A Simpler Approach to Interactive Topic Modeling
Current topic modelling algorithms have proven useful for the analysis of
large volumes of text. But despite their utility it can still be challenging
to make use of topic modelling, especially for projects requiring a more
interpretive approach. These challenges include awkward computational
workflows with long iteration cycles, limited affordances for editing and
interrogating the resulting models after they have been built, difficulty
interpreting the resulting models, and the difficulty of operationalising
complex numerical models as part of an answer to a research question.
To address these shortcomings we have taken a step back from the conventional
probablistic or linear algebraic approaches to topic modelling and developed
a new approach using a much simpler foundation of boolean information
retrieval and basic clustering of words into groups. Despite giving up on
much of the mathematical sophistication of existing approaches our approach
enables a new class of dynamic interaction between analyst and topic model,
and hopefully allowing for a more concrete link between the created model and
the research question.
We have implemented a prototype of our approach as a Python toolkit enabling
usage as a software library, command line tool, or through a web interface
for interactivity. We will demonstrate the effectiveness of this
implementation and method through a detailed case study on a complex social
media dataset, complemented by a live demonstration of what the analysis
process can look like. In doing so we will also talk about our plans for
future work and extension of this approach.
Recording link: https://acspri-org-au.zoom.us/rec/share/fe80_c8_scf5yQJEY_Vv6vJ6R7XFjJUKwZnkpSbPOus_gP1IIKiSQO8FF_LSLDiZ.j0lwiy1bIA1ehAAO?startTime=1669185660000