Dissertations, Theses, and Capstone Projects

Date of Degree


Document Type


Degree Name





Michael Mandel

Committee Members

Janet Dean Fodor

Jason Bishop

Subject Categories

Computational Linguistics | Linguistics | Phonetics and Phonology | Psycholinguistics and Neurolinguistics | Syntax


Prosody Syntactic Parsing Ambiguity Computational Disambiguation


Prosody has strong ties with syntax, since prosody can be used to resolve some syntactic ambiguities. Syntactic ambiguities have been shown to negatively impact automatic syntactic parsing, hence there is reason to believe that prosodic information can help improve parsing. This dissertation considers a number of approaches that aim to computationally examine the relationship between prosody and syntax of natural languages, while also addressing the role of syntactic phrase length, with the ultimate goal of using prosody to improve parsing.

Chapter 2 examines the effect of syntactic phrase length on prosody in double center embedded sentences in French. Data collected in a previous study were reanalyzed using native speaker judgment and automatic methods (forced alignment). Results demonstrate similar prosodic splitting behavior as in English in contradiction to the original study’s findings.

Chapter 3 presents a number of studies examining whether syntactic ambiguity can yield different prosodic patterns, allowing humans and/or computers to resolve the ambiguity. In an experimental study, humans disambiguated sentences with prepositional phrase- (PP)-attachment ambiguity with 49% accuracy presented as text, and 63% presented as audio. Machine learning on the same data yielded an accuracy of 63-73%. A corpus study on the Switchboard corpus used both prosodic breaks and phrase lengths to predict the attachment, with an accuracy of 63.5% for PP-attachment sentences, and 71.2% for relative clause attachment.

Chapter 4 aims to identify aspects of syntax that relate to prosody and use these in combination with prosodic cues to improve parsing. The aspects identified (dependency configurations) are based on dependency structure, reflecting the relative head location of two consecutive words, and are used as syntactic features in an ensemble system based on Recurrent Neural Networks, to score parse hypotheses and select the most likely parse for a given sentence. Using syntactic features alone, the system achieved an improvement of 1.1% absolute in Unlabelled Attachment Score (UAS) on the test set, above the best parser in the ensemble, while using syntactic features combined with prosodic features (pauses and normalized duration) led to a further improvement of 0.4% absolute.

The results achieved demonstrate the relationship between syntax, syntactic phrase length, and prosody, and indicate the ability and future potential of prosody to resolve ambiguity and improve parsing.