[developers] Lexical rules changing predicate symbols

[Ann Copestake]

Hi Emily et al,

A very very belated response to a message from months ago - I had 
started creating a long reply, thought I had lost it, and now discovered 
it hanging around, so thought I might as well send it. I feel I was 
agreeing with your message, but just trying to put some flesh on it from 
my own viewpoint.

Here are some thoughts, more related to trying to distinguish between 
the contexts where we might use different bits of machinery, but trying 
to link up with the theoretical issues.  I'm afraid I know so little 
about phonology that I have no clue about words-for-phonology, but I 
have some ideas about words-for-syntax vs words-as-lexicalised-units vs 
words-as-writing-convention. I'm going to say some thing which are maybe 
too basic below, but I feel it's important to be specific (and I never 
know exactly when people are going to disagree with me ...)

1. The criterion for representing something as a pattern (where I'm 
trying to use `pattern' as a general term, to include rules, types and 
sense-changes), rather than just listing, is that it should demonstrate 
at least semi-productivity - i.e., there is a pattern that is 
sufficiently regular that capturing it should be part of an ideal 
synchronic description of a language
       - of course there are different notions of productivity depending 
on whether one is thinking of an individual speaker or a language 
community.  If the former, there's the distinction between adult native 
speakers, children and second language learners - we can distinguish 
between idioms of encoding and decoding - we can also distinguish 
between the dimensions of psycholinguistic evidence, behaviour and and 
people's own self-awareness.  Computationally, the test is whether or 
not it's useful, and that may not exactly match any of these. 
Computational usefulness might be about coverage (in which case it's 
reasonable to balance with problems about ambiguity) but it could also 
be about learnability in some contexts (possibly not so relevant for 
DELPH-IN).

2. the (semi-)productive pattern can't be described as part of syntax 
without serious problems - most obviously because it's happening within 
a `word' (according to whatever notion of word is being worked with ...) 
- and it's not part of an inflectional morphological paradigm.  I would 
ideally want to have the option of treating some phenomena as 
essentially syntactic within a word-as-writing-convention, if the 
conventional word boundary notion just doesn't work well.  (The paper 
that Guy wrote for HPSG 2015 is relevant here.)  I don't think that 
really happens much in English, where at least partial lexicalisation 
seems to operate rather generally on words (and also on some phrases).  
But in other languages, it just seems to be the case that the notion of 
word really doesn't work in the same way.  That is to say - there's a 
notion of `word' which is a rather arbitrary artifact of writing systems 
and it's not at all clear we should spend much effort trying to preserve 
that, when we'd be better off segmenting and treating the segments as 
things that the syntax can operate on.  Whether or not we get 
lexicalisation, as evidenced by idiosyncratic syntax, semantics, 
phonology or whatever, seems a better guide to whether we should regard 
something as a word-for-syntax, because it will correlate with what 
humans learn as a unit.  However, for practical DELPH-IN grammar 
purposes, we are currently somewhat stuck with using the units we're 
given by the convention of a language, and then having to do other forms 
of processing to construct the words-for-syntax (in the case of 
multiword expressions) or using syntax-like lexical-rules or similar (in 
the case where the conventional word is too big).

3. but there is a relationship with syntax
- There are meaning shifts that I wouldn't suggest trying to capture at 
all in a DELPH-IN grammar because they don't interact with syntax, and 
hence we'd simply introduce ambiguity - they can be treated by 
operations on the meaning space that a single predicate symbol gives us, 
possibly as a post-processing step. This isn't a very clear division 
because there are some cases which do interact but do so in a rather 
marginal way.  For instance, strawberry the fruit is much more likely to 
be used as a mass term than strawberry the plant, and the fruit/plant 
polysemy is almost fully productive in English but is not found in all 
languages.  So this would meet my ideal linguistic criteria of a pattern 
but is too fine-grained for us to want to capture it in the ERG, given 
that the grammar allows many words to be general between mass and 
count.  In any case, as long as we can easily reconstruct that we have a 
necessarily mass-context, a necessarily count-context or a vague 
context, we can do what we need to with postprocessing.  The same 
applies with the good old grinding examples, of course.

4. By this point, we have a choice between a lexical rule with some sort 
of explicitly-specified semantic content and, on the new proposal, one 
that just changes the sense in the predicate (along with whatever 
syntactic effect we're aiming at capturing).  The former case is clearly 
appropriate for stuff which is very like a syntax operation but just 
happens to operate on units which are smaller than the conventional word 
boundary.  For instance, in Semitic languages which represent a subject 
as an affix on the verb, it seems clear this is something that we'd want 
to treat very like a syntactic rule, and our current notion of lexical 
rule is fine for that.

On the other hand, as you said, there are other processes where the 
pattern is regular but with some exceptions, and where the precise 
semantic effect is difficult to pin down.  Much of English derivational 
morphology falls into this class - it is useful to represent the 
relationship with the stem somehow, without claiming that we're 
capturing everything there is to say.  e.g., "unkind" definitely has 
some relationship with (a sense of) "kind", and it's useful to know 
about that, but there are some idiosyncratic aspects of its meaning.   
Similarly, I'd say it's useful to represent the relationship between 
nouns denoting dances and the corresponding verbs (tango etc), which is 
productive, but I'm quite content to do that via a predicate changing 
operation.

Evidence for a psycholinguistic distinction between these two classes is 
far messier than one would like, and in any case, `lexicalisation' can 
operate in situations like adjective noun combination, which we think of 
as clearly part of syntax, but I think we just have to accept that our 
representational devices have to cut the world into classes which miss 
some nuances.  At least until deep learning comes along to save us ...

Anyway, in terms of the current way I'm thinking about compositionality, 
predicate changing operations belong in Phase 1, which isn't subject to 
the same sort of monotonicity constraint as Phase 2.

All best,

Ann

On 20/01/16 22:11, Emily M. Bender wrote:
> Dear Ann, Dear all,
>
> Picking up this thread again, with apologies for the delay. Thank you, 
> Ann, for
> the detailed reply!  If I've understood correctly, non-monotonicity in 
> lexical rules
> is okay from a practical perspective because the search space will be 
> constrained
> (given that the predicate symbol isn't allowed to change, only the 
> sense) and
> from a theoretical perspective because: (a) the lexical rules are 
> modeling processes
> that are perhaps more diachronic than synchronic, such that we expect 
> some drift
> in the lexical semantics of the derived forms, (b) the leaps that 
> people make in
> assigning words new meanings (using metaphor, etc) probably can't be 
> modeled
> with strict compositionality anyway.
>
> That leads me to wonder whether we have two definitions of 'word' in 
> play that
> might not line up as well as we'd like: One has to do with syntax 
> (words are the units
> that syntax gets to play with) and phonology (words are the domains of 
> a certain
> set of phonological rules; though already these two don't line up 
> quite right).  The
> other has to do with lexicalization and words as units of 
> (potentially) idiosyncratic
> form-meaning pairing.  That's probably an over-blown way of saying: in 
> some languages,
> there's more "grammar" going on inside of what we might call words by 
> the first
> definition, and I would like to think that any fully productive 
> morphological process
> is still compositional.
>
> All of this is probably highly relevant to what Joshua is working on 
> with Lushootseed,
> but it might be hard to tease apart which of those transitivizing 
> suffixes are productive
> and which aren't...
>
> Emily
> ​