Programming, perception, and a priori postulates

I'm in the middle of working on my third, and final, qualifying 'exam'. I took a sit-down exam (no scare quotes) a little over a year ago, and I designed and carried out a study of speaker focus and fricative production during the Spring and Summer of this year. The final 'exam' will be much like the second one - I will design and conduct a research project from the ground up.

This afternoon, I was writing a Matlab script to do some corpus analysis. The general plan with this study is to investigate a couple of different kinds of frequency effects in speech perception. In terms of word recognition, there are a number of well documented frequency effects - on average, frequent words are more accurately recognized in noise identification tasks and responded to more quickly in lexical decision tasks than are infrequent words. Lower level (i.e., sub-word) frequency effects are, as far as I know, less well documented.

With regard to the qual, I am primarily interested in what I have been calling (phonological) contrast frequency. Phonologists call two words with distinct meanings and forms that are identical aside from a single feature difference at a single location a minimal pair. For example, the words 'sue' and 'zoo' - [su] and [zu] - mean two very different things, and the only difference in form is that, at the beginning of the word, the former has a voiceless fricative whereas the latter has a voiced fricative.

In its simplest form, the contrast frequency for a given pair of speech sounds is the number of minimal pairs involving these sounds. You can very likely come up with other minimal pairs involving 's' and 'z', but it would be very hard for you to come up with many minimal pairs for, say, the sounds at the beginning of 'this' and 'think'.

My third qual will address at least one possible psychophysical effect of differences in contrast frequency. Of course, I first have to establish that there are suitable differences in contrast frequency for me to employ in a perception experiment. I was working on this today, using the Hoosier Mental Lexicon, a 20,000 word dictionary that includes machine readable phonemic transcriptions and word usage frequencies, among other informations. It has a good track record, having been put to good use in, for example, word recognition work documenting the effects of lexical neighborhoods (I'll likely post about this at a later date).

I want to use the HML to tally some contrast frequencies so that I can use the best possible pairs of sounds (i.e., those that will maximize the effect I am looking for) to carry out a psychophysical experiment. It turns out to be less than entirely straightforward to tally contrast frequency, mostly because you have to make a number of potentially unwarranted assumptions about the organization of speech sounds (and words) in the mental lexicon.

In general, the idea of contrast frequency seems straightforward - simply count the number of minimal pairs for a given sound. Getting a machine to count the number of minimal pairs is reasonably easy. But what about pairs of words that are nearly minimal pairs (e.g., 'this' and 'think')? It seems to me that, if I'm interested in the relationship between 's' and 'z', I should take into account the relationship between every pair of words with one member containing an 's' and the other a 'z' - 'sue' vs. 'zoo', 'sing' vs. 'zing', sure, but 'ask' vs. 'as' and 'safe' vs. 'zap', and all the rest, too. But if I'm going to take all the occurences of these sounds into account, I have to devise a measure of how similar these two words are (i.e., how important the differences are), and how the location of the 's' and the 'z' in their respective words affects this.

So far, I've written code that will find all the occurences of any given pair of sounds. It then takes each occurence of one of them and, for each occurence of the other, compares their environments - the sounds that come before and after the pair of interest. I've been thinking of various ways to weight the value of a difference in environment according to how far from the pair of interest the difference occur, as it seems reasonable to assume that the immediate environment plays a more important role in contrast frequency. If two sounds in a non-minimal pair are in completely different environments, they will hardly seem contrastive at all. If these sounds are in a minimal pair, they are the very definition of contrastive. In between these two extremes, I assume there is some in-between level of 'contrastiveness', so it seems like a good idea to take these cases into account along with the true minimal pairs.

I've also thought how nice it would be if the transcriptions in the HML included syllable affiliation information for each consonant. It seems reasonable to assume that two sounds in a non-minimal pair would be 'more contrastive' in some sense if they were both in the 'same' syllable position in their respective words. Unless I code this into the HML myself, though, it isn't going to play a role in this project.

By writing code to get a computer to carry these functions out, I have forced myself to make explicit a number of assumptions about how speech sounds are organized in the mind. These assumptions inform a number of potentially important decisions I have to make. To name three, I have to decide how to weight segmental distances when tallying environment differences (should I weight with an exponential decrease or the reciprocal of the number of segments?), how to deal with word edges (if, after aligning occurences of two sounds, the word edges do not line up, how many difference-tallies do the misaligned edges count for?), and how (or whether) to factor in usage frequencies and morphosyntactic properties (do I incorporate raw usage frequencies, the logarithm of raw usage frequencies, and/or the relative proportion of content vs. function words when tallying a pair's contrast frequency?).

The next step is to fix a silly indexing mistake I made (I had to leave promptly at 4:30 to go eat carnitas, and so could not finish the code today), see what the numbers look like for some potentially interesting pairs of sounds, then check the literature on 'functional load', a notion that is likely closely related to my 'contrast frequency'.