Talk:Dependency grammar

Random Billboard
The sections on "Implementations" and "External Links" primarily seem to be a billboard promoting random scientific work stemming from both the computational and linguistic realms. Essentially everything under implementations appears to be data-driven parser generators that are completely agnostic to underlying grammar formalisms. External links mostly hold links to randomly chosen treebanks (UD being an exception, and most likely deserving its own section in the article). imo they should all be truncated and, where it make sense, be incorporated into sections of the overall DG article that motivate the links.79.204.248.31 (talk) 00:43, 27 February 2016 (UTC)

hello pls check the .pnj file at "constituents" page. the second from the end is added by mistake. the correct png is found on this page. — Preceding unsigned comment added by 77.127.219.254 (talk) 00:34, 24 February 2021 (UTC)

LG demo removal
Is there a reason Ddxc removed the LG demo link? –jonsafari (talk) 04:00, 14 January 2008 (UTC)

New article is a copy from ... ?
This article has been completely replaced by text that has been obviously copied from somewhere else (note the missing inline images and the textual references to these!). The text was added at the very end of August 2009 by User:Tjo3ya. I am slightly concerned about WP:COPYVIO. I would like to know who the original author of the text is! linas (talk) 04:14, 9 September 2009 (UTC)
 * Yeah, the references to images and their relative position to text is a tell tale sign of copyright infringement. I'm not able to find the text in Google books though. But many books are yet not indexed there. I think this article should be reverted to the stub version of June 11. Pcap ping  11:33, 11 September 2009 (UTC)

Suggested staring point for anyone game to tackle the rewrite
Hi, I'm not up to the task myself but if anyone else is theres some good online material in the document http://www.infoamerica.org/documentos_pdf/bar03.pdf that covers Dependency grammar history from medieval times through to computer science. The current article doesn't even hint at the origins.

Are the sentence diagrams in the article correct?
I'm no expert on this topic but if a head word is supposed to define the role the words that depend upon the head word, why in the world is an auxiliary verb being used as a head word in nearly all of the examples? The word that determines the roles of the various phrases in the sentence is the _semantic_ portion of the verb, not the auxilary verb, i.e. in "We are trying to understand this sentence". "are" (the auxiliary verb) has no control over the other phrases that can be present - "are trying" is a modifier on the verb "understand". It determines timeframe and aspect of "understand", i.e. it tells use whether we talking about something we plan to do, are currently doing, are beginning to do, are continuing to do, are in the process of completing, have completed.

On the other hand "understand" determines whether we can use bare nouns as a verbal complement or need to insert a bunch of prepositions. The test of this is that I can replace "are trying" with a variety of time/aspect/mood modifiers (modals, auxiliary verbs) and the phrases "we" and "the sentence" will not change their structure. On the other hand, if we replace "understand" then we may or may not be able to keep "we" and "the sentence". For example, if we replaced "understand" with "go" then we would need to prefix any verbal complement with a preposition "to, towards, from, away from, etc etc".

If I were diagramming that sentence I would do something more like this:

understand -- we           are trying to         the sentence (subject)    |    |                | are trying           sentence |               |                    to               the ( time/aspect       (verbal complement)                 signifier )

Hello ???,

The criteria you are using to determine dependencies are semantic in nature. Most work in modern dependency grammars is concerned primarily with syntactic criteria, however. Both types of dependencies (semantic and syntactic) are certainly valid, but again, most work in modern dependency grammar is focusing primarily on syntactic dependencies.

The dependency trees in this article and in other articles in Wikipedia show syntactic dependencies. These dependencies are supported by the results of standard constituency tests. Any complete subtree of a dependency tree is a constituent, and most of these complete subtrees are positively identified as constituents by constituency tests. Take the example sentence that you reproduce here. When the finite verb, the auxiliary are, is the root of the entire tree as shown in the article, the subject pronoun we, and the non-finite verb phrase trying to understand the sentence are constituents (= complete subtrees). The results of many constituency tests support this understanding of the structure. For instance, they positively identify the non-finite verb phrase trying to understand the sentence as a constituent:


 * a. ...and trying to understand the sentence, we (certainly) are. - Topicalization
 * b. What we are doing is trying to understand the sentence. - Pseudoclefting.
 * c. We are doing that. (doing that = trying to understand the sentence) - Proform substitution
 * c. What are you doing? - Trying to understand the sentence? - Answer fragment

These four tests all support the analysis in the article, where the auxiliary verb are is the root of the entire sentence structure, which renders the non-finite verb phrase trying to understand the sentence a constituent.

Compare this situation with the structure you suggest. Your tree shows are trying to as a constituent. No constituency test (with the one exception of coordination, perhaps) identifies are try to as a constituent, a fact that you may test for yourself. Furthermore, few if any modern grammars (be they constituency- or dependency-based) view are trying to as a constituent.

I suggest also reading the articles on the verb phrase and immediate constituent analysis. The empirical evidence supporting the finite verb as the syntactic root of clause structure is strong.

Finally, I can comment that a section should be added to this article that mentions the distinction between semantic and syntactic dependencies. I have been intending to do this for awhile, but have not yet gotten around to it. --Tjo3ya (talk) 20:20, 25 July 2012 (UTC)

This makes no sense, Tjo3ya. Also, you just assert your position as a postulate without explaining why it should be adopting. In the example, understand is the verb indicating the process/action that actually happens. Consider the following:
 * We understand the the sentence.
 * We are understanding the sentence.
 * We try to understand the the sentence.
 * We are trying to understand the sentence.

In all such cases, the verb is understand. The subject we and the object the sentence both depend on understand (not on are or try). Your view, again presented without any justification, seems to rest on "prehistoric", "blind", "nonsensical" grammatical analysis of the kind abandoned for ages, as far as I know (are is not the real verb just because it is conjugated, lol!). Modern grammar approaches try instead to make sense. denis &#39;spir&#39; (talk) 10:08, 28 March 2014 (UTC)

Thank you for your reply
Thank-you for the reply. But I really am not following your line of reasoning here.

(a) why do you conclude the analysis is semantic? The labels on that diagram identify syntactic roles, not semantic content. Which words belong to which syntactic roles isn't a semantic issue either. A combination of phonetic properties, sub-morphemes, and part of speech tagging can be used to combine words into phrases and to define the role of each phrase in the sentence -- as amply demonstrated by our ability to comprehend the nonsense poem "The Jaberwocky".

(b) Any diagram that could be transformed algorithmically to create those four sentences would count as a head by the standard you just presented. With two small modifications to my diagram and no change in head word, it would be trivial to construct an algorithm to make those four sentences using "understand" as a head word. All you have to do is move the parenthesized labels an add an "aux verb" label to the word "are"

(head) | understand - |            |                      | (subject)  (time/aspect)          (verbal     ) |       (signifier  )          (complement ) |            |                      |   we       --           sentence |            |             |         (aux verb)  (w/o aux verb)   the |            |            are         trying |                       to

With those small changes your four sentences become simple substitutions:

(i) Topicalization: and [time aspect signifier minus aux verb] [head] [verbal complement], [subject] [aux verb] (ii) Pseudoclefting: What we are [time aspect signifier] do is [head] [verbal complement] (iii) Proform substitution: [subject] [aux verb] doing that. (iv) Answer fragment: What are you doing? [time aspect signifier minus aux verb ] [head] [verbal complement]

(c) Isn't your choice of transformational tests assuming a priori that this sentence as a copula with "trying" acting as a modifier? In that case all time/aspect/mood alterations only involve the verb "to be". However, there is another reading: that of progressive present.

English is inherently ambiguous about whether an "ing" word is acting as an adjective or the second half of a progressive present verb. Whatever head we chose therefore has to be associated with an algorithm/parse tree that works with both readings. If we view this as a progressive present verb then we also need to be able to support transformations of mood/aspect/timeframe on the verb "try" and not just the verb "are". Many of these transformations cause the verb "are" to go away but they never cause the verb "try" to disappear. On the other hand, if we view this as a copula these transformations both "are" and "try" are persistent. Therefore the phrase "try to understand" as the head is compatible with both readings but "are" as the head word is not.

For example, to change this sentence to an optative mood we'd get "We {would like|want|wish} to try to understand the sentence". The word "are" disappears completely from the sentence and is replaced by "would like to". Change it to simple past you get "We tried to understand the sentence". "are" also disappears, but this time it isn't even replaced by anything. Instead we get an inflected form of "tried".

(d) It was my understanding that one of the goals of dependency grammar was to adapt the insights of functional grammar to the more structured needs of computational linguistics. Languages have such a wide variety of ways of representing the notions of timeframe/aspect/mood. If you have any intention of using your parse tree as the basis for automated translation, it is a dangerous strategy to make standalone words that set timeframe/aspect/mood into top nodes in your tree. You end up with source languages missing nodes that the target language needs or target languages needing to insert null nodes for the nodes that simply aren't relevant to their particular combination of syntax/morphology just to avoid losing a whole branch of the tree.

Some languages use auxiliary verbs and some don't. If you make an auxiliary verb the head simply because it is the carrier of the subject inflection, you will end up with different heads when you useo a language that doesn't use auxilary verbs as the starting point for your parse tree. "am walking" in English is one word in Latin (ambulo), German (laufe or gehe depending on dialect) and one word in Hebrew (holekh or holekhet depending on the gender of the speaker). Swahili also uses one word "ninakwenda". In Hebrew, Latin, German, and Swahili, "are" would never appear as the head because it doesn't exist in the sentence.

If you try to claim "well, I can still parse out morphemes" that are equivalent to "am" and get the head that way. I'd say sometimes yes, sometimes no. German uses a single verb form for both simple present and present progressive. If you really want to stress that you mean the progressive sense you use an adverb like "jetzt". Sure there is a second word, but it isn't the one that carries the subject inflection the way "am" does. Hebrew also uses a single verb form for the simple present and present progressive. But when it wants to clarify it means ongoing action it uses a very different strategy : it converts walk to an infinitive and makes it the complement of an inflected verb. "holekh" means either "I walk" or "I am walking". To stress that walking is ongoing you say "Mamshikh=continue lalekehet=to walk". Mamshikh is no more the head word than "jetzt" is even though it is clearly the inflected word. It is a modifier on the lemma for "walk".

Or consider the conversion of an intransitive verb to a transitive verb with a third party actor. In English we convert a verb to its causitive sense by prefixing it with the inflected word "made" ( He ate ---> I made him eat ). In Hebrew you quite often modify the word stem and convert the verb from a Qal/Piel stem to a Hifil stem by adding the prefix "ha" and making certain internal vowel changes. Even though "try" is the word that gets inflected in the simple past in English and all tenses in Hebrew, I suggested "try" was a modifier on "understand" because its role is to frame the core activity/state discussed by the sentence. A few weeks before I read this Wikipedia article, I ran across an article on a tribal language in the Americas that used word stem to indicate effort in a manner that looked a lot like the causative inflections that I just mentioned in Hebrew. So it is possible that one language's "trying + infinitive" is captured in another language with a single lemma (root + "trying" morpheme).

(e) You seem to be defining finite verb in a manner that is language dependent and then trying to claim cross linguistic significance.

If by finite verb we mean the lemma that identifies state/action discussed by the sentence once we have factored out all syntactic/morphological elements meant to communicate mood/timeframe/aspect in a pro-forma fashion, then I would agree the evidence weights strongly in favor of this lemma controlling the sentence structure. The number of complements, particular combination of case markings, prepositions, post-positions, affixes, word order and other sorts of noun phrase role markers typically is determined by this lemma.

This lemma and the structure it creates is also permissive of a great many transformations. Most languages support a fairly standardized set of transformations between different timeframe/mood/aspect. You may need to know the morphological class of a word, to effect these transformations but you don't need to know the meaning. Throughout all these changes the selection of noun phrases, role markers, and verb lemma will stay the same even when the syntax and number of morphemes related to timeframe/mood/aspect varies wildly.

But if you are equating "are" with the finite verb then that is clearly NOT what you mean. Instead you seem to mean the word with subject inflections.

But if I may step back a bit, how would that work cross linguistically? If you are parsing a sentence how do you identify the subject? In ergative languages, there is a nominative case, but stative verbs agree with a word that has accusative case markings. In Swahili a verb can agree with the location where action takes place and not just the actor or the things affected. Most English speakers would not consider a location (adverb) a valid verbal subject and have trouble even conceiving of how that would work in Swahili.

Many languages inflect adjectives as well as verbs. These can agree with the same word as the verb. How do you tell which is verb and which is adjective? Is it really just inflection? Hebrew words acting as present tense verbs, adjectives, and animate nouns all use the same inflectional endings an all agree solely with the number and gender of the subject - NOT person. You have to use position and other syntactical clues, not subject-word agreement to pick out the verb in the sentence.

What about sentences where there is NO agreement at all? There's no subject inflection on any word in an English sentence that uses a modal, e.g. "I could run". What word would you then chose? Modern Swahili copulas have a variety of particles they use in place of the English "am/are/is". But they aren't inflected. No subject markers, no tense markers. Nada.

And what if a non-inflected or less inflected word controls the form of an inflected word? Consider French: in the compound past (aux verb + participle) the auxiliary verb agrees with person and number whereas the participle agrees only with gender. Since both agree with the subject, which do you chose? Maybe you should chose the auxiliary verb, since the auxiliary verb contains person as well as gender? But wait. You can't know what the auxiliary verb is without knowing the participles. In French the participle determines the choice of auxiliary verb (avoir, etre). Furthermore since both auxiliary verbs are irregular, you can't even guess what morphemes to use to represent person/number inflections on the auxiliary verb without the help of the participle. So now which is head and which is dependent?

It seems to me that equating finite verb with some sort of purely morphological criteria, does not survive well cross-linguistically. What works in one language is either too specific or too general to work in another language or simply irrelevant. It doesn't even work consistently within a language. Some English sentences have subject inflections and some don't. You need a higher level functional definition that is then operationalized in different ways. In one language it could be as simple as looking for a word with certain morphemes or submorphemes. In another language it might involve a multistage analysis that first uses various sorts of markers that classify sentence types and then uses other sentence type dependent clues to identify the word or words that compose the sentence's finite verb. Beth 87.68.215.176 (talk) 18:24, 1 August 2012 (UTC)

Three types of dependencies
Hello Beth,

Your message is long, too long, I think. It is difficult to see how to craft an appropriate response to the tremendous content. Let me summarize where I think the disagreement lies between our views. It boils down to this: I think you are failing to distinguish between semantic, syntactic, and morphological dependencies. For me, the following summaries hold:


 * 1. Semantic dependencies: determined by predicate-argument structures, the arguments of a predicate depend semantically on the predicate
 * 2. Syntactic dependencies: determined by distribution, the root word/morph of a given syntactic unit determines the distribution of that unit
 * 3. Morphological dependencies: when one word or part of a word influences the morphological form of another word, the latter depends morphologically on the former.

These dependency types can coincide, run opposite to each other, or be independent of each other entirely. If one fails to distinguish between them, unending confusion can be the result. Igor Mel'cuk's works have been insightful for establishing the necessity that one distinguish between these three types of dependencies. And again, most work in modern dependency grammars focuses primarily on syntactic dependencies.

There are relatively easy answers to some of your counterarguments. In particular, the string are trying to does qualify as a concrete unit of dependency syntax. It is a catena. By acknowledging catenae, I think one gains the ability to acknowledge as concrete syntactic units many of the word combinations you point to.

Your reluctance to position the auxiliary verb are as the root of all clause structure in your example runs counter to almost all work in modern syntax. This fact is true regardless of whether one chooses a dependency grammar or a constituency grammar.

Finally, my points here can be illustrated with examples from English or other languages, but these matters should be limited to a discussion of one or two examples at a time. --Tjo3ya (talk) 20:01, 1 August 2012 (UTC)

Reply
Tjo3ya writes --- Your message is long, too long... ---

So I will simplify: a) I see you defending a diagram of limited utility against a diagram with greater utility. b) It appears to you feel compelled to do this because of a fixed notion of what should be the head. This forces you to unnaturally separate an auxiliary verb from its participle and treat it as if it were no more part of the verb than "red" is part of the verb in the sentence "The ball is red with polka dots". The difference though is there is NO valid transformation along the lines of "The ball is red" to "The ball redded ...". but there is a valid transformation of "I am trying to ..." to "I tried to...". "is-red" and "is-trying" do not have the same behavior yet you insist it is valid to draw them as if they were identical.

Personally, I think showing a less powerful diagram at the top of the article doesn't do much at all to suggest to readers what dependency theory is capable of.

Tjo3ya writes --- Your reluctance to position the auxiliary verb are as the root of all clause structure in your example runs counter to almost all work in modern syntax. ---

Not that I see. What I see is variety:


 * Rambow et al. "A Dependency Treebank for English". Discusses dependency grammar research at U. Penn.  Please note that all the diagrams treat auxiliaries and modals as child nodes. (ftp://ling.upenn.edu/studentpapers/creswell/rambow-et-al.pdf )
 * Case and Gender: Concept Formation Betweeen Morphology and Syntax - http://books.google.co.il/books?id=5XaeNozadRMC ) :  Van Helden, p. 593 claims Tensier himself treated auxiliary verbs and prepositions as part of other nodes.

According to Van Helden ( pp. 675-676 ) the Russian school of dependency analysis has gone through various stages, first putting the subject at the head, then the whole verb phrase and then finally just the tense bearing portion at the head (Mel'cuk). But if you read on, it turns out that Van Helden's main point is simply that there are a lot of disputes about what should be the head.

This article, based on the work of Mel'cuk, also puts the aux at the head. http://acl.ldc.upenn.edu/P/P01/P01-1029.pdf. If you look at this article and the Rambow article you will see that (a) both are recent and both come out of one institution. If Upenn can tolerate two different views of what counts as the head of a dependency diagram, shouldn't Wikipedia articles represent the full range of viewpoints? Seems to me that the article ought to at least be discussing the fact that different theoreticians have different ideas about what should be the head.

Tjo3ya writes: --- The string are trying to does qualify as a concrete unit of dependency syntax. It is a catena.---

Agreed its a catena. But the diagram in the article does not indicate that. Graphically a catena is a subtree, not some arbitrary portion of a long chain of connected words. It has to be a subtree because otherwise there is no way to identify a unit to extract and transform.

Drawing the diagram as I did, with the string elements situated in a labeled subtree made it very easy to extract the phrase and manipulate it. The diagram supports all of the copula transformation, all of the tense/mood/aspect transformations for "is trying to", translations to several other languages that are structured in very different ways from Western European languages, and several other transformations as well: "I am trying to understand..." to "I do understand..." or "If only I understood".

By contrast the diagram in the article only supports transformations for a copula. There is no way to do the other transformations because the diagram has a single long line "am ..... the". There is no tagging or any other information that would allow a software program to isolate "am trying to" as a transformable element nor even "understand the sentence". Without that, there is no way to do any transformation other than the four copula transformations.You lose the ability to support the full range of transformations of which an actual human speaker would be capable.

Tjo3ya writes: --- I think you are failing to distinguish between semantic, syntactic, and morphological dependencies. --

Recognizing a distinction between dependency types won't solve the problem. All it does is allow you to let the morphological relationship between the French auxiliary and participle go one direction and the syntactic relationship go the other. The only thing that will solve the problem is a subtree that allows the extraction of "is trying to" as a unit.

Without that ability to extract a subtree you are limited in the kind of transformations you can do. In French you need to treat the aux + participle as a unit if you want to do the full range of tense transformations: simple present, simple past, future all use a single word in place of the two word combination "aux + participle". They may be written as two separate words, but they function as one.

The failure to couple the French auxiliary and participle also would result in incorrect translations between Hebrew and French. The Hebrew equivalent to "avoir/etre + participle" is a single word marked for person, number, and gender. The French auxiliary verb is only marked for person and number. The French participle is only marked for number and gender. But together they are marked for person, number, and gender just as the Hebrew verb. You can have an accurate translation from French to Hebrew and back again only if you take aux verb + participle as a unit even though it is written as two words.

Beth 87.68.215.176 (talk) 16:26, 6 August 2012 (UTC)

Constituency tests
Hello Beth,

Thanks for your interesting message. I appreciate the sources you have given to illustrate your point.

Rambow et al.'s analysis of dependency structure is challenged in major ways. The biggest difficulty for me with those structures is that the constituents (= complete subtrees) shown are NOT identified as constituents by a large majority of constituency tests. Take the example sentence The flight will have been booked from Figure 2. The tree shows each of the auxiliaries as a constituent, but let's focus on the finite auxiliary verb will.


 * The flight will have been booked.
 * a. *Will the flight have been booked. - Topicalization (unacceptable as a statement)
 * b. *It is will that the flight have been booked. - Clefting
 * c. *what the flight have been booked is will. - Pseudoclefting
 * d. *The flight does have been booked. - Proform substitution
 * e. What about the flight being booked? - *Will. - Answer fragment
 * f. *The flight have been booked. - Omission

The tests deliver consistent results. Based on these data, we can conclude that will is NOT a constituent. The exercise can be extended to the other two auxiliary verbs. Each of the auxiliaries alone is NOT identified as a constituent by the constituency tests, yet Rambow et al.'s analysis takes each of these words alone to be a constituent. Please note that these considerations are empirical. That is, I have produced an argument that is backed by empiricism. What empirical considerations can be produced to illustrate that we should view will and each off the other auxiliary verbs as constituents. If Rambow et al.'s analysis is to receive empirical support, concrete illustrations like this are needed, not vague claims about the ease of automatic tagging and translation in terms of transformations.

Concerning transformations, you seem to take it for granted that transformations occur. I think a majority of dependency grammars reject the concept of transformation as it is understood in derivational systems (e.g. TG, GB, MP). My particular stance is that a construction-based theory of syntax is the better approach.

A catena is any subtree, any subtree at all!. A complete subtree (= constituent) is a particular type of catena. See the article on catenae again. Every constituent is a catena, but there are many catenae that are not constituents. The point you make concerning simple vs. periphrastic forms is nicely accommodated in terms of catenae. The word combinations associated with periphrasis are catenae. See the article on periphrasis. While there may not at present be any software designed to acknowledge and manipulate catenae, it must be possible to produce applications that would do so. That can't be that difficult of a task for those who have the necessary programming knowledge.

Yes, Wikipedia should accommodate the minority stance. But let me emphasize in this regard that what you are arguing for is represented by a very small minority in theoretical linguistics. Almost all work in GB and MP positions auxiliary verbs above the main verb. Most dependency grammars do this as well - I'm thinking here in particular of the German schools such as Heringer (1996) and Eroms (2000). Word Grammar does it too, as does MTT, as you point out.

If you want to add a section to the article that points to the alternative analysis you propose, I will likely not attempt to remove it. I will, however, strive to include statements in the section to the effect that the analysis is not supported by empirical considerations such as the results of constituency tests. Furthermore, the section should include a tree illustrating the analysis.

Finally, I can point you to an interesting debate about the importance of constituency tests for our understanding of syntactic structure. See the at times heated exchange between me and Rjanag and Taivo here. --Tjo3ya (talk) 19:44, 6 August 2012 (UTC)


 * I need to qualify something I wrote yesterday. If you add a section demonstrating the analysis you prefer, I would not attempt to remove it assuming that it is supported by at least one or two prominent sources. The paper by Rambow et al. that you included above does not seem to appear in a peer-reviewed journal. Prominent sources are necessary to justify the presence of a section presenting the alternative analysis of sentence structure. If these sources exist, then I am actually in favor of including such a section, although again, I will likely strive to point out that the analysis is not supported by constituency tests. --Tjo3ya (talk) 19:57, 7 August 2012 (UTC)

Reply
Thank you again for your comments.

Tjo3ya says: --- A catena is any subtree, any subtree at all!. --

Of course it is, BUT

Mathematically a subtree is node X and all of its child nodes. Another equivalent way to state this is:  A subtree is Node X,  all of its leaf nodes, and all nodes that have to be traversed to reach the leaf nodes. Thus in the article's graph of "I am trying to understand the sentence" we have the following subtrees:


 * "am" = the entire tree
 * "I" = just the node "I" - it has no children
 * "trying" = "trying to understand sentence the"
 * "to" = "to understand sentence the"
 * "understand" = "understand sentence the"
 * "sentence" = "sentence the"
 * "the" = "the" – it has no children

No where in that list is something that ends with "trying" or "to". So how would you know to extract "am trying to" or even "am trying"?

The requirement of "all of its descendents" is needed because it gives the sub graph (= any connected subset of nodes in a graph) certain useful mathematical properties. One of the most important of these is that for any node in a directed acyclic graph there is one and only one subgraph that includes "all of its descendents". Thus the node's id is also the subgraph's id. No additional information is needed to know what nodes to extract.

But let's suppose I agree with you and say a catena=subtree can be any subgraph and not just the one mathemeticians call a subtree. Then you have another problem. If you remove the "all of its descendents" criteria then there isn't a neat way to identify the subgraph. Since you are allowing arbitrary exclusion of some descendent nodes but not others, you have to specify in your subgraph identifier not only the head node but a rule for including or excluding descendents. You can say "node X plus all children one to three levels deep" or "node X plus all chilren of the edges labeled "foo" and "bar" ) . There are as many rules as there are possibilities for selecting some children but not others.

The graph in the article does not provide any rationale for selecting some children of "am" and not others. So by either definition of a subtree, mathematical or otherwise, there is no basis for selecting out "am trying to" and treating it as a catena.

Tjo3ya says: --- The biggest difficulty for me with those structures is that the constituents (= complete subtrees) shown are NOT identified as constituents by a large majority of constituency tests. ---

The difficulty with Rambow's diagram has nothing to do with the choice of head, but rather his failure to give you enough information to isolate "will have been" and "will have" in subtrees. Once that's been done it is quite easy to algorithmically generate sentences that will pass Topicalization, Omission,  Passivation/Activation, and Answer Fragment tests.

booked flight        | the           | will |            have
 * been

Constituency tests for "will have been":


 * Topicalization: It [subtree=been] done, [node=booked] [subtree=flight] that is. => It will have been done, booked the flight, that is.
 * Passivation/Activation They [subtree=will] [node=booked] [subtree=flight]  =>   They will have booked the flight.
 * Answer fragment: Has the flight been booked?  It [subtree=been] by Tuesday =>  It will have been by Tuesday.
 * Omission:  This is tricky because you have to chose a mood that does not require a timeframe/aspect signifier.     One option is the command but passive verbs can't be turned into imperatives.  Turning the sentence into a subjunctive might do though:   Would that [subtree=flight] were [node=booked]!   =>  Would that the flight were booked.

Ultimately it boils down to this: a graph is only as useful as the information content it provides. This includes the end as well as the start of a catena.

No one is disputing the importance of constituency tests. They are a useful tool for seeing whether or not something functions as a unit. But they aren't always enough to see what is really going on in the language and how it functions. Languages do more than move around verbal complements and the tests you are considering as sufficient really do nothing more than shuffle and replace complements. Tense/state/mood conversions like "is trying" to "tried" or "will have been X" to "has been X" are just part of how we use English. They appear in the normal back and forth between individuals: Mom: "Will you clean your room?" Kid: "Mom, I already cleaned my room." Speech writers use them to retorical effect "I fought for lower taxes. I will always fight for lower taxes!". Therapists use them to build connections and create empathy, and so on.

Similarly, we can't just ignore cross-lingual issues. Conversions between languages aren't just for software: any bilingual or plurilingual person does them, especially if three or more conversation participants all are most comfortable in different languages. So do real time translators at meetings and at their leisure book translators. Models of language have to support these facts. This isn't a case of generative grammar vs. dependency grammar. There are simply some linguistic realities that need to be accounted for by any model. You don't have to be a nativist and assume some sort of internal neuronal transformation engine to acknowledge them as linguistic behavior.

As for the article itself: there are a lot more issues in my opinion than just one diagram. I've read a number of dependency grammar theory review articles recently and several of them do a much better job than this article at outlining the shared elements of a dependency grammar theory, the key debates it has engendered, the problems with early forms of tranformational/generative grammar they were trying to solve, the mathematical formalisms typically used and so on. When I have a bit more time, I'll be happy to put together a list of a few examples if you are interested.

Beth 77.124.109.209 (talk) 11:12, 9 August 2012 (UTC)

Subtrees
Hi Beth,

Thank you for the comments about subtrees. It appears that my use of the term subtree has been inconsistent with how the term is used in graph theory - although I will be checking with another source beyond Wikipedia for verification. What you state is a subtree in graph theory, I have been calling a "complete subtree". I am happy with this correction. This exchange is now profitable for me in this regard.

But in the area of constituency tests and what they tell us about constituent structure, I don't think we are making any progress. You seem to be producing an argument that is contrary to what has been established and widely accepted by the vast majority of work in theoretical syntax in the past 40 years. More importantly, your use of constituency tests baffles me. Checking the test string will have been in the example discussed above, this is what constituency tests tell us:


 * The flight will have been booked.
 * a. * Will have been the flight booked. - Topicalization
 * b. *It is will have been that the flight booked. - Clefting
 * c. *What the flight booked is will have been . - Pseudoclefting
 * d. *The flight does (so) booked. - Proform substitution.
 * e. What has occurred with the flight and booking? - * Will have been . - Answer fragment
 * f. *The flight booked. - Omission

Based on these results, we have no reason to view will have been as a constituent, contrary to what you seem to claim above. My use of these tests is consistent with how they are used in most syntax and grammar textbooks. See the article on the constituent and the sources on constituency tests cited there (I am largely responsible for the content of that article and the citations listed). There are literally dozens and dozens of syntax textbooks that employ them in the manner that I have done here.

I fully believe that you are right when you state that it is more difficult to computationally pick out a given non-constituent catena than it is to pick out a constituent catena (= your subtree). But I am wondering whether that is enough motivation to posit the existence of syntactic structures that are not supported by empirical considerations (e.g. constituency tests). I cannot imagine that computational applications of dependency theory are going to profit in the long run if they model syntactic structures in a way that is not supported by everyday empiricism.

Your main argument for the structures you propose seems to be about computational applications of dependency structures. In this area, my very limited knowledge of the field suggests that your approach is also contrary to the current state of the art. I know, for instance, that Joakim Nivre, a leading voice in the field, and his collaborators also position auxiliary verbs above the main verb.

As to your last comment about the quality of the article, I support any efforts to improve the coverage and presentation of the theory, and you of course have just as much right to edit the article as me. But according to Wikipedia policy, disputes about content should be resolved above all by basing the presentation on prominent literature that can be cited. In this area, I will scrutinize additions to the article. --Tjo3ya (talk) 18:39, 9 August 2012 (UTC)


 * Tjo3ya writes: --- I know, for instance, that Joakim Nivre, a leading voice in the field, and his collaborators also position auxiliary verbs above the main verb. ---


 * Joakim Nivre includes auxiliary verbs in the group of language features with no clear consensus:
 * "there are also many constructions that have a relatively unclear status. This group includes constructions that involve grammatical function words, such as articles, complementizers and auxiliary verbs, but also structures involving prepositional phrases. For these constructions, there is no general consensus in the tradition of dependency grammar as to whether they should be analyzed as head-dependent relations at all and, if so, what should be regarded as the head and what should be regarded as the dependent.  For example, some theories regard auxiliary verbs as heads taking lexical verbs as dependents; other theories make the opposite assumption; and yet other theories assume that verb chains are connected by relations that are not dependencies in the usual sense."  (quoted from "Dependency Grammar and Dependency Parsers https://files.ifi.uzh.ch/cl/kalju/Courses/2006_DG_Tartu/ToRead/05133.pdf  ).


 * Tjo3ya writes: ---more difficult to computationally --
 * I said something much stronger. That it was unsolvable. There isn't enough information to pick out the nodes since one knows the root (top node) of each subgraph but not which nodes to exclude. "more difficult" would mean that you could find the answer if your program could run long enough.  Sorry to nit on you, but the distinction between "unsolvable (undecidable)" and "difficult (computationally hard)" in computer science is quite important.  I encourage you to look up terms like undecidable and "computationally hard" in whatever source you trust.


 * Tjo3ya writes: --- More importantly, your use of constituency tests baffles me. ---
 * My main point was that "will have been" is clearly functioning as a unit. You can see that from the fact it moves as a unit.  As for your demonstration that the tests fail ... of course the tests will fail if they are narrowly defined in such a way that they only are applicable to noun phrases.    No surprise there.  "will have been" is a unit, but it isn't a noun phrase.


 * In any case, this is off topic.  The real issue we need to address is your statement "well if you add Rambow, I'll have to add my constituency test argument".  A review article is supposed report on a range of views that are considered part of the theoretical tradition. It is not supposed to be a battle ground for which approach is the best approach. Rebuttals are out of place. You aren't actually including an alternate point of view if your sole purpose is to treat it as a straw man to bash down.


 * What should be happening in that article is something more along the lines of the Nivre article I just cited. It is a good example of neutrality in action. I'm sure Nivre has his own preferences, but he also understands that when he is writing an article reviewing theoretical approaches and applications he has to report not evaluate.  You don't see him saying that its non-normative to have a different view than he does about where to put the auxiliary verb.


 * Here's a proposal – why not do an end run around this whole issue and just diagram sentences that don't have any disputed constructs, e.g. sentences that use a simple verb and have simple noun phrases as subject and object.  And better yet,  annotate the graph with something that says what theory of dependency (and what strata) was used to define the arrows.  Even something as simple as the direction of the arrow on "the tree" depends very much on how you define dependency, i.e. by which theory you are using to draw the graph.
 * Tjo3ya writes: --- As to your last comment about the quality of the article... ---
 * I have more to say about that, but lets take one thing at a time and resolve the graph issue. The changes that need to be made to this article are a lot deeper than adding a few paragraphs with citations.  — Preceding unsigned comment added by 77.124.109.209 (talk) 19:03, 16 August 2012 (UTC)

Where's the literature?
Hi Beth,

Where's the literature that takes auxiliaries to be dependents of full verbs? You have produced an unpublished paper by Rambow et al. who do it that way and you have now cited Nivre, who states that some DGs do it that way, but Nivre himself assumes the opposites in his works. Furthermore, Nivre does not cite anyone in the passage you quote who actually does it that way. And indeed, I have just located an article in which Joshi and Rambow position the auxiliary above the main verb:

http://www1.cs.columbia.edu/~rambow/papers/joshi-rambow-2003.pdf

For me to believe that there is a significant contingent in DG that takes auxiliaries as dependents of main verbs, I need to see actual illustrations in peer-reviewed journal articles or in books from prominent publishers.

Compare your position with mine. The leading voices in DG on the theoretical side (e.g. Hudson, Kunze, Mel'cuk, Starosta, Gross, Eroms, Heringer, Engel, Kahane, Gerdes, etc.) all position auxiliaries above full verbs, and on the computational side, Nivre and his colleagues appear to also position auxiliaries above full verbs. I can easily list the sources and page numbers that back up my claims.

Your statement that the challenge of picking out verb catenae is an undecidable problem is suspect for me. Taggers can tag the input. Once the words are tagged, I can hardly imagine that it would be impossible to pick out verb catenae. In English for instance, the highest verb in a verb catena is usually the left-most verb, and from there, the verb catena cascades downward to the right. Given these easily observable facts, isolating these verb catenae should not be an overly difficult task.

Your suggestion for a compromise does not appeal to me. DG has progressed very far from the simple tree structures that you suggest should be produced in the article. DG is now a full blown theory of syntax and grammar that can compete with the best constituency grammars. The compromise you suggest would generate the opposite impression. In fact the trees you want to produce would hearken back to the early 1970s. Furthermore, I do not see the need to compromise in this area because you have failed to produce significant peer-reviewed literature that backs up the type of analysis that you prefer.

I may not respond to further messages unless they include some significant literature that backs up your stance or unless some third party voices enter the exchange. I will, however, defend the current content of the article against any attempts to change it. --Tjo3ya (talk) 23:06, 16 August 2012 (UTC)


 * Tjo3ya writes: Your suggestion for a compromise does not appeal to me


 * Huh? How does labelling a graph as "drawn according to so-so-so's theory for strata X" make the article look less sophisticated?  To me, that makes it more informative. it means I can go look up that theorist and understand why he argues for such a graph.  Don't you think your readers deserve that opportunity?  Had you done even that, I might not have responded at all to this article.  I would have simply looked up the theorist.


 * On the topic of simple sentences, I fail to see how using simple sentences is going to make DG look unsophisticated. There is no reason its rich range of ideas on various grammar topics can't be discussed in sections outlining the different views on each of those topics (auxiliaries, conjunctions, and more).  In those sections it would be very appropriate to illustrate different approaches to more complex sentence structures.


 * Simplicity often communicates better because it lets people focus on the actual point you are trying to illustrate, rather than get caught up in debates in their mind. None of your diagrams are making a point specific to the modelling of auxiliary verbs so why should you even care whether or not the sentence has an aux verb?  The points those diagrams were illustrating can be made just as well without using aux verbs, so why use them?


 * Your statement that the challenge of picking out verb catenae is an undecidable problem is suspect for me. 


 * First of all, the diagram in the article isn't tagged. Second, even when there are tags, the tags aren't enough.  Even with tagging, a programmer still needs a rule for deciding which node-tag combinations to include or exclude from the sub-graph.  The only reason mathematical sub-trees can get away without such a rule is that they have an implicit rule of "everything".   If you don't want everything, you need to explain what you don't want.


 * Without that information (the rule) the catena is no more extractable than is determining whether "x" is an integer when the only thing I know about "x" is that "x=y" and "y is some member of the set of real numbers".  Or alternatively, its like saying "get me some files in subdirectory 'foo' ".  Sure the files have types and names. But that isn't going to tell me whether you want the files that contain music or the files that are 1 to 3 levels deep or something else entirely.


 * So again, if your goal was to show a powerful graph that illustrates the strength of DG, you haven't done it.


 * Tjo3ya writes: Compare your position with mine.


 * My position is simply that debate (both current and historical) needs to be acknowledged and discussed. I don't particularly care what the content or sides of the debate are so long as the debate is acknowledged and all opinions properly cited. It isn't healthy or honest for an article to present a field as monolithic when it is not. In fact it makes the field look dead and uninteresting, rather than rich and vibrant.  If one wants a field to look mature, one needs to present it as something other than dogma.


 * When I read an article like this (http://www.cl.uzh.ch/studies/theses/lic-master-theses/lizgerold.pdf), I see someone thinking hard about epistemological issues. That interests me.I'm interested not just because of the ideas it expresses, but also because it means he had an advisor and an academic environment that supported his wanting to do so. A mature interesting field needs to be aware of its own epistemological challenges.  In my opinion a field is described better by the questions it asks than the answers it gives.  If this Wiki article was the only article I'd ever read on DG, I'd be running miles away from it on the supposition that it had become the latest religion.  There isn't a single sentence in this Wiki article that demonstrates that DG has the capacity to be self critical beyond wanting to expand the scope of the problems it addresses.


 * As far as deciding what is and what is not a debate, the WP:SYN policy explicitly states that editors should not do their own synthetic analysis of a field. They are supposed to report on other people's synthesis.  In the context of an article reviewing a theoretical domain, I take that to mean that we need to be citing neutral lit reviews and not our own opinions about what counts as a debate or not.  If recognized authors who are qualified to review the field say that such and such is a debate, that makes it a debate.  Maybe they meant a current debate.  Maybe they meant a historical debate.  Either way it is a debate and needs to be acknowledged.


 * Nivre and Van Helden aren't the only survey sources I could cite to support that there are debates both historic and current within DG, including debates over aux verbs. But until we agree on what sort of sources need to be cited, there isn't much point in bringing on more sources.


 * What I'm also not sure of how to handle is your notion what counts as support for your opinion. You cite Nivre in support of their being a lack of legitimate debate when he's gone on record saying that he believes there is legitimate debate. You cite Rambow in support of there being a lack of debate, even though there is evidence he goes either way, sometimes using aux at the head and sometimes not. Wouldn't it be more reasonable to consider him a living example of someone who thinks there is no "one way"?


 * Furthermore, what motive would Nivre possibly have for acknowledging the legitimacy of an opinion that differs from his own if there weren't sources showing the existence of that opinion? I think we can safely assume that the literature exists, or at least the Nivre believes it exists, even if he didn't go into detail.


 * I'm not suggesting you eliminate a discussion of aux verbs from the wiki article. But given that two qualified sources assert that this has been a problematic issue (Nivre, Van Helden) the appropriate place to discuss this issue is in a section devoted to historic and current trends in the handling of aux verbs.


 * Beth 77.124.109.209 (talk)


 * PS: I note that you respond each time with a new top high level topic. Can we please indent responses under the main original topic, either using colons or subtitles?  All of these top level entries are one discussion and so belong grouped together. You can create a subtitle by adding an extra "=" before the section title. "===" will create a subtopic one level in from "==".  "====" will create a subtopic one level in from "===", and so on.  — Preceding unsigned comment added by 77.124.109.209 (talk) 15:06, 17 August 2012 (UTC)

Introductory paragraph
I added the introductory paragraph, describing what a dependency grammar actually is. The article doesn't actually present an introduction or summary of what the term dependency grammar means. But I'm not a linguist, so I had to go out on the web and try to develop my own understanding of the topic. Hopefully, the introduction I added is a concise statement that allows a non-linguist at least to grasp what the rest of the article is talking about. — Preceding unsigned comment added by 76.210.149.11 (talk) 03:42, 13 September 2012 (UTC)


 * Hello ???,


 * In my opinion, the paragraph you added was not helpful. That paragraph basically put the cart before the horse, since it jumped right to an example sentence without providing a lead in. An understanding of dependency grammar, and of any other topic, is built up incrementally over time by reading and pondering the data and issues surrounding the data. Checking other sources on the internet and elsewhere is of course good. That's part of building understanding.


 * Concerning this specific article, perhaps you can provide some feedback about why you don't/didn't understand when you first read it. A key aspect of understanding dependency grammar is to know that it is quite different from a constituency grammar (= phrase structure grammar). Constituency grammars are generally more common in Anglo-American linguistics, and it is therefore important to know that one is dealing with an approach to syntax that one may not have encountered before in language and/or linguistics courses. --Tjo3ya (talk) 05:43, 13 September 2012 (UTC)

I want more info or a replacement of the word flatter as you stated in the and yes it did help but I'm not a fan of syntax in linguistics since seemed so lengthy and complex   — Preceding unsigned comment added by 75.61.138.63 (talk) 00:28, 5 April 2015 (UTC)

Correctness of examples in "Representing dependencies" ?
The example representations use The conventions can vary. and place can at the head of the tree. Is this really correct? In my view, vary is the verb, thus the head of the sentence; can modifies it by introducing a modal variant. Would you say that in She's eating an apple., is is the head? in my view, is only introduces an aspect variant of eat. denis &#39;spir&#39; (talk) 09:45, 28 March 2014 (UTC)


 * There are numerous linguistic arguments demonstrating that the finite verb is the root of the clause: the results of constituency tests, subject-verb agreement, V2 word order in Germanic languages, position of negation ne...pas in French. There is only one finite verb per clause (barring examples involving coordinate structures), but there can be numerous nonfinite verbs. Placing the finite verb as the root of the clause is consistent with the vast majority of work in syntax in the past 35 years, in particular with work in phrase structure grammars (PSGs). Every modern PSG that I am aware of assumes the finite verb to be the highest verb in the structure.


 * Your question may be stemming from dependency parsers in computational linguistics. (At least) one prominent dependency scheme, the Stanford dependency scheme, takes the full verb as the root. The motivation for doing this is to reduce the dependency distance between content words. From a linguistic point of view, there is no argument for doing that. Other dependency schemes for parsing natural language take the finite verb as the root of the structure. If you would like me to elaborate on any of these points, please say so. I will be happy to back these points up with examples.


 * Corncerning clitics, there is no problem assuming that a clitic auxiliary verb is the root of the structure. The examples in the article (here at Wikipedia) demonstrate how this is done. --Tjo3ya (talk) 16:49, 28 March 2014 (UTC)


 * It seems to me that your argumentation stems from taking for granted abstract, meaningless (literally!) approaches to grammar à la Chomsky; as if language was not supposed to signify (to mean); a path that, i guess, has abundantly proved to be wrong and a dead end and is quickly beeing abandoned by all kinds of linguistic domains and schools, including natural language processing in software (perhaps, it's previous stronghold).


 * So-called syntactic schemas or rules or categories actually nearly always are semantic. Take for instance SVO order: it is used to indicate which part of a verbal sentence is the subject (often agent, else theme) and which is the object (or attribute, or identifier...); both S & O typically are noun phrases, thus we need to make a distinction. In other languages, indeed, the distinction is made by case alterations, in some (eg german) by both order and case. These distinctions are purely semantic, however for some reasons called syntactic. Syntax is mostly a system of common, conventional means to signify meaning (semantics).


 * There are nevertheless apparently pure syntactic, or rather grammatical, rules, the ones that force us to introduce "unmeant" elements, or more generally alterate our thought. For instance:
 * She said she would leave early.
 * Here, the speaker probably did not mean:
 * That person, who probably has XX sexual chromosomes and is considered as woman, said "self" would leave early.
 * or anything similar. The speaker instead probably just meant
 * That person said "self" would leave early.
 * but was forced by english grammar to irrelevantly introduce sex/gender into their picture. (This is certainly both a consequence and a cause of our sexist ideology, but this is not the topic I'm discussing here.) Right? Such schemas may be called syntactic; but as shown even this case is not semantic-free, indeed it carries ideological dogma/belief (that "man" vs "woman" are fundamentally natural categories, and always fundamentally relevant, like say chemical elements, as opposed to ideological categories).


 * Another example is French ne which (in modern language) is imposed by official grammar rather than carrying any meaning: "Pourvu qu'il ne vienne pas !"; "Je craignais qu'il ne vienne." In such a case, there really is no meaning (in modern French): this part of the shema is purely syntactic.


 * To sum up: your distinction between syntactic and semantic (and other) is basically wrong, in my view. In nearly all cases, a syntactic schema is a general expressive schema for a semantic schema; at times, an alteration of the meaning is imposed by the language's rules. In very rare cases, some aspect of a schema carries no meaning at all: this is exceptional.

denis &#39;spir&#39; (talk) 18:48, 28 March 2014 (UTC)

Denispir, above all, please do not associate my views on syntax with the Chomskyan tradition. I am decidedly anti-Chomskyan, I am a DG guy. To be a DG guy means that one probably disavows oneself from the Chomskyan tradition, which I do, again, decidedly. Much of Chomskyan syntax is nonsense in my view.

But I'm not sure where your argumentation is headed. You are emphasizing the importance of semantic units. That is of course a legitimate thing to do. But if one does that in DG, one is interested more in semantic dependencies, i.e. dependencies between predicates and their arguments. Most work in DG focuses, instead, on syntactic dependencies. Note that the article mentions both types of dependencies; it focuses much more on syntactic dependencies because that is where the emphasis traditionally lies. Let's take an example:

(1) Fred is revising the text.

The predicate "is revising" takes the arguments "Fred" and "the text". We probably agree about that, because from a semantic point of view, most would agree that "Fred" and "the text" are the arguments of the predicate "(is) revising". But "Fred" is a syntactic dependent of "is", not of "revising". We know this in part because of the subject-verb agreement. "Fred" agrees with "is", not with "revising". This reasoning is syntactic. We also know that "the text" is a syntactic dependent of "revising", in part because the three words act as a single syntactic unit, e.g. "... and revising the text, Fred (certainly) is". We have diagnostics for syntactic structure that we use to reveal the syntactic structure of sentences. The stance that you seem to be advocating for appears to want to ignore these diagnostics.

But before we continue with this exchange, I must state that any changes to the article as it currently stands would have to be backed up by good literature. In other words, if you are going to argue that this article on DG should be changed/expanded/improved, I will likely emphasize that whatever changes are made, they must be backed up by good literature. --Tjo3ya (talk) 19:29, 28 March 2014 (UTC)

Recent additions
The additional information that has recently been added to the article does not fit into the greater whole; it is, rather, a bit redundant. I therefore do not think it improves the article. I am going to remove it now. If there is disagreement, let us first discuss the issue here. --Tjo3ya (talk) 02:37, 8 December 2014 (UTC)


 * I understand that the information I added was already mentioned in the article in an implicit way, but I think an article on dependency grammars needs to explicitly say what a dependency relation is, and even what a grammar is, in the interest of clarity. The way I realised this was by speaking with user JMP EAX, who commented on the lack of clarity in the article.


 * Actually, the reason we were talking about it in the first place is that I started an article which needed a definition of dependency relations. Rather than repeating the definition in the article I started, I figured it'd be best to link to the Dependency Grammar article.
 * Christian Nassif-Haynes (talk) 03:40, 16 December 2014 (UTC)

I think I can help with some of the questions in the other page. Cross-serial dependencies are indeed concerned with syntactic dependencies. In fact most work in DG is focusing on syntactic dependencies. Other types of dependencies clearly exist, though, and acknowledging these other types is important, because if one fails to do so, one can get confused about syntactic dependencies. Some of the work in computational circles (e.g. the Stanford/Google crowd) is making this mistake. It is mixing up semantic dependencies with syntactic dependencies. It is precisely for this reason that I added the sections on other types of dependencies (semantic, morphological, prosodic). Igor Mel'cuk is the one who has written most about the different types of dependencies. His work is, however, too often ignored by those working in computational circles.

That being said, please comment further on what is unclear about the article. The article should be accessible to a large audience. I am not opposed to adding or changing something. What I am opposed to is additions to the article that do not fit into the whole. --Tjo3ya (talk) 10:09, 16 December 2014 (UTC)


 * If you want to know why JMP EAX thought it was unclear, you'd have to speak with him/her. Personally, I think it's just unclear because, like I said, I can't see anything in the article which explicitly mentions what a dependency or grammar is. The "Syntactic Functions" section comes closest to defining the dependency relation because you can see the dependencies labelled---so now, if the reader if clever, (s)he can infer that dependencies describe the (syntactic) relationships which pairs of words have with each other. Actually, I think something like this does a concise job of describing dependency grammars in that it's very explicit and describes new phrases and concepts fully when they're introduced. In particular, the "connections" section gives a good sense of what a dependency relation is. Of course it's nowhere near as in-depth as the DG article, which is why the DG article should be clarified. I'd go so far as to say all the relevant content from the Lucien Tesniere article should be taken from it and merged with the DG one.


 * Now, let me try and give you a concrete example of something I think is unclear (in the DG article). If you look at the very first sentence, you can see that the phrase "dependency relation" is introduced. In the sentences which follow, some facts are stated about dependency relations, but they're never defined. It's a bit like someone asking what a cow is and getting the reply, "cows are ruminants. They have four legs"—Of course I'm wiser for having this new knowledge, but I still don't have a feel for what a cow really is—it could look like a grass-eating Siamese cat for all I know!


 * Christian Nassif-Haynes (talk) 13:48, 16 December 2014 (UTC)


 * Thanks for your comments. I have made a couple of changes to the introduction in an attempt to accommodate some of your points. You will probably not be satisfied by these minor changes, however. In the scheme of things, I think your comments are motivated in part by a minor misinterpretation that you may have picked up elswhere. You seem to overemphasize the role that the syntactic functions play in dependency theory. Many DGs produce tree structures similar to most of the trees in the article (e.g. Tesniere 1959, Starosta 1988, Heringer 1996, Gross 1999, Eroms 2000) insofar as the trees do not include the labels showing the syntactic functions. And some prominent DGs that prefer the dependency arcs also tend to omit the labels for the syntactic functions (e.g. Matthews 1981, Hudson 1984, 1990). The syntactic functions do play an important role in DG theory, but they do not constitute the main thrust of what a dependency grammar is. The main thrust of a DG concerns how it is grouping the words; it is grouping them directly, instead of via the intermediate nodes associated with the constituency relation.


 * Concernig a definition of dependency, producing one would be fraught with difficulties. Theoreticians wrangle incessantly over definitions, and any definition produced in the article would certainly be unsatisfactory to many DG theoreticians.


 * Concerning the article on Lucien Tesniere, I am responsible for its content. Including that information in the DG article would be redundant. Furthermore, Tesniere's DG differs from most modern DGs in significant ways. --Tjo3ya (talk) 07:15, 17 December 2014 (UTC)


 * Concerning the word `constellations'. I have tried to Google this word in relation to grammars and I cannot find any information. This word really does need a definition. I am very interested in grammars and am surprised to find a word that I have never seen before in this context. Can someone please either define the word in this context or give a reference so that I can do this? Thanks for an otherwise good article. Jimmy3987 (talk) 05:10, 22 December 2017 (UTC) jimmy3987 22 December 2017.
 * Think of syntax trees, like the ones in the article. They are similar to the constellations of stars in the night sky. It is a simple metaphor; it should not be over interpreted. Note, however, that there has been a concrete distinction drawn using a similar term, namely configuration. See non-configurational language. --Tjo3ya (talk) 01:22, 23 December 2017 (UTC)

Identity of Phrase structure (at least some) and dependency grammar? (Asking more out of interest than anything else)
Hi, looking at the nice graph comparing the two competing (and apparently at odds) theories, it looked to me as if the two were - well - identical? Obviously not with the number of nodes they have, but it seems as if the one (Dependency grammar) could be expanded into the other and the other (Phrase structure) folded into the one ? I mean by means of a simple rule to follow, ie without "human" intervention (rather by an ~ algorithm).

I might look into this more myself still, but not just now :). In the meantime, if anyone can comment, I'd be interested, thanks! 37.209.42.230 (talk) 14:12, 13 August 2015 (UTC)


 * Yes, dependency structures can be mechanically translated to constituency structures, and if constituency structures are entirely endocentric (i.e. headed), they can be mechanically converted to dependency structures. The difference, though, lies with the number of groupings (complete subtrees). Since constituency allows the number of nodes to exceed the number of words, the structures can be very tall, whereas the one-to-one restriction (words to nodes) on dependency structures forces one to assume relatively flat syntactic structures. This is a big difference. The constituency structures that result when one translates from dependency structures are flatter than most constituency grammars want to assume. Much of modern syntactic theory in the Chomskian tradition assumes strict binarity of branching, which results in very tall tree structures, with a lot of nodes. There is no way for dependency to acknowledge so many groupings. Further, all dependency structures are endocentric (i.e. headed), whereas constituency structures can also be exocentric. This is a further big difference. --Tjo3ya (talk) 15:59, 13 August 2015 (UTC)
 * Cool, thanks for the thorough answer! I don't think I've worked out / followed you on all the differences yet, but that's a great head start :). I'll have to see if I can put it to some use. Regards 37.209.42.230 (talk) 11:56, 17 August 2015 (UTC)