SCIL was the first major project I worked on as part of the LACAI Lab. The goal of SCIL was to identify social roles of individuals within dialogues (such as Leader, Influencer) by analyzing the language used by those individuals (the frequency of certain words, how many words a speaker introduced into the conversation, how speakers influenced the topics of conversation, etc...). SCIL was initially an IARPA project in the late 2000s and early 2010s. Since then, a lot of advancement had been made in Natural Language Processing and Computational Sociolinguistics. The project was initially implemented in Java, so my task was to re-implement SCIL in python as that had become the main language for NLP research.

I chose to re-implement SCIL as a python package to increase the ease of use and to increase the modularity of the toolkit. My main goals were to create an implementation that provided results that were at least as good as the original while also vastly increasing the speed of the program and increasing its functionality. I was able to achieve these goals, adding several new metrics it can calculate, vastly increasing the speed of the toolkit (metrics that took minutes to compute now took seconds), and achieving performance that was as good as the original implementation (on shared metrics). A description of these metrics can be found in the PyPI and Github pages linked below.

In addition to the toolkit, I also worked on an auto-tagger model for identifying metaTags of turns in dialogues with a fellow lab member. We achieved moderate success, creating a model that did produce the correct metaTags significantly better than random and was sufficient enough, in most cases, to produce accurate results with SCIL. As a note, metaTags were required by some metrics but were rarely the only determining component of those metrics, so even if the tags weren't fully accurate, the metric wasn't necessarily that innacurate. The main hinderence to the development of this model was a lack of ground-truth training data. Links to the tagger can also be found below.