Hussein Ghaly presents on a topic in computational linguistics.
In the face of ambiguity: can prosody improve parsing?
In this research, our goal is to improve parsing using prosody. While parsing has traditionally been applied to written language, we believe that spoken language contains additional cues that can be useful in parsing. Prosody is a property of human speech that consists of acoustic patterns of utterances including its timing, pitch, and loudness, and which is related to syntax (Shattuck-hufnagel and Turk, 1996). One use of prosody is to disambiguate sentences containing syntactic ambiguities (Wagner and Watson, 2010). Some algorithms have been developed to predict sentence prosody from syntax (Watson and Gibson, 2004).
In computational linguistics, parsers face difficulty with syntactic ambiguities, such as prepositional phrase (PP) attachment (Kummerfeld et al., 2012). Therefore, since prosody can help disambiguate some syntactic ambiguities, prosody could be used to improve parsing by resolving these ambiguities.
We work on the Switchboard Corpus, which contains audio recordings of spontaneous conversational sentences annotated with ground truth parses and prosodic information (ToBI annotations). We propose here improving parsing through prosody using an ensemble of dependency parsers (spaCy, ClearNLP, and Google SyntaxNet), and a re-ranking algorithm to identify for each sentence the most likely parse hypothesis from the three parsers. By incorporating prosodic information into this re-ranking process, we can improve unlabelled attachment score by 0.57%.
All are welcome!