From: Christopher Potts
Date: July 24, 2011 22:13:51 MDT
To: Class list
Subject: Computational Pragmatics: student suggestions implemented

Computational Pragmaticists!

I'm writing to describe two improvements I made to the course codebase and webpages based on student suggestions from Friday's class.

1. Joseph observed that the current Switchboard Dialog Act code and experiments treat '+' as just another tag, though its interpretation is "continued from previous by same speaker".  I've now modified swda.py so that one can follow these + tags back to their source.  The implementation is not as intuitive as I would like (to avoid a lot of modifications to the code at this point), but it works well enough:

a. Instantiate the corpus as usual:

corpus = CorpusReader('swda')

b. When you iterate with corpus.iter_utterances() or corpus.iter_transcripts(), you can supply a keyword argument follow_plus.  If follow_plus=False (the default), + is left alone. If follow_plus=True, then + is converted to its source act tag.  For example:

# Count the tags, following + to its source:
d = defaultdict(int)
for utt in corpus.iter_utterances(follow_plus=True): 
	d[utt.act_tag] += 1
	# Or, for DAMSL, which is also affected:
	# d[utt.damsl_act_tag()] += 1

I adjusted tag_counts() in swda_functions.py so that tag_counts(follow_plus=True) will do the counts this way, whereas tag_counts() or tag_counts(follow_plus=False) will leave + alone.


2. During our discussion of the word-tag clustering experiment, I wasn't able to articulate why I normalized the vectors by length rather than by turning them into probability distributions.  Relatedly, I did not have an explanation for what the x-axis meant in figure 2:

http://compprag.christopherpotts.net/swda-clustering.html#NORMED

Ariel explained it to me after class, though: the function length_normalization is sensitive to the extent to which the values are distributed across the vector: it delivers higher values for vectors that are highly distributed.  Here's an illustration in Python:

import numpy

def length_normalization(vec): 
	vec = numpy.array(vec)
	return vec / numpy.sqrt(numpy.dot(vec, vec))

## Most distributed; sum = 40
sum(length_normalization([8,8,8,8,8]))
2.2360679774997898

## Less distributed; sum = 40
sum(length_normalization([10,10,10,10,0]))
2.0

## Even less distributed; sum = 39.999
sum(length_normalization([13.333,13.333,13.333,0,0]))
1.7320508075688772

In terms of figure 2, we see that 'absolutely', 'exactly', and 'definitely' occur with basically only 1 act tag (it happens to be 'aa'), whereas 'oh' is widely distributed relative to the act tags.

I've added a note explaining this to the caption of figure 2.

---Chris