<html>
  <head>
    <meta content="text/html; charset=utf-8" http-equiv="Content-Type">
  </head>
  <body bgcolor="#FFFFFF" text="#000000">
    Hi Emily,<br>
    <br>
    I'll be on (non-emailing responding!) vacation for three weeks from
    this weekend, so thoughtful discussion would have to wait till after
    that.  But quick comments below again.<br>
    <blockquote
cite="mid:CAMype6cMLGYr-ZsipjaYox5Gr3o25oUTAkPkr7rZ8Q1e_WX+HA@mail.gmail.com"
      type="cite">
      <div dir="ltr">
        <div class="gmail_extra">
          <div class="gmail_quote">
            <blockquote class="gmail_quote" style="margin:0 0 0
              .8ex;border-left:1px #ccc solid;padding-left:1ex">
              <div bgcolor="#FFFFFF" text="#000000">
                <div>
                  <div class="h5">
                    <blockquote type="cite">
                      <div dir="ltr">
                        <div class="gmail_extra">
                          <div class="gmail_quote">
                            <blockquote class="gmail_quote"
                              style="margin:0 0 0 .8ex;border-left:1px
                              #ccc solid;padding-left:1ex"><span>
                                <blockquote class="gmail_quote"
                                  style="margin:0 0 0
                                  .8ex;border-left:1px #ccc
                                  solid;padding-left:1ex"> The UW group
                                  has been reading and discussing
                                  Copestake et al 2001<br>
                                  and Copestake 2007, trying to get a
                                  better understanding of the MRS<br>
                                  algebra.  We have a few questions---I
                                  think some of these issues have been<br>
                                  proposed for the Summit, but I'm
                                  impatient, so I thought I'd try to get<br>
                                  a discussion going over email.  UW
                                  folks: Please feel free to chime in<br>
                                  with other questions I haven't
                                  remembered just now.<br>
                                  <br>
                                  The two big ones are:<br>
                                  <br>
                                  (1) Copestake et al 2001 don't
                                  explicitly state what the purpose of
                                  the<br>
                                  algebra is.  My understanding is that
                                  it provides a guarantee that the MRSs<br>
                                  produced by a grammar are well-formed,
                                  so long as the grammar is<br>
                                  algebra-compliant.   Well-formed MRS
                                  (in this sense) would necessarily<br>
                                  have an interpretation because the
                                  algebra shows how to compose the<br>
                                  interpretation for each sement.  Is
                                  this on track?  Are there other
                                  reasons<br>
                                  to want an algebra?<br>
                                </blockquote>
                                <br>
                              </span> We have never managed to prove
                              that MRSs constructed according to the
                              algebra will be scopable, but I think that
                              is the case.  But more generally, the
                              algebra gives some more constraints to the
                              idea of compositionality which isn't the
                              case if you simply use feature
                              structures.  It excludes some possible
                              ways of doing semantic composition and
                              therefore constitutes a testable
                              hypothesis about the nature of the
                              syntax-semantics interface.  It also
                              allows one to do the same semantic
                              composition with grammars in formalisms
                              other than typed feature structures.<span><br>
                                <br>
                              </span></blockquote>
                            <div><br>
                            </div>
                            <div>I'm still trying to understand why it's
                              important (or maybe interesting is</div>
                            <div>the goal, rather than important?) to
                              exclude some possible ways of doing</div>
                            <div>semantic composition.  Are there ways
                              that are problematic for some reason?</div>
                            <div>Is it a question of constraining the
                              space of possible grammars (e.g. for</div>
                            <div>learnability concerns)?  Related to
                              issues of incremental processing?</div>
                          </div>
                        </div>
                      </div>
                    </blockquote>
                    <br>
                  </div>
                </div>
                The following may sound snarky but I really don't mean
                it this way - would you have the same type of questions
                about syntactic formalism?  Because I can answer in two
                ways - one is about why I think it's important to build
                testable formal models for language (models which aren't
                equivalent to general programming languages, so more
                constrained thn typed feature structures) and the other
                is about what the particular issues are for
                compositional semantics.  I don't want to reach the
                conclusion that semantic composition requires arbitrary
                programs without looking at more constrained
                alternatives.  So yes: learnability, processing
                efficiency (human and computational), incrementality and
                so on, but one can't really look at these in detail
                without first having an idea of plausible models.<span
                  class=""><br>
                  <br>
                </span></div>
            </blockquote>
            <div><br>
            </div>
            <div>Not snarky and totally a fair question.  Our syntactic
              formalism in fact isn't constrained.</div>
            <div>Part of what's appealing about HPSG is that the
              formalism is flexible enough to state</div>
            <div>different theories in.  <br>
            </div>
          </div>
        </div>
      </div>
    </blockquote>
    <br>
    I guess for me, the typed feature structure formalism was developed
    (largely) independently of HPSG and what's attractive about it is
    that it seems to be a particularly good programming language for
    grammar engineering.  <br>
    There are restrictions on what people are willing to call HPSG, but
    these are flexible/inconsistent, and the community (or subparts of
    the community) changes its mind sometimes.  I'm not at all unhappy
    with this, as long as there is progress.  Anyway, some of these
    restrictions can be elegantly encoded in typed feature structures
    and others can't.  Working out what the restrictions amount to
    formally should be an objective, even if we don't then change
    formalism.<br>
    <br>
    <blockquote
cite="mid:CAMype6cMLGYr-ZsipjaYox5Gr3o25oUTAkPkr7rZ8Q1e_WX+HA@mail.gmail.com"
      type="cite">
      <div dir="ltr">
        <div class="gmail_extra">
          <div class="gmail_quote">
            <div>(As opposed to certain other approaches that have to
              hard code </div>
            <div>theoretical claims into the formalism.) </div>
          </div>
        </div>
      </div>
    </blockquote>
    <br>
    they could all be encoded in typed feature structures, of course,
    but in some cases the claims/constraints can't be represented
    particularly nicely in TFSs.  for example, if we're representing a
    categorial grammar, we have to implement forward and backward
    application in feature structures, and there's an implicit claim
    there are no additional rules.  We do a similar thing with our
    grammar rules, of course, but HPSG allows for differing number of
    constructions in a grammar without a claim that the framework has
    changed, while in categorial grammar the formalism and framework is
    defined by the rule inventory.  However, if at some point we really
    do decide we know how to do syntactic subcategorization (say), it
    would make sense to work out precisely what we're doing and relate
    it to other frameworks.  <br>
    <br>
    <blockquote
cite="mid:CAMype6cMLGYr-ZsipjaYox5Gr3o25oUTAkPkr7rZ8Q1e_WX+HA@mail.gmail.com"
      type="cite">
      <div dir="ltr">
        <div class="gmail_extra">
          <div class="gmail_quote">
            <div>So at that point, I think it makes sense to see the</div>
            <div>algebra as a theory of composition that can be
              implemented in the HPSG formalism </div>
            <div>(or others).  </div>
          </div>
        </div>
      </div>
    </blockquote>
    <br>
    except that we don't formally constrain composition in TFS in the
    way that the algebra assumes ... <br>
    <br>
    <blockquote
cite="mid:CAMype6cMLGYr-ZsipjaYox5Gr3o25oUTAkPkr7rZ8Q1e_WX+HA@mail.gmail.com"
      type="cite">
      <div dir="ltr">
        <div class="gmail_extra">
          <div class="gmail_quote">
            <div>I'm totally with you on building testable formal models
              (and testable at</div>
            <div>scale), so the program of formalizing the general
              approach of the ERG and then </div>
            <div>seeing whether the whole grammar can be made consistent
              with that set of 'best practices'</div>
            <div>makes a lot of sense.  I guess the piece that's new to
              me is why this should be considered</div>
            <div>for composition separate from the rest of the grammar. 
              <br>
            </div>
          </div>
        </div>
      </div>
    </blockquote>
    <br>
    The first point would be that different bits of the syntax need
    different types of restriction - e.g., subcategorization is
    different from agreement, although they obviously interact.  The
    second point is that what we're trying to accomplish with semantics
    is different, specifically that the primary interest is relating
    structures for sentences/phrases to structures for words and that we
    never filter structures by the compositional semantics.  This is
    actually something that has a reflex in the implementations, since
    we don't need to compute the compositional semantics until we have a
    complete parse.<br>
    <br>
    <blockquote
cite="mid:CAMype6cMLGYr-ZsipjaYox5Gr3o25oUTAkPkr7rZ8Q1e_WX+HA@mail.gmail.com"
      type="cite">
      <div dir="ltr">
        <div class="gmail_extra">
          <div class="gmail_quote">
            <div><br>
            </div>
            <div>(Also, for the record, I'm always skeptical about
              arguments grounded in learnability,</div>
            <div>since I think they require a leap that I'm not ready to
              make that our models are actually</div>
            <div>relatable to what's 'really' going on in wet-ware.  </div>
          </div>
        </div>
      </div>
    </blockquote>
    <br>
    actually that wouldn't follow - learnability results are
    fundamentally about information rather than implementation.  But I
    don't actually have anything useful to say about learnability.<br>
    <br>
    <blockquote
cite="mid:CAMype6cMLGYr-ZsipjaYox5Gr3o25oUTAkPkr7rZ8Q1e_WX+HA@mail.gmail.com"
      type="cite">
      <div dir="ltr">
        <div class="gmail_extra">
          <div class="gmail_quote">
            <div>But processing efficiency, incrementality,</div>
            <div>etc are still interesting to me.)</div>
          </div>
        </div>
      </div>
    </blockquote>
    <blockquote
cite="mid:CAMype6cMLGYr-ZsipjaYox5Gr3o25oUTAkPkr7rZ8Q1e_WX+HA@mail.gmail.com"
      type="cite">
      <div dir="ltr">
        <div class="gmail_extra">
          <div class="gmail_quote">
            <div> </div>
            <blockquote class="gmail_quote" style="margin:0 0 0
              .8ex;border-left:1px #ccc solid;padding-left:1ex">
              <div bgcolor="#FFFFFF" text="#000000"><span class="">
                  <blockquote type="cite">
                    <div dir="ltr">
                      <div class="gmail_extra">
                        <div class="gmail_quote">
                          <div> </div>
                          <blockquote class="gmail_quote"
                            style="margin:0 0 0 .8ex;border-left:1px
                            #ccc solid;padding-left:1ex"><span>
                              <blockquote class="gmail_quote"
                                style="margin:0 0 0 .8ex;border-left:1px
                                #ccc solid;padding-left:1ex"> <br>
                                Subquestions:<br>
                                <br>
                                 (1a) I was a bit surprised to see the
                                positing of labels in the model.  What<br>
                                would a label correspond to in the
                                world?  Is this akin to reification of
                                propositions?<br>
                                Are we really talking about all the
                                labels here, or just those that survive
                                once<br>
                                an MRS is fully scoped?<br>
                              </blockquote>
                              <br>
                            </span> the model here is not a model of the
                            world - it's a model of semantic structures
                            (fully scoped MRSs).<span><br>
                              <br>
                            </span></blockquote>
                          <div><br>
                          </div>
                          <div>Oh, I missed that in the paper.  So are
                            we then talking about "interpretation"</div>
                          <div>because the fully scoped MRSs are assumed
                            to have interpretations via </div>
                          <div>a second layer of modeling?</div>
                        </div>
                      </div>
                    </div>
                  </blockquote>
                  <br>
                </span> yes<span class=""><br>
                  <br>
                  <blockquote type="cite">
                    <div dir="ltr">
                      <div class="gmail_extra">
                        <div class="gmail_quote">
                          <div> </div>
                          <blockquote class="gmail_quote"
                            style="margin:0 0 0 .8ex;border-left:1px
                            #ccc solid;padding-left:1ex"><span>
                              <blockquote class="gmail_quote"
                                style="margin:0 0 0 .8ex;border-left:1px
                                #ccc solid;padding-left:1ex">  (1b) How
                                does this discussion relate to what Ann
                                was talking about at IWCS<br>
                                regarding the logical fragment of the
                                ERG and the rest of the ERG?  That is,<br>
                                if all of the ERG were
                                algebra-compliant, does that mean that
                                all of the ERSs<br>
                                it can produce are compositional in
                                their interpretation? Or does that
                                require<br>
                                a model that can "keep up"?<br>
                              </blockquote>
                              <br>
                            </span> it's really orthogonal - what I was
                            talking about at IWCS was about the complete
                            MRSs.<span><br>
                              <br>
                            </span></blockquote>
                          <div><br>
                          </div>
                          <div>Got it.</div>
                          <div> </div>
                          <blockquote class="gmail_quote"
                            style="margin:0 0 0 .8ex;border-left:1px
                            #ccc solid;padding-left:1ex"><span>
                              <blockquote class="gmail_quote"
                                style="margin:0 0 0 .8ex;border-left:1px
                                #ccc solid;padding-left:1ex"> (2)
                                Copestake et al state: "Since the
                                constraints [= constraints on grammar
                                rules<br>
                                that make them algebra-compliant] need
                                not be checked at runtime, it seems<br>
                                better to regard them as metalevel
                                conditions on the description of the
                                grammar,<br>
                                which can anyway easily be checked by
                                code which converts the TFS into the<br>
                                algebraic representation."  What is the
                                current thinking on this?  Is it in fact<br>
                                possible convert TFS (here I assume that
                                means lexical entries &amp; rules?) to<br>
                                algebraic representation?  Has this been
                                done?<br>
                                <br>
                              </blockquote>
                              <br>
                            </span> `easily' might be an exaggeration,
                            but the code is in the LKB, though it has to
                            be parameterised for the grammar and may not
                            work with the current ERG. You can access it
                            via the menu on the trees, if I remember
                            correctly.  The small mrscomp grammar is
                            algebra compliant, the ERG wasn't entirely
                            when I tested it.<span><br>
                              <br>
                            </span></blockquote>
                          <div><br>
                          </div>
                          <div>In the spirit of keeping the discussion
                            going without delays, I haven't actually</div>
                          <div>played with this yet.  But: accessible
                            from the trees seems to suggest that the</div>
                          <div>testing takes place over particular
                            analyses of particular inputs, and not
                            directly</div>
                          <div>on the grammar as static code analysis. 
                            Is that right?</div>
                          <div><br>
                          </div>
                        </div>
                      </div>
                    </div>
                  </blockquote>
                  <br>
                </span> I think one could use static analysis based on
                that code on a grammar which was very careful about not
                having pointers into the semantics other than those
                licensed by the algebra.  Either no such pointers at
                all, or ones in which there was a clear locality, so it
                was never possible for information to sneak back into
                the semantics bypassing the official composition.  
                Proving that can't happen with a grammar like the ERG is
                difficult.<br>
                <br>
              </div>
            </blockquote>
            <div><br>
            </div>
            <div>Would it be possible at least to detect pointers into
              the semantics, for hand-investigation?</div>
            <div><br>
            </div>
          </div>
        </div>
      </div>
    </blockquote>
    <br>
    well, yes, you'd just have to traverse the graph, I guess.  I think
    there actually is some code that might do what is needed, related to
    the check that ensures the semantics can be separated off from the
    syntax.<br>
    <br>
    All best,<br>
    <br>
    Ann<br>
    <br>
    <blockquote
cite="mid:CAMype6cMLGYr-ZsipjaYox5Gr3o25oUTAkPkr7rZ8Q1e_WX+HA@mail.gmail.com"
      type="cite">
      <div dir="ltr">
        <div class="gmail_extra">
          <div class="gmail_quote">
            <div>Emily</div>
            <div> </div>
            <blockquote class="gmail_quote" style="margin:0 0 0
              .8ex;border-left:1px #ccc solid;padding-left:1ex">
              <div bgcolor="#FFFFFF" text="#000000">
                <blockquote type="cite">
                  <div dir="ltr">
                    <div class="gmail_extra">
                      <div class="gmail_quote">
                        <div>Thanks again,</div>
                        <div>Emily</div>
                        <div><br>
                        </div>
                      </div>
                    </div>
                  </div>
                </blockquote>
                <br>
                you're welcome!  <br>
                <br>
                All best,<br>
                <br>
                Ann<br>
                <br>
              </div>
            </blockquote>
          </div>
          <br>
          <br clear="all">
          <div><br>
          </div>
          -- <br>
          <div class="gmail_signature">
            <div dir="ltr">Emily M. Bender<br>
              Professor, Department of Linguistics<br>
              Check out CLMS on facebook! <a moz-do-not-send="true"
                href="http://www.facebook.com/uwclma" target="_blank">http://www.facebook.com/uwclma</a><br>
            </div>
          </div>
        </div>
      </div>
    </blockquote>
    <br>
  </body>
</html>