Seminar • Data Visualization — Human-Centered Machine Learning: Approaches for Mixed-Initiative Topic Model Refinement

Thursday, April 22, 2021 12:00 pm - 12:00 pm EDT (GMT -04:00)

Please note: This seminar will be given online.

Mennatallah El-Assady, Research Associate
Data Analysis and Visualization, University of Konstanz
Visualization for Information Analysis Lab, University of Ontario Institute of Technology

Applying topic modeling algorithms to analyze the content of text corpora is a prevalent task in the humanities and social sciences. However, as their results are typically subjective and domain-dependent, there is no single ground-truth segmentation of a corpus that can be used to optimize such models. Hence, refining these models has relied on humans externalizing their knowledge to adapt the results to their domain understanding. To tackle this challenge, I rely on human-centered machine learning, contributing novel human-in-the-loop paradigms to explain, diagnose, and refine topic models.

In this talk, I present four different workflow and interface designs, each tailored to a different user group. First, data scientists can use the progressive learning approach to give feedback on the relation between inputs and outputs efficiently. Next, we show all considered model decisions while reconstructing conversational text data to allow analysts to understand the model’s decision space. Afterward, I present an approach using an intelligible model for machine learning experts. Based on live updates, the experts can investigate the model and see the impact of their interactions before applying them using speculative execution. Lastly, for domain experts, I introduce Semantic Concept Spaces, a workflow for applying a machine teaching paradigm to capture the experts’ knowledge.

This series of approaches introduces different workflows but achieves comparable refinement results. This allows machine learning experts and non-experts alike to rely on tailored human-AI interactions to refine topic models for their data and tasks.


Bio: Mennatallah El-Assady is a research associate in the group for Data Analysis and Visualization at the University of Konstanz (Germany) and in the Visualization for Information Analysis lab at the University of Ontario Institute of Technology (Canada). She works at the intersection of data analysis, visualization, computational linguistics, and explainable artificial intelligence. Her general research interest is in combining data mining and machine learning techniques with visual analytics, specifically for text data. In particular, she is researching methods of the automatic analysis and visualization of transcribed verbatim text corpora.

She has gained experience working in close collaboration with political science and linguistic scholars over several years, which led to the development of the http://lingvis.io/ platform. More recently, she has been working on establishing the expandable AI framework http://explainer.ai/. El-Assady has co-founded and co-organized several workshop series, notably http://vis4dh.org/http://visxai.io/, and http://argvis-workshop.lingvis.io/.


To join this seminar on Zoom, please go to https://zoom.us/j/96656478119?pwd=TDhaZVNhelpuN3U5NFE5RUJKckxMdz09.