Brandon Leung

Brandon Leung

Passionate in computer vision research & software engineering. M.S. machine learning & data science, UCSD.

Statistical Linguistic Analysis for User Chat Message Logs

📅 Feb. 2021 - Jul. 2021 • 📰 Slides • Github

In this project, I contributed to the development of an interactive dashboard to analyze user chat logs and describe their linguistic behavior. It is meant for smaller companies and organizations to easily understand the linguistic patterns of their user base. Many applications are integrated with social chatting systems, including video games, dating apps, and social media. With this dashboard, even verbose chatters with thousands of logged chat messages can be summarized, evaluated, and compared with other users at a glance. This is enabled through sentiment analysis, clustering, style transfer, and generative modeling. Then, downstream use cases include flagging/suspending/banning toxic users, recommending advertisements or posts to users, and even using their “virtual” chatbot counterpart to predict their behavior to new inputs. The dashboard is powered with Jupyter Notebooks and Voilà, tested with pytest, documented with Sphinx, and deployed using AWS (EC2 and S3).

Within my group of collaborators, I mainly worked on the UI, deployment, and generative modeling. The generative modeling utilizes the BlenderBot chatbot based on RoBERTa, along with unsupervised style transfer based on a Seq2Seq Transformer from Krishna Et. al. As a proof of concept, a Discord chat dataset was used.

The initial user interface. The user uploads a chat log to start the analysis.
The Patterns & Clusters tab provides basic statistical information from the uploaded chat log. Additionally, the chatter is plotted in 3D using TF-IDF vectors and k-means clustering. This allows one to find other chatters similar to the uploaded chatter.
The Sentiment tab classifies all the messages from the uploaded chatter's logs, into one of five emotions: sadness, joy, neutral, anger, and fear. The top 5 sentences of each emotion can be shown.
The Chatbot tab provides an interactive stylized chatbot, which one can use to query novel, style-consistent responses from based on the uploaded chatter's logs.
The deployment architecture for the dashboard.
More examples of generative modeling from the stylized chatbot, trained on several users' chat logs.

Other collaborators on this project: Shivad Bhavsar, Rex Chen, Jiayi Luo, Rongxiang Zhang, Kevin Youssef.