John Wilkerson on "Exploring Structural Features for Position Analysis in Political Discussions" by Cäcilia Zirn, Michael Schäfer, Simone Paolo Ponzetto, and Michael Strube:

This paper investigates whether it is possible to extract positional information from FOMC speeches. The authors first make a distinction between speeches based on the linguistic structure of the meetings. Specifically they argue that longer prepared speeches will be more revealing of positions than the shorter spontaneous responses that tend to follow the longer ones. The argument is thoughtful and interesting.

They then compare the cosine similarity of long speaker speeches on the assumption that similar words are an indicator of similar policy positions. These results are then compared to a Reuters scaling of FOMC members with respect to inflation concerns.

The next step in the project will be to incorporate whether the responses signal agreement or disagreement with the prepared speech, and presumably whether that information is also helpful for distinguishing speaker position.

Overall this is a nice project and it is good to see this interest in bringing in contextual information such as speech length and speech sequence. It would also be interesting to see a similar effort in an area, such as congressional floor speeches, where there is a substantial body of research but not along these lines that I am familiar with.

I do have a couple of substantive comments. The first is that FOMC speeches are different from congressional floor speeches. Congressional floor speeches are political. They are public and the policy details have pretty much been decided. Legislators are taking positions, so it makes sense that their words might predict their 'ideology.'

FOMC meeting are private so that politics plays lee of a role in decisions. I don't doubt that board members have ex ante preferences regarding inflation and the like, but it seems less obvious to me that speech similarities reflect positions. Perhaps, for example, topics are what bind members, either because certain members tend to be experts on certain topics, or certain members tend to be preoccupied with certain topics. To me that seems different from position.

Nonetheless, this is an interesting project and I am especially interested in what might emerge from the subdialog investigation. Thanks for participating!