Using Prosody to Improve Dependency Parsing – My SP2020 Conference Paper and Presentation

After finishing my PhD dissertation (check my blog post about it), I submitted a paper based on it to the International Conference on Speech Prosody (SP2020), which was to be held in Tokyo last year. I was so excited that the paper was accepted 😃! However, with COVID-19 and everything that happened, the conference was not held as a physical/in-person event 🙄, but mainly publishing the submitted papers and video presentations about the papers from the authors, so I had to figure out how to do that 🤔.

Here is a link to my published paper at the conference.

Here is the youtube video presentation about my paper:

To sum up the content and contribution of this work, here is a little summary:

In speech, people can make pauses or breaks between words (prosodic breaks). The distribution of these breaks is influenced by syntax. When using dependency structure to represent syntax, we can see interesting patterns: if there is a dependency relationship between two words, then there is very little likelihood of having a prosodic break between them, and vice versa. The highest likelihood of a prosodic break between two words is when the first word depends on a word further to the left, and the right word depends on a word further to the right, as we can see between “that” and “what” below:

Using these patterns, and the actual speech data about sentences and words in the Switchboard Corpus, we were able to build a system that can take the prosodic information and use it with syntactic information to select the most likely parse tree from a group of possible parse trees generated by an automatic parser. So in principle, prosody can improve dependency parsing 😃.

However, this will need to be taken further, with more corpora, and possibly more languages, and I’d be glad to collaborate with other researchers in this regard.