Nov 18
======

Peperkamp/etal:2006
-------------------

How many of the 46 phonemes in the pseudo-language overlapped with
French phonemes? How would the results change if there was a larger or
smaller overlap with the real-world language also being tested?

In Experiment 2, the second filter is described as considering
"allophonic distributions of two segments as spurious if for each of
the five articulatory components, the allophone is more distant from
its contexts than the default segment." Thus, the allophone has to be
different (and more distant) to the default segment in each component
to be filtered. But as some of the properties are binary, allophones
of some categories won't be filtered at all. For example, as vowels
are voiced and therefore have the same value for the component
"voicing", they can never be more distant from the context than the
default segment in each component.  Furthermore, the description in
the text does not match the formular in Appendix C which uses an
existential quantifier, not a universal quantifier.

The acquisition of allophonic rules: Statistical learning with
linguistic constraints What application could a statistical algorithm
that acquires allophonic rules have in our everyday life? Is it still
useful more than 15 years later, e.g. in TTS applications?

What are the unique characteristics of infants' acquisition of
allophonic rules?

I would be interested to know how the algorithm would perform on
different data, for example when looking at different allophonic rules
/ less frequently occurring phonemes. Is it possible to establish the
minimum amount of data required for successful detection of the
allophonic distribution?

In this paper the authors make an assumption that allophones appear in
different contexts. What if they put them in one context, for example?
Would the model recognize them after the learning process?

I read Peperkamp et al. (2006) for this week's seminar session and
found its approach and results very interesting especially because
such experiments with an algorithm are novel to me. I have no question
per se but it would be nice if we could clarify the mathematical parts
of the algorithm and discuss more interpretations of the results (Fig
1. and 2.).

What is the benefit of using a pseudo-language in the first experiment?

Even though allophones and default segments usually have
non-overlaping distribution (e.g. [k]-[c]), this is not always the
case. For example in Greek there is an allophonic relationship between
[l] (default segment) and [ʎ] (allophone) but it is possible for the
allophone to appear in words the default segment is expected. Will
such pairs be found as near-complementary distributions how can we
make sure that they are not dismissed by a model?

Does the conclusion they came to hold true for all languages or only French?

The authors state that only after adding linguistic constraints on the
form of possible allophonic rules, the algorithm got better at
detecting spurious allophonic distributions. How would this translate
to a child which is learning the language, where does it get these
constraints from?