Re: [Link Grammar] Trouble Building 4.7.10. Patch Available?

[Linas Vepstas]

Hi,

On 23 March 2013 12:15, Jen Seale <jennife...@gmail.com> wrote:

Hi, Is there a patch for the following? (see highlighted) Thanks so much, Jen 

  CCLD   liblink-grammar.la

Undefined symbols for architecture x86_64:

  "_build_disjuncts_for_dict_node", referenced from:

     -exported_symbol[s_list] command line option

ld: symbol(s) not found for architecture x86_64

The following should fix this: 

edit the file link-grammar/link-grammar.def

delete the line that says "build_disjuncts_for_dict_node", it should be line 183

 

I assume you are not trying to build the "corpus statistics" version. If you are, then I need to provide a more complex fix.  I'll provide a version 4.7.11 later today or tomorrow.

--linas

[Linas Vepstas]

I just posted version 4.7.11 which contains the one-line fix suggested below.

-- Linas

[Jen Seale]

Thank you, L. I'm excited to try it out with the corpus stats. - J

[Linas Vepstas]

On 26 March 2013 11:15, Jen Seale <jennife...@gmail.com> wrote:

Thank you, L. I'm excited to try it out the corpus stats. -J

For parse ranking, they're not impressive, underwhelming, even.

Far more interesting is the automatic wordnet word-sense tagging.  This has OK accuracy but what makes it remarkable is that its "instant" or "free", and requires no computation: just a table lookup.   The idea is kind-of neat:  The link-grammar disjuncts are kind-of-like extremely fine-grained part-of-speech tags -- e.g. it identifies a word as not just a noun, but a noun used in a very specific way.   Turns out these correlate reasonably with wordnet senses:   If a word gets used in some particular way (e.g. idiomatically, or a collocation) chances are good there is only one meaning that is appropriate for that usage.

-- Linas

[Jen Seale]

Hi L,

Hope all is well! Apologies on the belated (and slightly relevant)
response here... I've been trying to differentiate between senses in
contexts that seem (possibly) to require class/type links, modifying
the dict to do so... I'm working on parsing a limited corpus, so this
may actually make some sense to implement.

Take the following examples, for instance:

1. It is just white.
2. The color is just white.
3. The color is just the white that I have have been thinking of.
3. He was just white with rage.
4. He was just white.
5. He was just the kind of white I was thinking he would be.

In these contexts, to delineate between the cases when 'white' is an
adjective, and when it is a noun, I was thinking of making two 'be
verbs'.
Be1: connects a noun that represents a class to a noun that is a kind
of that class
Be2: connects a noun to its attribute

Also, I was thinking of creating the following kind of links (using
full words for clarity sake):
'Color-' 'Type of color+'

So:
Be1: ('Color-' & 'Type of color+') or ('Type of color-' & 'Color
+').
White.adj: ('Type of color-' or 'Type of color+)
He.pronoun - no 'Color-' or 'Color+', so it will never connect to Be1
It.pronoun - would have to have all the class and type connectors,
also disjuncts for not types. X/ Further resolution/ambiguity
handling necessary.
'Color' and synonyms, etc.: ('Color-' or 'Color+')

All of this is very nascent, and I'm just beginning to work on the
'kind of white' scenario, but at very first blush:
of1: ('Color-' & 'Type of color+')

Anyhow, a simple problem I am facing (and perhaps what I should have
just said in the first place) is that I can't get words with new links
to parse. The 'primer' online seems to indicate that one may simply
create new links at will. New links I have created show up when I '!!
word' in the shell, but, again... no parses. Is there something I am
missing?

Best!
- J

On Mar 26, 12:26 pm, Linas Vepstas <linasveps...@gmail.com> wrote:

[Jen Seale]

Important Correction:

> Be1:  ('Color-'  & 'Type of color+') or ('Type of color-'  & 'Color
> +').

> White.noun: ('Type of color-' or 'Type of color+')
> He - no 'Color-' or 'Color+', so it will never connect to Be1
> It - would have to have all the class and type connectors,

> also disjuncts for not types.  X/  Further resolution/ambiguity
> handling necessary.
> 'Color' and synonyms, etc.: ('Color-' or 'Color+')

[Linas Vepstas]

Hi,

Sounds reasonable.  I've noticed before that colors are very oddball words; they claim to be adjectives, yet are frequently used as if they were nouns.  Its a strange situation, and I don't know enough linguistics to think clearly about the 'right' answer.

Or 'best' answer... 

  

Anyhow, a simple problem I am facing (and perhaps what I should have
just said in the first place) is that I can't get words with new links
to parse.

 Yeah. Been there, done that.

 

The 'primer' online seems to indicate that one may simply
create new links at will. New links I have created show up when I '!!
word' in the shell, but, again... no parses. Is there something I am
missing?

Every now and then, I trip over this problem and get very frustrated.  Typical causes are:  

-- I messed up the order of the links viz  A+ & B+ versus B+ & A+

-- The test sentence cannot possibly link to the left-wall.   This is a silent failure, that can drive you nuts.  Make sure some word in the sentence can link to wall, and make sure there are no crossing links that clash with this  Turn on !wall to make sure.  This is you most-likely problem.

-- The test sentence is being rejected by post-processing.  Turn on !bad to see the rejected linkages.

My suggestion: start very simple: create two new words: e.g. 

blorp: CCC+ & Wd-;      % pretend blorp is "white"

asdf: CCC-;                    % pretend asdf is "be"

and go from there. You'll quickly see which change it was that broke things.

-- Linas

[Linas Vepstas]

Hi,

On 2 May 2013 12:45, Jen Seale <jennife...@gmail.com> wrote:

Important Correction:

> Be1:  ('Color-'  & 'Type of color+') or ('Type of color-'  & 'Color
> +').
> White.noun: ('Type of color-' or 'Type of color+')
> He - no 'Color-' or 'Color+', so it will never connect to Be1
> It - would have to have all the class and type connectors,
> also disjuncts for not types.  X/  Further resolution/ambiguity
> handling necessary.
> 'Color' and synonyms, etc.: ('Color-' or 'Color+')

Avoid using capitalized words.  Link-grammar has fairly complex processing for capitalization, trying to make sure such words are either at the start of a sentence, or are proper nouns, or follow a colon ...  trying to cheat will just cause a mess.

The new link should be all-upper-case.  I suggest CLR+ and CLR-  There is some very special processing for mating together link types with mixed upper/lower names, including wild-card matching.   Don't use lower-case until you understand how it works, and are at the point where you really need to  enforce some more subtle linkage rules.

-- Linas

[Jen Seale]

Thank you, Linas! That was very helpful. Making viable links now. -
Jen

On May 2, 10:35 pm, Linas Vepstas <linasveps...@gmail.com> wrote:
> Hi,
>

[Linas Vepstas]

I can't tell if you plan to submit your changes back to the main project.  I encourage submissions and fixes: these benefit both you (when new versions come out) and everyone else.

However, for the kind of changes you are proposing, I may be very wary; I would likely reject large, unexplained changes. Changes are easier to deal with when they are a sequence of patches: each change is something small, easy to verify, easy to understand.  So once you think you have something working and submittable, then let me know, and we can review and discuss.  If you plan additional, ongoing work, then I cannot overstate the importance of using some change-management tool.  In this case, I would recommend using 'quilt'  to maintain a sequence of patches.  quilt is ideal for maintaining a stack of patches that are going to be emailed about.

-- linas

--
You received this message because you are subscribed to the Google Groups "link-grammar" group.
To unsubscribe from this group and stop receiving emails from it, send an email to link-grammar...@googlegroups.com.
To post to this group, send email to link-g...@googlegroups.com.
Visit this group at http://groups.google.com/group/link-grammar?hl=en-US.
For more options, visit https://groups.google.com/groups/opt_out.

[Jen Seale]

Ah, yes! I'm not planning on submitting these changes to the main
project. The one described above is pretty highly experimental! If I
hit upon something more like what you are describing, I'll let you
know. Best, Jen

On May 5, 1:35 pm, Linas Vepstas <linasveps...@gmail.com> wrote:
> I can't tell if you plan to submit your changes back to the main project.
>  I encourage submissions and fixes: these benefit both you (when new
> versions come out) and everyone else.
>
> However, for the kind of changes you are proposing, I may be very wary; I
> would likely reject large, unexplained changes. Changes are easier to deal
> with when they are a sequence of patches: each change is something small,
> easy to verify, easy to understand.  So once you think you have something
> working and submittable, then let me know, and we can review and discuss.
>  If you plan additional, ongoing work, then I cannot overstate the
> importance of using some change-management tool.  In this case, I would
> recommend using 'quilt'  to maintain a sequence of patches.  quilt is ideal
> for maintaining a stack of patches that are going to be emailed about.
>
> -- linas
>

[Linas Vepstas]

Yes, please.

Yes. What your described really does sound experimental, (i.e. plenty of ways to do it wrong) and so that did make me worried; but under the philosophy of 'share alike', if you do get something reasonable and working, I would like to be able to incorporate it .. if for no other reason than if would make it easier for you when new version come out.  So if you get to the point of something that works, please do share.

If you were in Austin, I'd even suggest meeting ... but perhaps some other day, then.

-- Linas