International Conference on Dependency Linguistics in Prague

The second International Conference on Dependency Linguistics was held in Prague, 27-30 August. There were a total of 31 talks, six of them very short, but supplemented by posters that were available for viewing during the whole conference. I think all of the papers were presented by a single speaker, although more than half of them had at least two authors.  Two of them had six authors, and one five. In most of these cases, only one of the authors was at the conference, perhaps because funding was tight.

The invited speakers were Dick Hudson and Aravind Joshi. Hudson has of course been working on dependency grammar longer than anyone outside of Russia and has quite a few books on his own version of dependency grammar in print. Joshi is more of a computational linguist than a theoretician and is known as the creator of Tree-Adjoining Grammar (TAG). There are quite a few websites about this, and it appears that the idea of TAG can be applied to derive DG representations as well as PSG representations.  

The conference proceedings, edited by Eva Hajičová, Kim Gerdes, and Leo Wanner (who were also the main conference organizers), were prepared in advance and distributed to the participants as they arrived. The book has a total of 33 papers, so there are two, one by Petr Homola and Matt Coler and one by Amba Kulkarni, which were not presented at the conference at all. Tim Osborne presented three papers, one of them a five-minute one. The team of Anna Nedoluzhko and Jiří Mirovský presented two papers; I think each of them did one of them.

The above numbers aren’t very impressive; I think the first international dependency conference had more speakers, and the conference book is about 20 pages longer. But the range of places the speakers came from this time is more noteworthy. There were speakers from China, Korea, Turkey, Finland, the United States, and especially India, as well as quite a few European countries.  

Lunch was provided to all participants. There was a dinner on the final evening at a restaurant next to the Vitava River, about a mile from the conference hall.   

The conference overlapped with the conference on Meaning-Text Translation, which started on the last morning of the DG conference and continued for another day or so afterwards.

The next International DG conference is planned for Upsala, Sweden.  Joakim Nivre, one of the invited speakers at the first conference, has consented to be at least one of the organizers.

The proceedings for both DG conferences are still available on the internet. Just type in: depling 2013 or depling 2011 to get to them.

dependency in morphology

It has long been recognized that there are many parallells between syntax and morphology. For this reason, the concept of ‘morphosyntax’ to refer to the intersection of the two, or perhaps in some cases to analyses which could not be clearly classified as either syntax or morphology.

Talmy Givon, for example, has argued that morphology is fossilized syntax.

In the paper cited at the end of this blog, I have argued that relationships between morphs can be characterized in the same way as relationships between words -- using the concept of dependency. Consider the following assumptions:

1. Morphemes have grammatical categories.

2. Morphs within a word are related to each other by the concept of dependency and this relationship can be modeled as a dependency tree.

3. A dependent of a given morph can be either its complement or an adjunct.

4. The category of the parent morph determines the category of the entire word.

5. Predicate-argument structures reflectng the way the meanings of the individual morphs combine with each other can be created on the basis of the dependency tree for each word.   

I will demonstrate these assumptions in my analysis of the word inevitability.

                                              NOUN: ity

                                             /

                         ______ADJ: abil

                       /                /

           MOD:in   VERB: evit

                        

Semantics: ity (un(abil(evit)))

The predicate argument structure of the above can be created using two rules also used  for syntax:

downward semantics: (i) the SEM of a complement becomes the argument of the SEM of its head.

So the SEM of abil becomes the argument of the SEM of ity   ity(abil) .  Likewise, the SEM of reli becomes the argument of the SEM of abil   abil(evit).  Combining these two applications of downward semantics, we get  ity(abil(evit)).  

This would give us the analysis of evitablity, but of course the constituent which is the child of ity would in that case not include the prefix in.

upward semantics: (ii) the SEM of an adjunct takes the SEM structure consisting of the SEM the SEM its parent and the SEMs of any complements of this  parent as its argument.

in is an adjunct of its parent abil. So its SEM takes the combination of the SEM of abil and the SEM of its complement reli as its argument. This gives us:  in(abil(evit)), which is also the SEM structure of inevitable This entire unit combines with the SEM of ity, as in the discussion after (i), to produce the SEM of the entire word.

When we talk about the inevitability of X, we mean that X cannot be avoided.  The sentence and the word both have almost the same predicate-argument structure.  The differences are that:

(i) the avoider, expressed as the agent of can avoid in the syntax, cannot be easily linked to the word structure of inevitability  

(ii) can replaces abil

(iii) avoid replaces evit

(iv)  not replaces in

(v)  the syntax refers to a state of affairs, whereas the word refers to this state of affairs as a property (ity) of the overall situation under discussion.
(vi) there is no overt passive morheme in the word. 

There is no full word verb evit in English (although there are full word equivalents in Esperanto(eviti), French (e'viter), Italian (evitare) and Spanish (evitar), at least), so we can use the verb avoid to translate it.  

For a fuller exposition of all this, with examples in a variety of languages, and also some more speculative ideas on this topic, see:

maxwell, dan, 2003: the concept of dependency in morphology. Agel, Vilmos, and Eroms, Hans-Werner. dependency and valency: an international handbook of contemporary research. 678-684. Berlin: de Gruyter.

The volume of articles cited above, as well as the accompanying volumes, seems to be well-known in Europe, but less so in the rest of the world, perhaps because it costs more than $600.  That is probably beyond the budget of most individual scholars, but university libraries are more likely to have it.  The individual paper can be purchased at the De Gruyter website for 30 euros or $42, however.

             

more on citations lists

It occurred to me after my last post that even a search engine like Microsoft Academic might in some cases find titles that had little to do with the topic intended by the searcher.  So I took a look at some of the titles that the search came up with.  The main new finding is that "principles and parameters" does indeed come up with quite a few titles that have nothing to do with that framework in linguistics. This seems to be because this title does not include the word "grammar", which would filter out articles that had nothing to do with grammar. So of the 200 articles that the search came up with for these words, somewhere between 130 and 140 of them actually have something to do with the linguistics framework of that name. Here is the breakdown by academic field.

Arts and Humanities: 44/45
Computer Science: 21/32
Biology 1/7
Engineering 0/21
Medicine 12/12
Physics 0/7
Multidisciplinary 53/71
Social Science: 1/2

This does not seem to be true for the other frameworks I checked on, since there titles all do include the word "grammar" or "syntax".  I didn't check all the titles, but did check several pages for some of them and didn't find a single publication that was clearly unrelated to the linguistic framework that it was found under.

Other topics that might be investigated relative to citations lists:

1. The number of citations for each publication.

2. Which journals turn up in these lists?

citations lists

I'm not sure what citations lists are used in the academic world these days, but there are several options on the internet. The one with the biggest numbers and presumably the least accurate is simply Google.  The Wikipedia article on citations lists warns against putting faith in general search engines like Google for searches on academic topics, simply because the numbers get inflated by factors which have nothing to do with your topic.  Still, it is interesting to see what Google gives for various linguistic frameworks.  Notice that the quotation marks should be part of the item searched.  The numbers given below vary slightly from day to day, but seem to be generally fairly stable.

"dependency grammar"   62,100
"construction grammar"  111,000
"principles and parameters"  96,200
"head-driven phrase structure grammar"  60,300
"lexical functional grammar" 88,700
"minimalist syntax"  49,800

Now what happens when we search on an engine designed for academic topics, like "Microsoft Academic"?   One difference is that the numbers are much smaller.  Another is that a breakdown in provided according to academic area.  A distinction is made between "publications" and "citations". If quotation marks are used, only numbers for publications are given. Here is what I got today:

"dependency grammar"  324  (computer science: 262, arts and humanities: 12,...)
"construction grammar"  191 (computer science: 52 , arts and humanities: 39,...)
"principles and parameters" 200 (computer science: 32, arts and humanities: 45,...)
"head-driven phrase structure" 214 (computer science: 114, arts and humanties: 23,...)
"lexical functional grammar" 229 (computer science: 139, arts and humanties: 21,...)
 "minimalist syntax" 37 (computer science: 1, arts and humanities: 15,...)

I also did a search in Microsoft Academic without quotations.  In that case, numbers are given for both publications and citations.  there is sometimes (for dependency grammar, construction grammar, and lexical functional grammar) also a graph, showing the increase in the numbers of citations over the years.  The numbers come out only slightly larger in most cases, but go up into the tens of thousands for head-driven phrase structure grammar and principles and parameters.   I don't know how to interpret that.

Chapter from 1995 text on DG now available online

In 1993-94, I wrote a booklength text about dependency grammar and called it "Unification Dependency Grammar". This appeared on two websites, first in 1995 in the Ohio-State linguistics department website, then about two years later also on the Dependency Grammar website. This was housed first with some sort of computational linguistics project in Stuttgart, Germany, then at the Charles University in Prague. I think it was sponsored by the linguistics department, or at least in a department which employed several linguists.  Various other work on dependency grammar also appeared on this website. For several years, anything on the website could be downloaded.  But, as already reported here in an earlier blog, the German student with the technical expertise to maintain the website went back to Germany in about 2001, and by 2004 or so, it was no longer possible to download anything from it. Somewhere around this time, my text also disappeared from the website of the Ohio-State linguistics department. This was a blow to the DG movement, since information about DG was not so easy to find.  In particular, there was almost nothing about DG used as a form of generative grammar outside of Hudson's and Mel'c'uk's work.

I have now been able to make the first chapter of the text on UDG available at: 

https://docs.google.com/document/pub?id=1-wPfZ-34JP_g7OgglQJ7cL_uf9SbY93AHQf5HCca_YQ

This took a fair amount of work, since I had only a paper copy of this text available. I had to scan it as a PDF file, convert this to a Word Document, and then copy it into my Google Drive. The main problem was that all diagrams came out quite garbled. Although Google drive has no tools for drawing trees, I was eventually able to make them all legible using the same tools that I used in the original text (forward slash, backslash, vertical line, underscore, and hyphen -- all right from the keyboard).

The chapter available from the above link has nothing to say about UDG (which I now call DCxG: "Dependency Construction Grammar"), but rather makes a number of general observations about DG and compares DG trees to corresponding constituency structure trees. In recreating this chapter, I changed almost nothing involving the content of the original text. I do remember removing one footnote which no longer seemed correct. Also, the final page of the chapter seems to have been lost, so I had to reconstruct what I said there as much as possible.   

Dependency Structure and Semantic Structure

One of the most important advantages of dependency grammar over constituency grammar is the relative ease of linking the syntactic tree of any sentence to a logical expression representing the meaning of that sentence.  In this blog, I will present the essentials of how this works. Let us call the logical expressions representing the meaning of any sentence its SEMANTIC structure, or SEM structure for short. A two formative sentence consisting of two simple words or one word consisting of two morphemes will have a SEM structure of exactly two parts: probably one predicate and one argument. If P is the predicate and A the argument, then it will be written as follows:  P(A).The DG tree of this sentence will have exactly two nodes. Here are some examples with corresponding DG trees. For the purposes of this discussion, I am using the Osborne-Gross style of dependency trees, although I omit the dotted lines from the nodes to the linear positions of the words (see previous two blogs).

Come here!    (b)    In Paris            (c) I do           (d)   venu! (Esperanto for “come!”

Come                  in                                 do                      -u ( volative marker)

      \                          \                            /                       /            

       here                 Paris                  I                      ven  (come)

The corresponding SEM structures, where the words in parentheses are the meanings of the corresponding words, are as follows:

     Here (come)        in(Paris)                    do(I)              u(ven)

What makes the algorithm simpler is that in a DG tree there are no phrasal nodes.  Phrasal nodes are a complication for any algorithm which creates a logical expression.  Any constituency structure tree will have at least one phrasal node in addition to the two lexical nodes  The lexical nodes are always at the bottom of the tree and the phrasal node uniting them at the top. The ones with a verb might have both a S node and a VP node, as well as the two lexical nodes.  In expressions like these with only two formatives, it might be possible for the algorithm to ignore any phrasal nodes and say that any nominal element is the argument and the other its predicate.

But if the sentence is somewhat longer this strategy becomes more complicated,  because there seems to be no good way to specify which nominal element becomes the argument of which predicate and also what the relationships among the larger units thus created are.  Consider the following sentence, for example.

The student stood on the table.

Given the following rules based only on syntactic categories and linear order:

(i)     A determiner is a predicate taking the following noun as its argument.

(ii)   A preposition is a predicate taking the following noun as its argument.

      (iii) A noun and its modifiers are the argument of the following verb.

(iv)  A prepositional phrase following a verb is a predicate taking the verb and its complements as its argument.

It would be possible on the basis of these rules to construct the following predicate argument structure (the numbers below the formula aim to make it easier to see which closing parenthesis corresponds with which opening parenthesis):

( the (student) ) stood (on (the (table)  ) )

1      2       -2  -1          3    4      5    -5-4-3

But it is clear these rules do not make any use of the phrasal nodes in the constituency structure tree of this sentence.  Furthermore, these rules are rather specific and idiosyncratic. There would have to be a lot more of them to get a way of creating predicate-argument structures for a moderately wider range of English sentences.

But if we create a dependency tree for the same sentence, we need only two very general rules.

                           stood

                    /                    \

     student                            on

      /                                               \

The                                                      table

                                                          /

                                                   the

The rules are:

(i)     A complement is the argument of its predicate parent

 (ii) an adjunct is a predicate taking its parent (and any arguments the parent may have) as its argument.                                        

We can use these rules step by step to get the predicate argument structure above.

A.    ‘The’ is an adjunct taking its parent noun as its argument, giving us ‘the(student)’ and ‘the (table)’ by rule (ii)

B.      ‘The table’ is the argument of its parent ‘on’ , giving ‘on(the (table))’ by rule (i)

C.     ‘the student’ and ‘on the table’ are arguments of their parent predicate ‘stood’ again by rule (i).  

Now suppose we add the adverb ‘quietly’ to the above sentence. This gives:

The student stood quietly on the table.

or

The student quietly stood on the table

or

The student stood on the table quietly

It doesn’t matter which of these versions we use. The adverb is a child of the verb ‘stood’ in all cases.  And by rule (ii), this adverb is a predicate which takes the rest of the sentence as an argument:

quietly ( (the (student)  )    stood (on (the (table )   )   )   )

          1 2     3       -3 -2            4    5     6     -6 -5 -4 -1         

WORD ORDER IN DEPENDENCY GRAMMAR

WORD ORDER IN DEPENDENCY GRAMMAR (II)

 If the word order of a given construction is fixed, as it usually is in English and similar languages, then there is no problem making the connection between the rule and its realizations in specific sentences. The rule simply specifies the order required and (parts of) the specific sentences replicate it.  So, for example, the rule for the definite article would specify that it appears to the left of its head noun (or dependent noun, for those who favor the analysis in which determiners are heads).  

But what happens in constructions which allow more than one linear order of head and dependent?

In English, for example, many adverbs can either precede or follow the verb they modify.  In Russian, the order of subject, verb, and object in a transitive sentence is said to be free:  any one of the six logically possible orders can occur in a given sentence.  All the inflections are also the same, so they apparently all involve the same construction, then it appears that the rule must allow for both/all the possible orders, even though only one of them is chosen in any given sentence.   

Until I became familiar with the work of Osborne and Gross, my way of doing this was to simply omit the linearization markers ‘<’ and/or ‘>’  in the rule.  Without these restrictions, the rule is interpreted to mean that any linear order of the nodes in the rule is possible, at least for two generational trees. 

Osborne and Gross deal with trees for whole sentences, but not the rules for generating them, so I was left to my own resources about how to resolve the following dilemma:  if no linearization markers are used in the representation of the linear order of nodes in a dependency tree,  as in the work of Osborne and Gross, then there is nothing to omit in the rules involving free word order.  At first I took this fact to be evidence that I should stick with my own approach to linearization, since there seems to be no way to generalize the Osborne-Gross approach to deal with free word order.  

An alternative would be to allow linearization markers be used only to specify the order of words in the realization of constructions involving free word order.  Realizations of constructions involving strict word order could use the Osborne-Gross system of specifying linear order on the basis of the position of the dependents relative to the head. 

Another alternative is to assume that word order is never really free, once semantic and pragmaticl factors are taken into account. Once these are specified in the rule, then it may well turn out that the linear order of the words is entirely determined.   This would allow the Osborne-Gross approach to be used everywhere.  To take the example of Russian transitive sentences mentioned above, one might hypothesize that the word representing old information is always first and the one representing new information is always last.

This idea, that syntactically free word order is not free once other factors are considered, has to be an assumption at this point, since our understanding of these factors at present is probably not sufficient to formulate it as a specific testable hypothesis.  But there are many cases in which it is clearly true, and I don’t know of any clear case in which it is clearly not true.  But since the work of Pollard and Sag 1986, it has been a goal of many syntacticians to incorporate all of these factors into our rules.  If this can be done, the above hypothesis becomes testable.