Conversational alignment
Switch branches/tags
Nothing to show
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Failed to load latest commit information.
Childes_Results
Childes_userInfo
models
papers
parsers
shells
wordlists
.gitignore
LICENSE
README.md
test.sh

README.md

alignment

Conversational alignment

Code examples - Reading in a csv https://docs.python.org/2/library/csv.html - Grouping list of dictionaries https://www.daniweb.com/software-development/python/code/216750/group-a-list-of-dictionaries-python

Formulas - TRUE_POWER (Outputs the expected results) - DNM (Formula used in Echoes of Power

changelog

6/26/2015

Derived formula for Echoes of Power: powerProb = ((# of times marker is said by A and B)(# of words A says to B))/((# of words B says to A)(# of times A says the marker)) baseProb = (# of times B says the marker)/(# of words B says) prob = powerProb - baseProb

6/29/2015

Figured out that the formula we were working off of was incorrect.

Instead of calculating probabilities using tokens, we need to calculate probabilites using utterances powerProb = (# of utterances where A and B both say marker)/(Number of utterances where A says the marker) baseProb = (# of utterances where B says the marker)/(# of utterances that B says to A) prob = powerProb - baseProb

7/13/2015

Figured out that the formula we were working off of was incorrect.

Our new formula is:

powerProb = log((# of utterances where A and B both say marker)/(Number of utterances where A says the marker)) baseProb = log((# of utterances where B says the marker and A doesn't)/(# of utterances that A doesn't say the marker)) prob = powerProb - baseProb

7/20/2015

SHARED_CODE: Fixed a bug which was reducing our BNOTA

7/22/2015

SHARED_CODE: Added DNM formula option and options to change smoothing values

7/23/2015

TWTR: Updated readCSV to discard more utterances which quote a speaker

- Check for unicode quotes
- Check if speakers acknowledge each other
- Check if utterance contains "RT"
- Check if utterance contains "[mention]: "

7/27/2015

- No longer discard if speakers don't acknowledge each other