (Akbik et al., 2015). The project represents a merge of the Universal Dependencies version of the English Web Treebank (EWT; Bies et al., 2012; Silveira et al., 2014) and PropBank (Palmer et al., 2005). The EWT corpus includes sentences from five web registers including blogs, newsgroups, emails, reviews, and Yahoo! Answers.

Creating an ASC treebank

1. For each sentence in the training section of EWT, we extracted the large-grained argument structures using the default PropBank semantic role labels (e.g., ARG0-Verbsense-ARG1).

2. We converted the large-grained arguments to fine-grained semantic role frames (e.g., agent-V-theme), using relation mappings from PropBank frame files (Palmer et al., 2005), which also draws on information in FrameNet (Fillmore et al., 2003) and VerbNet (Schuler, 2005).

3. Based on a discussion of ASC categorization between the authors that included co-annotation of 100 sentences, we manually assigned an ASC to each semantic role frame that occured at least times in the corpus (n = 355). For example, the semantic role frame theme-Verb-attribute was annotated as an attributive construction and agent-Verb-theme was annotated as a transitive simple construction.

4. In some cases, the corpus analysis indicated that particular semantic role frames could represent multiple ASCs. This most often occurred in cases where a fine-grained semantic role for a particular argument of a particular verb was unavailable in PropBank, leading to an underspecified semantic role frame. In these cases, the use of each ‘semantic role frame + verb’ combination that occurred at least twice in the treebank was checked and each was assigned an ASC.

5. Particularly ambiguous cases were resolved through discussions between the authors.

6. As a final step, we conducted spot checks which led to a small number of corrections. This approach resulted in the categorization of 94.1% of the ASCs in the treebank. Any sentences that included uncategorized ASCs were omitted from further analysis.

7. We also evaluated the quality of the semi-automated annotation process (for more information, see Kyle & Sung (2023)).

ASC representation in Treebank

ASC	Most frequent verbs	Total Freq	Train	Dev	Test
TRAN_S	have, do, say	12,431	9,965	1,213	1,253
ATTR	be, seem, look	6,004	4,723	648	633
INTRAN_S	go, work, come	2,754	2,200	289	265
PASSIVE	attach, do, call	1,818	1,481	167	170
INTRAN_MOT	go, come, get	1,098	915	88	95
TRAN_RES	let, make, get	977	795	90	92
CAUS_MOT	take, put, send	675	546	64	65
DITRAN	give, tell, ask	534	448	40	46
INTRAN_RES	become, go, come	146	121	9	16
Total		26,437	21,194	2,608	2,635