Wesley Scivetti

Wesley Scivetti

PhD Student | Computational Linguistics

Posts by Collection

portfolio

publications

Meaning Representation of English Prepositional Phrase Roles: SNACS Supersenses vs. Tectogrammatical Functors

Published in DMR, 2023

Wes Scivetti, Nathan Schneider

We present a conversion scheme between two multilingual semantic annotation frameworks, the Semantic Network of Adposition and Case Supersenses (SNACS) and the Prague Czech-English Dependency Treebank. ocusing on prepositional semantics, we find considerable one-to-one overlaps between labels of the schemas, but find that the many supersenses in the configuration branch of the SNACS framework does not correspond well to any set of tags in PCEDT.

Download Paper

UCxn: Typologically Informed Annotation of Constructions Atop Universal Dependencies

Published in LREC-COLING, 2024

Leonie Weissweiler, Nina Böbel, Kirian Guiller, Santiago Herrera, Wesley Scivetti, Arthur Lorenzi, Nurit Melnik, Archna Bhatia, Hinrich Schütze, Lori Levin, Amir Zeldes, Joakim Nivre, William Croft, Nathan Schneider

We present a first attempt at using Universal Dependencies to annotate a range of constructions (as defined by Construction Grammar) using queries over edges and morphological lasbels in UD. We present an initial pilot across 5 constructions and 10 languages.

Download Paper

GDTB: Genre Diverse Data for English Shallow Discourse Parsing across Modalities, Text Types, and Domains

Published in EMNLP, 2024

Yang Janet Liu*, Tatsuya Aoyama*, Wesley Scivetti*, Yilun Zhu*, Shabnam Behzad, Lauren Elizabeth Levine, Jessica Lin, Devika Tiwari, Amir Zeldes (*equal contribution)

We present GDTB, a genre diverse discourse relation resource in the Penn Discourse Treebank (PDTB) style. We present empirical evaluations on the new resource, showing that mixed training on PDTB and GDTB leads to optimal performance.

Download Paper

Multilingual Supervision Improves Semantic Disambiguation of Adpositions

Published in COLING, 2025

Wes Scivetti, Lauren Levine, Nathan Schneider

We present the first set of multilingual fine-tuning and evaluations for the SNACS framework. We reveal, via a comparative analysis on parallel data, that the relative frequencies for different supersenses are highly dependent on language. Using data from 5 languages and a robust hyperparameter sweep, we substantially outperform the state-of-the-art on SNACS classification, finding that optimal performance is achieved by joint fine-tuning on all languages at once.

Download Paper

Construction Identification and Disambiguation Using BERT: A Case Study of NPN

Published in CoNLL, 2025

Wes Scivetti, Nathan Schneider

We present a set of probing experiments targeting both the syntactic and semantic properties of the Noun-Preposition-Noun (NPN) construction. Using a set of manually annotated data extracted from the Corpus of Contemporary American English (COCA), we find that probes trained over BERT representations are sensitive to both the form and function of the construction.

Download Paper

Unpacking Let Alone: Human-Scale Models Generalize to a Rare Construction in Form but not Meaning

Published in EMNLP, 2025

Wes Scivetti, Tatsuya Aoyama, Ethan Wilcox, Nathan Schneider

We evaluate human-scale (BabyLM) language models on the extremely rare let-alone construction, finding that they master a range of syntactic properties, but are not sensitive to the construction’s semantics. We then perform a set of Filtered Corpus Training (FiCT), finding robust performance on constructional syntax even in the absence of direct observation of let-alone or related Paired Focus and Comparative Constructions.

Download Paper

talks

teaching