How could a speaker use a DM grammar?

Distributed Morphology as a production architecture

Utku Turk

LING 449U, How words are built

2026-10-01

First, an ordinary DM tree

cats: a root, a nominal categorizer, and plural Number.

Before insertion, NumP contains nP with root CAT and n, plus Num plural. After insertion the same nodes receive kæt, zero, and z. Phonology yields cats.

The hierarchy stays the same here. Pronunciation is supplied at PF.

Why leave room for morphological operations?

Syntax and pronunciation may package features differently.

Operation What changes on the PF branch?
Fusion Two terminals become one insertion site.
Fission One terminal supports more than one insertion site.
Impoverishment A feature is deleted before insertion.

Fusion: two terminals, one insertion site

Italian am-a-te, ‘you.PL love’, present indicative.

A local PF subtree with separate T present and phi second-person plural becomes a subtree with a single T/phi terminal carrying present, second person, and plural.

The features survive; their packaging changes.

Fission: one terminal, two insertion sites

Toy paradigm: one agreement bundle, two endings.

Agr first-person plural under X splits into Person first and Number plural, realized separately as ka and mu. This is an invented paradigm.

ka ↔︎ [1st person] · mu ↔︎ [plural] → stem-ka-mu

Impoverishment: fewer features at PF

Spanish le dio un libro, ‘gave her a book’, but se lo dio, ‘gave it to her’.

In a simplified Spanish clitic analysis, DAT is removed from the first clitic before an accusative clitic. Both nodes remain, but the first no longer licenses le; se appears instead.

Delete [DAT] before an accusative clitic → se lo, not *le lo.

What changed?

Terminal sites Features at PF
Ordinary insertion Same Same
Fusion Fewer Combined
Fission, as drawn More Distributed across sites
Impoverishment Same in our example Reduced

Your turn: record

Keep a record of the meeting.

Please record the meeting.

  1. Draw a syntactic tree for each use in DM.
  2. State what is shared and what differs in the two structures.
  3. Propose a rule for the pronunciation in each context, including stress.

An answer: record

Root RECORD combines with n in the noun and v in the verb. The noun has initial stress; the verb has final stress.

For √RECORD:

  • In an n context, stress the first syllable: REcord.
  • In a v context, stress the second syllable: reCORD.

The vowels also differ: /ˈrɛkɚd/ versus /rɪˈkɔɹd/.

Your turn: separate

The files are separate.

Please separate the files.

  1. Draw a syntactic tree for each use in DM.
  2. Identify the shared root and the different categorizing heads.
  3. Propose a pronunciation rule for the final syllable. Does stress change?

An answer: separate

Root SEPARATE combines with a in the adjective and v in the verb. The adjective ends in schwa-t; the verb ends in ay-t.

For √SEPARATE:

  • In an a context, realize the ending as /ət/.
  • In a v context, realize the ending as /eɪt/.

Both forms retain initial primary stress.

Your turn: nominalizations

arrive → arrival

collect → collection

destroy → destruction

  1. Draw a DM tree for each noun.
  2. State a rule choosing the nominal suffix in each context.
  3. Explain any changes needed in the stem or its pronunciation.

An answer: nominalizations

Nominal heads combine with ARRIVE, COLLECT, and DESTROY. Their suffixes are al, ion, and ion; DESTROY uses the stem destruct-.
  • n → -al with √ARRIVE; n → -ion with √COLLECT and √DESTROY.
  • √DESTROY → destruct- in this nominal context.
  • Stem-suffix phonology gives collection /kəˈlɛkʃən/ and destruction /dɪˈstrʌkʃən/.

Why connect DM to production?

Both accounts distinguish building a grammatical representation from giving it a pronunciation.

Production question DM’s resources
What gets selected? Roots and grammatical features
How does it combine? Syntax and morphological operations
How does it sound? Vocabulary Insertion and phonology

DM inside the Formulator

Conceptualizer activates roots and features in List 1. These feed the Formulator: syntax, Morphological Structure, Vocabulary Insertion using List 2, and phonology. The output passes to the Articulator.

DM supplies the Formulator’s grammatical machinery.

What changes from a lemma-based account?

A familiar production account A DM production account
Select a lemma with grammatical properties. Select an abstract root and grammatical features.
Use lexical information to build the utterance. Combine roots with functional structure.
Retrieve the phonological form. Insert exponents into the resulting structure.

A root does not already contain its final category and inflected form.

Late pronunciation still needs early root identity

Selecting √KATZ ‘cat’ is different from selecting a generic ROOT slot.

  • The intended concept activates a particular root.
  • Syntax combines that root with grammatical structure.
  • Vocabulary Insertion supplies the appropriate pronunciation.

1. Start with the message

We want to say: Die Katzen schnurren.
‘The cats are purring.’

  • We are talking about cats.
  • There is more than one.
  • The listener can identify which cats.
  • They are purring now.

2. Select roots and grammatical features

Contribution from the message Input to grammatical encoding
Cats; purring √KATZ, √SCHNURR
More than one cat [plural], associated with √KATZ
Identifiable referents D[definite]
Present time in this example T[−past]

These are abstract inputs. Their pronunciations have not been inserted.

3. Build the nominal structure

Combine √KATZ with the nominal categorizer n.

Root KATZ combines with n, highlighted in red, to form nP.

The result is nP: a root in a nominal context.

4. Add plural Number

Combine the nominal structure with Num[plural].

Num plural is added above the existing nP containing root KATZ and n.

The structure now refers to multiple cats. [plural] is still a feature.

5. Add definiteness

Combine D[definite] with the plural nominal structure.

D definite is added above the existing NumP, creating a DP for the identifiable plural cats.

This DP identifies the cats the speaker is talking about.

6. Build the predicate and add its subject

Combine √SCHNURR with v, and place the cats’ DP in the subject position.

A vP contains the already built subject DP and a verbal structure made from root SCHNURR and v. A triangle abbreviates the subject's internal structure.

The structure connects who is purring with the purring event.

7. Add tense

Combine the predicate with T[−past], appropriate to the present-time message.

T minus past is added above the vP containing the subject DP and the purring predicate. T appears to the right before movement.

The clause now contains both the event and its temporal specification.

8. Establish the main-clause order

The verbal head raises through v → T → C. The subject DP moves to the front.

In the resulting CP, the subject DP occupies the first position and C contains the complex verbal head root SCHNURR plus v and T. A lower TP contains unpronounced copies.

Result: subject DP + finite verbal head.

9. Add an agreement site

On the pronunciation branch, add Agr, an agreement terminal next to T.

A local verbal head first has v and T minus past. Morphology adds an empty Agr terminal as a sister of T.

There is now a place for the verb’s agreement features.

10. Copy number; assign case

Copy [plural] from the subject onto Agr and D.

In the subject domain, D receives plural from the nominal structure and the DP has nominative case. In the verbal domain, Agr receives the subject's plural feature.

The subject DP also receives nominative case from its clausal context.

11. Fuse tense and agreement

Combine the adjacent T and Agr terminals into one realization site.

Separate T minus past and Agr plural terminals fuse into one T/Agr terminal containing both features.

[−past] and [plural] both survive.

12. Insert the root pronunciations

The Vocabulary supplies forms for the selected roots in their grammatical contexts.

Abstract root and context Inserted form
√KATZ in the nominal context /katsə/, Katze
√SCHNURR in the verbal context /ʃnʊr/, schnurr-

These pronunciations realize the roots that were selected earlier.

13. Realize the grammatical features

Realization site Exponent Result
D[definite, plural], nominative context /diː/ die
Nominal [plural], with √KATZ /-n/ Katze-n
T/Agr[−past, plural], verbal context /-ən/ schnurr-en

The two plural endings have different forms and different insertion contexts.

14. Prepare and articulate the sound sequence

/diː/ + /katsə/ + /n/ + /ʃnʊr/ + /ən/

↓

Die Katze-n schnurr-en.
‘The cats are purring.’

Phonology organizes the spoken forms. The articulator executes the plan.

Must planning follow the derivation?

We built nP → NumP → DP, then a clause, then movement.

Does a speaker have to plan it in that order?

Could planning begin with where the constituents will end up, with their lower positions already linked to those destinations?

Could a familiar frame be ready to reuse?

Suppose a speaker often uses this pattern:

[subject DP] + [finite verb] + …

Could its structure become familiar enough to retrieve as a frame, with places for new roots and features?

How much of the sentence would still need to be planned?

Plan the destination and its lower positions

Place the cats’ DPᵢ at its planned destination; reserve its lower positions.

A proposed CP planning frame has a subject DP at the front and an unfilled verbal complex in C. Below C, TP contains a linked subject trace and T prime. Inside T prime, vP contains another subject trace and a verbal structure with an unfilled root and a lower v trace. A lower T trace is also reserved. The DP's internal structure remains unfilled.

Filled: the subject’s identity and destination. Open: internal content.

How much structure comes with a trace?

Do we plan only the lower positions, or their surroundings too?

Possible plan What is reserved in advance?
Dependency links A destination and its linked lower positions.
Local structure Those positions plus sisters, complements, and attachment sites.
A fuller frame Clause structure, argument positions, and feature relations.

Where would you leave the tree underspecified?

What would distinguish these planning options?

Compare reuse of surface order, movement dependencies, and complement structure.

  • Would the same order help when the predicate needs different complements?
  • Would a frame help before its roots and number features are selected?
  • Which unfilled positions could remain open when speaking begins?

A reusable plan must still supply what DM needs for realization.

Now let something go wrong

Intended: “I wanted to tie up the dog.”
Produced: “I wanted to bark at the dog.”

German: anbind-en ‘tie up’ → anbell-en ‘bark at’.

The concept DOG can activate a related, unintended action.

Which part of the model needs access to these conceptual associations?

A wrong root can get the right morphology

Intended: versuchte … zu kommen, ‘tried … to come’.
Error onset: kam, ‘came’.

A simplified TP keeps T past fixed while root VERSUCH, try, is replaced by anticipated root KOMM, come. The resulting past form is kam.

The displaced root is pronounced in its new grammatical context.

What exactly was left behind?

Root in the finite position Past realization
Intended √VERSUCH ‘try’ versuch-te
Error √KOMM ‘come’ kam

The stranded element is [past]. Its pronunciation depends on the root that occupies the context.

Why would “the speaker just kept the suffix -te” miss the pattern?

A root can also land in a new category

Intended: einen Blick geworfen, ‘thrown a glance’.
Produced: einen Wurf geblickt, approximately ‘glanced a throw’.

The intended nominal root BLICK and verbal root WERF exchange contexts. WERF under n is realized as Wurf; BLICK under v receives the participial realization geblickt.

The grammar keeps working after an error

Altered representation Ordinary grammatical consequence
√KOMM in a past context Past realization: kam
√WERF in a nominal context Nominal realization: Wurf
√BLICK in a participial verbal context Participial realization: geblickt

Accommodation: the produced form fits the context created by the error.

Your turn: roots, tense, and pronunciation

Finite position: past Infinitival position
Intended schien ‘seemed’ zu drohen ‘to threaten’
Produced drohte ‘threatened’ zu scheinen ‘to seem’
  1. What moved between the two positions?
  2. What stayed in the finite position?
  3. What determines the two new pronunciations?

What does this architecture explain?

  • Selection: the message activates specific abstract roots and features.
  • Combination: syntax places roots in grammatical contexts.
  • Realization: ordinary morphology produces context-appropriate forms, even when an earlier choice was wrong.

Explain kam using these three steps. Where did the error occur?

Do roots and features need separate stores?

A speaker can produce the wrong root with the intended inflection, or the intended root with the wrong inflection.

Resource What it supplies
Root store Lexical content
Feature store Grammatical specifications
Vocabulary Phonological exponents

Independent errors can motivate separate representations.

Revising a developing utterance

“The girl reads … writes a letter.”

Conceptual planning ⇄ DM formulation ⇄ Local phonology → Articulation

Cognitive control can influence selection, mapping, and revision.

Which choice could be revised while the rest of the structure stays available?

Exit ticket

  1. Where does DM fit into this production architecture?
  2. Why must a root’s identity be available before its pronunciation?
  3. In versuchte → kam, what changes and what stays?
  4. Why can the error’s morphology follow from ordinary realization?

Sources: DM and production

Further reading: possible implementations

Extension: the same derivation in vectors

Die Katzen schnurren.

  • Maintain small pieces of structure, such as the subject DP and the finite verbal head.
  • Give roots and features content vectors.
  • Bind those vectors to positions within the structure.

The representations can be numerical while retaining grammatical distinctions.

A position is a role in the tree

A local tree for the nominal KATZ root labels NumP with the root address epsilon, nP with L, Num plural with R, KATZ with LL, and n with LR.

ε = root position · L = left branch · R = right branch

Bind a node’s content to its path. The same root could occupy a different path.

Bind content; retrieve with a position

\[ W = \sum_i \mathbf{v}_i\mathbf{p}_i^{\mathsf T} \qquad \widehat{\mathbf{v}}_i = W\mathbf{p}_i \]

\(\mathbf{v}_i\): what occupies a node. \(\mathbf{p}_i\): where it belongs.

In our utterance Content to bind
Subject’s nominal root position √KATZ
Subject’s Number position [plural]
Finite head’s verbal root position √SCHNURR
Tense position [−past]

A numerical example: copy plural to agreement

Columns are Num and Agr. Rows are plural and nonpast.

\[ W_{\text{before}}= \begin{bmatrix}1&0\\0&0\end{bmatrix} \quad\longrightarrow\quad W_{\text{after}}= \begin{bmatrix}1&1\\0&0\end{bmatrix} \]

Cue the Agr column: \(W_{\text{after}}\begin{bmatrix}0\\1\end{bmatrix} =\begin{bmatrix}1\\0\end{bmatrix}\) → [plural].

Number remains on the noun and is now available for verbal agreement.

Morphological operations become changes to bindings

DM operation One possible operation on the code
Fusion Replace two terminal bindings with one combined feature bundle.
Fission Rebind specified parts of one bundle to separate terminals.
Impoverishment Remove a specified feature from a PF-side content vector.

For our verb: T [0, 1] + Agr [1, 0] → T/Agr [1, 1].

Retrieve the bindings and realize them

Retrieved content and context Vocabulary result
√KATZ, nominal context /katsə/
[plural], at this nominal site /n/
√SCHNURR, verbal context /ʃnʊr/
[plural], at fused T/Agr /ən/

Together with D → /diː/: Die Katzen schnurren.

Could an LLM contain comparable information?

Two things to look for:

  • A direction or subspace that distinguishes a grammatical feature across many different words.
  • Attention-head outputs or other components that help use that information when predicting a form.

A syntactic head such as Num and an attention head are different kinds of object.

Begin with controlled morphological contrasts

Number Prefix Candidate continuation
Singular Die Katze schnurrt
Plural Die Katzen schnurren
Singular Der Hund bellt
Plural Die Hunde bellen

Train an analysis on some roots; test it on different roots and forms.

Find a candidate number direction

At layer \(\ell\), compare paired representations:

\[ \mathbf{d}_{\ell}=\operatorname{mean}_{\text{pairs}} \left(\mathbf{h}_{\ell,\mathrm{PL}}- \mathbf{h}_{\ell,\mathrm{SG}}\right). \]

Can a simple readout recover number for roots it has never seen in training?

Then ask which head outputs and layers carry or transform that information.

Change the candidate; measure the form choice

  1. Run singular and plural prefixes and save their activations.
  2. Patch a candidate head output or feature subspace from the plural run into the singular run at the matched position.
  3. Recompute the continuation probabilities.

Does preference shift from schnurrt toward schnurren?

What would connect the findings back to DM?

Finding What it would support
Number can be decoded across roots. A reusable grammatical distinction.
A selective intervention changes inflection. That information contributes to form choice.
The same feature predicts different exponents in different contexts. Context-sensitive realization.

Identifying Fusion, Fission, or Impoverishment would require additional, operation-specific tests.

Sources for the vector and LLM extension