German-like subordinate word order: Preventing subordinate phrases from firing twice



Suppose I want to have a different word order for subordinate clauses, like in German. This is for the Grammar Matrix, so has to be a basic, simple analysis that will work cross-linguistically and won’t get in the way of anything, so we are not necessarily talking about implementing German as in e.g. Fokkens’s work.

It seems straightforward to have special phrase structure rules then for subordinate clauses, like so (suppose the matrix word order is V2):

subord-comp-head-phrase := basic-head-1st-comp-phrase & head-final &

subord-subj-head-phrase := decl-head-subj-phrase & head-final &

However, this way, there will be 2 trees for the sentence:

subject1 knows that object subject2 took

Specifically, “subject2 took” (“noun2 verb” in the screenshot below) will be licensed by either the subordinate-subj-head rule or the normal subj-head rule:

Am I right in thinking that this is undesirable?

And that I should perhaps restrict the additional phrase structure rules as follows:

subord-comp-head-phrase := basic-head-1st-comp-phrase & head-final &
MC - ] ].

subord-subj-head-phrase := decl-head-subj-phrase & head-final &
MC - ] ].


Can you constrain the normal comp-head and head-subj rules to select for MC +? That might generalize better if other matrix/suborod word order pairs are added down the line.


So, the subordinate S is itself licensed by a subordinate phrase structure rule. It is therefore [ MC - ], but it needs to go through a normal comp-head rule to combine with a C. So normal phrase structure rules should not be restricted to [ MC + ], I don’t think.

Another way of looking at it: The difference in the two trees (which I assume is a spurious difference) is low: there are two ways to license the lowest branching V which then becomes part of the subordinate clause. But the [ MC - ] constraint emerges later in the tree, unless there is nominalization or something of the sort going on. In this case, we have fairly free word order, so it is probably not possible to tell that we are inside a subordinate clause as early as in the lowest branching V.


I think you are right actually and something like that is needed…

Because the way I did it first, the grammar will also parse:

subject1 knows that subject2 took object

which should be rejected. But it won’t be if a normal head-subj can fire for the low “v1 noun”.

Except it can’t exactly be constraining the normal rules to [ MC + ] since, for V2, it will clash with the specifications for nexus phrases which must be [ MC na ]. Looking into Fokkens now.


If you are able to constrain the normal rules to MC +, you could just have a special head-comp rule for the CP. Do you know why/where the nexus phrase is constrained to MC na? I don’t see that in matrix.tdl. Is it added by the word order library or is that coming from one of Fokkens’ grammars?


The nexus-phrase rules are used only in V2 word order (so yes, added by the WO library). However, I do not need them for the subordinate clauses, I suppose.

What presents a problem is actually constraining head-final phrases to [ MC bool ]:

head-final-head-nexus := head-final &

This means a normal main clause will unify with a complementizer that wants a [ MC - ] complement.

After reading a footnote in Fokkens (2014), looks like the reason for this is to accommodate languages which can, in fact, exhibit V2 order in both matrix and subordinate clauses.

So, for my German-like option, looks like all I may need to do is change it back to [ MC + ], as you suggested :). And then have the complementizer select for [ MC - ] things.


…Just the above, however, won’t fix accepting ungrammatical strings like

subject1 knows that subject2 took object

unless I also do something like:

subord-comp-head-phrase := basic-head-1st-comp-phrase & head-final &

subord-subj-head-phrase := decl-head-subj-phrase & head-final &

Does that look sensible?.. Insisting that, in a subordinate clause, the lower level constituent (such as a VP) is also licensed by a subordinate rule?


That does seem sensible to me. But makes me wonder why MC isn’t being identified between the head daughter and mother anyway. In matrix.tdl I see it identified in the opt and extracted comp phrases. Does anyone else know why this identity is not in the basic head-comp/subj phrases? Should it be?


I think here’s why (from matrix.tdl):

; ASF (2008-11-03) for the current v2 analysis, we need head-comp-phrase
; that does not pass up the MC feature…token id between mother and head
; daughter removed


Is there any further explanation of why MC shouldn’t be passed up? I’m struggling to think of a case where we would want to call a clause a matrix clause at one level but not at another level… It might be that it’s useful for Antske Fokkens’s analysis, but that calling it MC is confusing.

Specifically for (formal) German, the more precise statement is that verb-final word order is needed for subordinate clauses with a conjunction or complementiser. Without a complementiser, a complement clause looks just like a matrix clause:

  • subject1 knows, that subject2 object took
  • subject1 knows, subject2 took object


Yes, I think that is the case.

Here’s an excerpt from Antske’s thesis:

The MC feature was originally introduced to distinguish main clauses from subordinate clauses. In the Grammar Matrix customised verb second analysis, it registers which rule was last used. The desired behaviour is that exactly one constituent appears before the verb that heads the sentence. This is achieved by making sure that the head of a clause cannot attach to any other element as soon as it has been head-final once.
Verbs start out with the value [MC na]. Both head-final and head-initial rules require this value on their head-daughter. The head-initial rules pass the value of MC up from their head-daughter to the mother. The head-final rules, on the other hand, assign the value + to MC,15 so that neither head-final rules nor head-initial rules can apply thereafter. The root condition in the grammar specifies that only sentences with the feature-value pair [MC +] are acceptable main clauses. The grammar will thus only accept sentences where the head-final rule has applied exactly once, placing the head of the sentence in second position. […] The word order analysis using MC to create verb second word order straight- forwardly extends to verb-final order in subordinate clauses. Verbs and verbal phrases in the Right Bracket bear the value [MC −]. Subordinate conjunctive markers require this value on their complement. Informal verb second subordinate clauses are licensed by a unary rule.


A related issue comes up if the grammar has auxiliaries. The comp=head rule doesn’t require that the MC values of the head-dtr and non-head-dtr match, as this would be undesirable for a clausal complement taking verb, which is MC + and its daughter is MC -. However, for auxiliaries, we do want the MC values to match to avoid spurious parsing as shown below where the lower VP in the first tree is licensed by comp-head and the lower VP in the second is licensed by subord-comp-head.

Perhaps it’s sufficient to just have COMP head constrained to pass identify MC, and we don’t need a subord-comp-head at all? Just head-comp would have to be constrained to MC +?


If we always want to have the same MC value on auxiliary and its complement, we could constrain that on the lexical type for the auxiliary? Rather than on the head-complement rule.

As for the different phrase-structure rules, I will revisit the question of why I am having them :).


Basically, it seems like Antske had reasons to disallow the identity between mother’s and daughter’s MC. That was needed for her V2 analysis, and I can’t say I fully understand it as of yet. But since we probably don’t want to replace her analysis, I thought it would be easiest to add subordinate phrase structure rules which do identify mother and daughter’s MC and insist that it is -. Guy is also asking above why not identify it throughout.


Really, it seems like the quote above (from Antske’s thesis; click on my userpic in the top right corner of this message to see it) is pretty much all there is, regarding the GM implementation of V2. (Note that this is not her primary analysis of the German word order; it is an alternative one used specifically for the GM). I don’t understand how the analysis extends to the subordinate clauses. If V-final rules require [ MC + ], then how would a [ MC - ] thing go through them? Is there an additional rule? That’s what I assume for now. She says clearly a head-final rule can only apply once, so there is no way for it to license both the subordinate and the martrix clause. (Also, it seems pretty clear at this point that Guy is right and the MC feature is not used in this case to distinguish between the main clause and the subordinate clause, but rather to register which rule has applied last. So perhaps we should not be trying to use it for both purposes? But even if we put in an additional feature, I still don’t see how we would be able to license both a subordinate and a matrix clause under current V2 analysis without positing additional phrase-structure rules. Which brings us back to the question about how to not have them fire spuriously, such as with auxiliaries.)


Thanks for posting that extract. It really sounds like we have two very different uses of the MC feature here, and I think it would be cleanest to separate them.

If V-final rules require [ MC + ], then how would a [ MC - ] thing go through them? Is there an additional rule? That’s what I assume for now.

It certainly sounds like it! Either that, or the verb is not actually treated as the head of the clause – potentially, the conjunction could be treated as the head instead. But that would make the grammar look quite unlike other GM grammars.


To summarize this thread: the solution appears to be:

  1. Have separate phrase structure rules for the subordinate clauses, and have them [ MC - ] (and the complementizer selects for [ MC - ] ).

  2. Have auxiliaries select for [ MC na-or-- ] which prevents them from combining with a structure which had been licensed by a [ MC + ] rule, which kind of makes sense.