The thesis presents a system that automatically converts a recording of a debate into a structured overview of the presented argumentation. We built a
system that automatically performs the whole process from video to readable
report and leaves the judgment to the reader. It accepts a link or an uploaded
file, produces a transcript with speaker separation and analyzes the argumentation in successive steps, five for a debate between two speakers and four for
a single speaker. It first extracts arguments with their premises, then finds
logical fallacies in them, checks the verifiable facts in the premises against
scientific and online sources, links rebuttals to the arguments they attack and
writes a summary. The central design decision is the dividing line between
judgment and computation. Where the language model makes a judgment,
it only chooses from predefined sets of values and gives no numerical scores.
The numbers in the report are computed by the system, which counts these
values, links them across steps and displays them. We evaluated the system
with five repeated analyses of a debate and a single-speaker talk. Almost
all main topics appeared in every repetition, while the number of arguments
and detected fallacies varied between repetitions. Analyzing one recording
costs less than two euros.
|
