Home » Articles » NG2. What does sentence structure signify?

NG2. What does sentence structure signify?

This is the first substantive piece for LanguidSlog.  It’s quite short and should be easy for the reader.  Books on syntax typically jump straight into tree diagrams or nested brackets to represent sentence structure.  The graphical conventions are not difficult.  But the purpose of those representations isn’t self-evident.  Let’s look at some issues.

Sentence structureStructure and meaning

Linguists discern structure in sentences.  They then show that differences in sentence structure correlate with differences in meaning perceived by native speakers of that language.  They therefore assume that the meaning of a sentence is given by its structure.  The phonological string (the sounds or the signs) is merely how the meaning/structure is encoded for efficient transmission from speaker to hearer.

Some might quibble with the following model but most theories imply sentence processing is a sequence of steps like this:

meaning intended by speaker

=A=> structure

=B=> phonological output

===> phonological input

=C=> structure

=D=> meaning comprehended by hearer

The cognitive state initially producing the sentence triggers a process of construction (step A); and the cognitive state finally produced by the sentence results from a process of deconstruction (step D).

Others theories could imply that the initial and final cognitive states are the structure itself operating directly in the mental architecture (A and D do not occur).

None of these possibilities has ever been explained clearly enough for me.  That’s unhelpful but forgivable because ‘meaning’ is inaccessible.  Conjecture about it can only be expressed by reverting to language and that is rarely convincing.

No such vagueness applies to steps B or C.  Actually the scholarly literature pays more attention to C because authors usually start with a paradigm sentence and thence develop a structure to suit.  And for a computational linguist bravely trying to build a parser based on grammar, identifying relations within the phonological string is the primary aim.  This blog will therefore also concentrate on step C and assume any theory falsified in those terms must fail whatever it might claim for A, B or D.

Structure and depiction

Structure might be used merely to depict a sentence, with no assumption that it actually participates in the production or comprehension of that sentence.  That is not to say trees and nests are strictly for the birds.  Structures are useful – perhaps essential – to enable syntacticians and students to discuss sentences.

But a theory denying anything like steps A to D could not claim to be explanatory without some other means of showing how, for a sentence, the whole is greater than the sum of the parts.  This blog will eventually reveal such a theory.

Next time

My scepticism needs a firm basis.  We must therefore give due consideration to the possibility that structure does play a part in comprehension.  The next two pieces will discuss what would need to be stored from an incoming sentence in step C.

Mr Nice-Guy


  1. MW says:

    Finally some common sense in the world of linguistics. I’ve read a lot of stuff over at languagelog.com that just doesn’t seem to address the fundamental issues in a remotely plausible way. Look forward to your next chapter Mr Nice Guy…!

  2. SH says:

    “That is not to say trees and nests are strictly for the birds” made me laugh rather more than it should have. Odd looks from others in the cafe.
    It seems to me that there’s plenty of syntactic theory out there that focuses on how speakers/writers encode their meaning, rather than how listeners/readers decode it. The main focus of a lot of theoretical syntax is “of all the ways language could be put together, it seems peculiarly limited. Why is that?”
    The meaning of a sentence may not be entirely given by its structure, but its structure is certainly restricted. It’s possible that this isn’t of interest to you, of course.

    • Mr Nice-Guy says:

      Thanks, SH.

      You can look at the relationship between the phonological string and the structure drawn by a linguist from either viewpoint. Why do I allege that more emphasis is given to comprehension than to production? Perhaps merely because in the books the sentence usually precedes the tree. Also I studied production via Levelt (1989) – not really a syntax book.

      I agree about the ‘main focus’ of syntacticians. They have done a good job in documenting the peculiar limitations. The reason behind these limitations is the way language is constrained. This blog will pursue the idea that the constraints are more severe than is often assumed.

      My ‘given by its structure’ referred to the unconscious processing in the hearer’s head in normal language use. Of course, the structure written down by a linguist will ‘give’ the meaning of a sentence – because the linguist already comprehends it.
      Your ‘not…entirely’ is intriguing. Sure, the total cognitive effect on the hearer includes things from prosody, body language, mutual knowledge etc. Those things tend to be covered by other sub-disciplines, but did you mean something more mysterious in syntax?

      • SH says:

        Nothing so promising, I’m afraid. I was just referring to lexical choice, morphology, pragmatics, that sort of thing. Syntax is a lovely neat system on its own, but doesn’t operate in a vacuum (more’s the pity).

  3. MS says:

    You can either regard tree structures as representing the actual meaning of the sentence, where the tree represents the structure of a formula in some logical language in the same way that parentheses do in arithmetic. Or you can regard the tree as representing the way the words are put together to build that meaning.

    These two interpretations are closely related, but they are different: there may be more than one way of putting the same meaning together.


    • Mr Nice-Guy says:

      Thank you. That’s a good point except that ‘represents’ is not clear to me. I go on to claim that structure cannot participate in human sentence processing although obviously it has pedagogic value. Do please continue reading the blog and point out flaws in the argumentation.

      Is ‘structure’ paradoxical? Language allows an infinite number of sentences. If meaning is given by structure (plus lexical stuff), wouldn’t the brain need to hold an infinite number of structures? That can’t be right so perhaps structure is constructed from/deconstructed to atoms of meaning. But that would imply (for comprehension) parsing of the sentence to form the structure and parsing of the structure to deliver the ‘atoms’. Wouldn’t a better hypothesis be words straight to atoms?

  4. MS says:

    Re ‘“represents” is not clear to me’ … No-one has ever provided a convincing explanation of how children can learn any language they are exposed to other than in terms of access to structure. That structure doesn’t have to be uniquely linguistic and it is not impossible that there is another way. But the burden of proof is on you.

    Re ‘Is “structure” paradoxical?’ … The paradox is resolved by the idea of grammar. A grammar is a very simple object mathematically that has exactly the property of capturing infinite sets of expressions with finite means. For example, ordinary arithmetic forms a language with an infinite number of expressions. But it can be described by and interpreted using a tiny grammar.

    The idea of grammar is quite separate from its use in pedagogy.


    • Mr Nice-Guy says:

      Thanks again. Your input is most encouraging.

      Re ‘No-one has ever provided a convincing explanation’ … Yes, I like your broadening of ‘structure’. But in the blog the reader’s default assumption should be that structure means a PSG tree.

      Structure in that narrow sense is what I attack in subsequent posts. The burden of proof is indeed on me.

      Re ‘The paradox is resolved by the idea of grammar’ … That’s fine as long as ‘idea of grammar’ is tenable. In calling grammar an object, do you mean that it is somehow held in mental storage, distinguishable from the data it processes? That can’t be the case unless storage is addressable. Language and its processing must be autonomous, using one primitive function that also drives other mental activity. One aim of the blog is to show how that could be possible. I hope you will keep reading it – and writing!

  5. MS says:

    Re ‘In calling grammar an object’ … I’m not sure that is what I meant by calling grammar an object, but I certainly believe it is held in memory, distinct from the data it processes. That is how it is possible to process an infinite variety of senses with finite storage.

    Re ‘unless storage is addressable’ … Not sure what you are getting at, but addressable storage sounds fine to me.

    Re ‘processing must be autonomous’ … Again not sure what you are getting at (‘autonomous’ seems to be incompatible with ‘function that also drives other mental activity’). But the general idea of language sharing properties with other mental faculties has always been attractive in evolutionary terms.


    • Mr Nice-Guy says:

      Thanks yet again for your hugely valuable comments!

      Re ‘grammar an object’ and ‘addressable storage’ … Good. We agree that, if program and data are separate, addressability is needed. How else could data be manipulated and how could logic have branching and looping? To me it’s axiomatic and not limited to any particular architecture – von Neumann or whatever.

      So the crux is whether mental architecture actually has storage that is addressable. If there were addressability I would expect to find something about it via Google. But there’s no trace.

      Re ‘autonomous’ … Yes, I didn’t express that well. I was trying to say that all mental processing works in the same way. Homo sapiens needed no qualitative change for encoding/decoding language, just more capacity to hold the necessary data. Isn’t that more plausible in evolutionary terms?

      The aim of the blog is show how human language processing can work with no addressability, no separation of program and data, no ghost-in-the- machine. I won’t get everything right but will show it’s feasible.

      I’ll not be popular. Most linguistic theories imply more complex processing – ‘movement’ for example. Such a theory can be descriptively adequate but not explanatory without addressability.

  6. MS says:

    Re ‘addressability’ … ‘Addressable’ is a very general notion. Vision is an example of a system in which there is a conceptual separation between infinitely varied scenes and a finite mechanism which isn’t von Neumann. You might say it is ‘content addressable’. Associative memory would be another example.

    Re ‘no qualitative change for encoding/decoding language’ … Well maybe not no change, but a very small change, maybe just a bringing together of previously disparate elements.


    • Mr Nice-Guy says:

      At the start of this correspondence I didn’t know whether you were pro or con. But your several inputs have certainly been very helpful. We’re now at the heart of the matter. I hope you can continue commenting – even if you can’t support my iconoclasm.

      Re ‘”Addressable” is a very general notion’ … Agreed. Indeed content-addressability is implied in the blog (from LS7 on) by my glib assumption of a concept for each ‘phonological word’ encountered in comprehending a sentence. Activation from relevant acoustic elements must converge on such concepts, the rest being dissipated. I then show how, together with a particular organisation of lexical material, that’s enough for the meaning to be delivered. The mechanisms are simple.

      A theory of sentence processing is too complex if an intermediate result is put to one side, processing continues on something else to produce another intermediate result, and then the two results are brought together somehow. That requires a separate program. Both the storing/retrieving of intermediate results and branching/looping within the program itself require addressable storage – as in RAM (random access memory). CAM can’t do that.

      In the first few posts in LanguidSlog I show that the ‘intermediate result’ problem occurs whenever you try to build a structure – even one representing a sentence as simple as ‘John kissed Lucy’.

      Theories I’ve looked at all have this problem – even the work of researchers looking at content-addressability. For example, I quite like Alcocer and Phillips (2012) but they get distracted by c-command which might better be ignored as an epiphenomenon.

  7. MS says:

    Well, I’m all for incrementality in NLP, because that seems to be an empirical fact. Whether you can do anything in sentence processing without structure building is another question. Frankly, I’m skeptical. C-command is a natural property of applicative structures.


    • Mr Nice-Guy says:

      Re ‘Frankly, I’m skeptical’ … Which blinding light will effect a Damascene Conversion? I’ve already tried addressable storage on you. Instead let’s go for the presumed connection between structure and meaning in human language processing. And let’s use ‘tree’ to avoid slippery definitions of ‘structure’.

      If a tree is built from an incoming sentence it might be used in either of two ways. One, the tree itself, including the lexical items hanging from it, gives the meaning directly. Two, the tree must be further processed to arrive at the meaning.

      One is untenable because an infinite number of trees and their meanings would have to be stored by the language user.

      Two is implausible. Why go through two stages of processing when one is enough? Have the psycholinguists come up with any evidence of an intermediate state?

      Syntacticians like trees. Trees can be shown to correlate with meaning. But you never see a syntactician draw a tree for a sentence whose meaning they don’t already know. Trees are derived from meaning, not vice versa. They are artefacts.

      Your epistles are greatly appreciated. Keep them coming!

  8. MS says:

    Re ‘Let’s use “tree”’ … Well, OK, but I prefer ‘structure’ because meaning representations might not be trees. (They might be directed acyclic graphs.) But I’m happy to talk about trees.

    Re ‘One is untenable’ … No, that is just wrong. That is what grammars are for. They support the analysis of infinite treesets with finite resource. See my earlier remarks about the (tiny) grammar for the infinite language of arithmetic expressions.

    Re ‘Two is implausible’ … I agree that #2 is implausible, but #1 is fine.

    Re ‘Syntacticians like trees’ … That is sort of true about linguists. They reify syntax in ways they shouldn’t, when they are really talking about intuitions concerning meaning. One level of representation related to meaning is enough.

    But you still have to *have* meaning representations.


    • Mr Nice-Guy says:

      Despite the ‘just wrong’, I think we’re converging!

      In human language comprehension, words are processed to deliver meaning. My #1 and #2 models both assume a complete sentence is parsed and a tree (or whatever) is built. I said that #1 gives the meaning directly. If there were ‘analysis of infinite treesets…’ then we would be talking about model #2.

      Neither model is credible. The third way is to abandon any idea that a complete sentence is processed before meaning is delivered. (Disfluent speech also supports this.) Instead we should look at how incremental processing could identify a succession of sub-assemblies, each delivering part of the meaning of the sentence. I’ve gone for the smallest sub-assemblies: dependencies as in dependency grammar.

      Would a DG ‘tree’ meet your requirement for meaning representation?

      Thanks again.

  9. MS says:

    Re ‘Would a DG ‘tree’ meet (my) requirement for meaning representation?’ … Sure. DGs usually mainly concern predicate-argument relations, and that isn’t the whole of meaning representation, but they are fine as MR as far as they go.


    • Mr Nice-Guy says:

      The ideas seem to have survived your cross-examination. But the jury is the linguistics establishment. What will be their verdict? Many will feel threatened by a whole-sentence structure becoming merely a descriptive artefact, not actually part of human sentence processing.

      My case is based on four points. Structure cannot be built without addressable, RAM-like storage. Structure as the end-point of processing is impossible, and as an intermediate point implausible. Lexical organisation must account for the many-to-many relationship between phonological words and meanings. Disfluency pervades all spoken language and how its meaning is conveyed must be central to any explanatory theory.

      Doesn’t this deserves better dissemination?

  10. MS says:

    I think it will be hard to get the linguists to listen. The computational linguists will be more receptive, but they have already thought of a lot of this.
    In either case, you will have to show that these ideas *work*, in the sense of actually solving some problem at a practical scale.

    For the linguists, that would involve for example, showing that you can handle the full range of phenomena exhibited by relativization (unboundedness, multiple dependencies, across the board extraction from conjunction, adjunct islands, etc.) For the computationalist, this would be showing that whatever you can do you can do over large training and test datasets.


    • Mr Nice-Guy says:

      I agree about the linguists. The best tactics would be to build a rule-based computational model that works so well that the rules are indisputable. How the linguists react is then inconsequential.

      While I’m comfortable with the prospect of designing the model, data capture would be a large and tedious business way beyond my own resources.

      The do-or-die tactics I’ve adopted instead are therefore to attack the linguists directly. Progress would help enlist support for the computational work – I hope. The attack is on the methodology of theoretical linguistics which, as I show in the blog (and summarise in my last email), is fatally flawed. The alternative ‘grammar’ I put forward is merely intended to show that a solution without those flaws is feasible, but I make no claim that it’s complete or definitive.


This site uses Akismet to reduce spam. Learn how your comment data is processed.