From an idea to a word

Griffin & Ferreira (2006), up to section 1.3

Utku Turk

LING 449U, How words are built

2026-09-03

Where we are

• What we want: how complex words get built while you are talking.

• We need a core piece: how a single, simple word gets out of your mouth.

• Let’s write down your questions first.

First, a game

Describe this picture

🐌 🍃

Write down random six words come to your mind.

• Nobody in this room wrote the same six words.

• It is clear that first two all of you shared: snail and leaves. but then?

Speaking is not listening, reversed

• Recognizing a word in your native language is fast, automatic, and needs no intention.
• Producing the same word needs an intention and takes up to five times longer.

Listeners start looking at the referent before the speaker has finished saying the word. Speakers need about 900 ms to start naming a pictured object.

Which is roughly how long it took you to get going on the snail.

Keep in mind that

• Almost everything in this chapter is about single, isolated utterances. We set conversations, narratives, turn taking aside for a second.
• Almost everything is about nouns.

I want you to keep on asking if any of this would easily survive in other situations?

The three steps

What are the ‘steps’ the chapter discusses?

The three steps

• Conceptualization, decide what to express
• Formulation, decide how to express it
• Articulation, say it

Properties in the chapter

• Property 1, the steps
• Properties 2 to 9, from meaning to word
• Properties 10 to 13, assembling the sounds
• Properties 14 and 15, how it all plays out in time

For today: we stop at Property 8, at the end of section 1.2.

Models disagree less than you would expect. They mostly agree on the facts and differ on the mechanism, and on which stage they bothered to model.

1.1 Word selection precedes sound assembly

Property 1

Speakers access words by meaning first, then work on the sounds.

Two kinds of evidence:

1. Speech errors come in types, and the types line up with the steps.
2. Time course in picture and word interference.

Errors come in types

Error type	Example
Semantic word substitution
Blend
Phonological word substitution
Morpheme substitution
Sound substitution

Errors come in types

Error type	Example
Semantic word substitution	“give him the yel- er the pink … the green jersey”
Blend	“Justi:ce, Justin and Travis”
Phonological word substitution	“the battle for the green journey [jersey]”
Morpheme substitution	“an awful lot of useless information. useFUL information”
Sound substitution	“the raid, uh, the roadie is accepting…”

Errors come in types

• Corpora like this are opportunistic. You get what people happen to say.
• The rates are low: roughly 2.5 word errors and 1.5 sound errors per 10,000 words of running speech.

Which errors do you think are undercollected, and why would that matter?

Most of the examples in Table 1 come from one cycling commentator at the 2005 Tour de France. Talking fast for six hours a day is very good for error corpora.

Why the types matter

• You can pick the wrong word and pronounce it perfectly.

• You can pick the right word and scramble its sounds.

• Those are two different failures, so there are two different steps.

Notice the fourth row. Morphemes substitute too, and they substitute separately from the stem. Hold onto that one.

Now we are going to run an experiment on you

The rules

1. A picture appears. Say its name out loud, immediately.
2. A word will appear on top of it. Ignore the word.
3. Do not think very hard.

Practice

RED

BLUE

GREEN

PURPLE

ORANGE

Practice 2

CLOUD

🪑

Website

https://www.utkuturk.com/pwi

Which one hurt?

The same trial, on a timer

picture

distractor

-400 ms -200 0 +200 +400

The time you introduce the distractor directly affects which step it interferes with.

What people previously found (idealistic)

• Two stages, and two separate levels of representation: word level and sound level.

1.2 Selecting a content word

Property 2: a web of words

• Semantic substitutions: calling a van bus.
• Blends: behavior plus deportment gives behortment.

Of course, we fail at selecting the ‘exact’ word.

Which raises the obvious question: what is a meaning, in these models?

Two possibilities

Non-decompositional

BIRD is one atomic lexical concept.

Neighbours come from links in a network: BIRD to ANIMAL to FISH to fish.

Decompositional

bird = HAS WINGS, HAS FEATHERS, SINGS SONGS.

Neighbours come from shared features: airplanes have wings, opera singers sing.

Two possibilities

Two possibilities

• Decompositional gets similarity for free: overlap in features.
• Non-decompositional has to build similarity out of network links.
• Both have to explain why the neighbourhood is context dependent. Blends only involve words that are interchangeable in that sentence.

For us: if a morpheme carries meaning, which of these two boxes does -ed go in? Does PAST have features? (Look at Kandel & Pañeda et al. 2025)

Property 3: similar meanings compete

• Words that intrude in substitutions share grammatical class, taxonomic category, and level of specificity with the target.
• Related distractors slow you down relative to unrelated ones.

But sometimes dog and bone do not compete. Why?

How models make words compete

• Activation only. Probability of selection tracks activation. Enough for error patterns, not enough for latencies.
• Lateral inhibition. The more active a word is, the harder it pushes the others down. Time to threshold is your reaction time. Kind of a race.
• WEAVER++. Two factors: a ratio of the target’s activation to everything else in the response set, plus a critical difference the target has to clear.

That last one makes a real prediction: one strong competitor is worse than two half strength ones.

Property 3, a small case: codability

• tooth, bomb: one dominant name, so fast.
• TV or television, weights or barbells: two good names, so slow.

Property 4: fenced in by grammatical class

• Nouns substitute for nouns, verbs for verbs.
  • What about interchangable words?
  • What about theories that says words are categoriless?

• Distractors interfere when they share grammatical class, verb transitivity, and other context relevant syntactic features.

• Even purely phonological substitutions tend to keep the category.

Property 4: one possible takeaway

The pressure to keep the grammatical class is stronger than the pressure to keep the meaning.

This is the first place syntax enters the model, and it enters as a frame with slots: a noun slot only accepts nouns.

Property 5: which meanings are easy

• Imageability and concreteness. vampire, wind, pine beat fear, sense, spirit.

• Predictability. The more predictable a word is in context, the fewer pauses and ums before it, and the faster the naming.

The twist: imageable words are produced faster and are more likely to be replaced by a semantically related word.

Richer meaning helps you find the word, and gives you more neighbours to trip over.

Property 5: exercise

• How can we explain these effects with decompositional or atomic meaning models?

Property 6: looking is not naming

• Distractor word banana while you say apple, that is interference.
• Distractor picture of a banana while you say apple, that is facilitation.

Words seem to get activated only when you intend to talk about the thing. Objects on their own do not push you into the lexicon.

Speakers even look right at an object while deliberately naming it wrong (Griffin & Oppenheimer 2003).

Property 6: except it is not clean

• Ignored objects show no semantic interference.
• But they do speed you up when their name sounds like your target, cap while naming a cat.

What does this suggest?

Every model routes phonology through the lemma. So if you got phonology from the ignored object, the lemma was active. And then where is the semantic interference?

Property 7: selection leaves a trace

• Repetition priming for object naming lasts months.
• Perseveration: name a zebra, then call the giraffe zebra.
• Semantic homogeneity: naming many animals in a row gets slower.

• None of this happens if you just read the word aloud or categorize it. You have to select it from a meaning.

Reading, phonological assembly, and sentence production all have learning models. Word production does not. The authors name this as a gap.

One more experiment. This one works on about half the room.

Do not say it out loud

What is the name of that tall guy who plays in Frankenstein?

If you have it, sit on it. If you are don’t remember it, raise your hand.

Hands up, answer these

1. What letter does it start with?
2. How many syllables?
3. What does it rhyme with?
4. Would you know it if I said it?

Property 8: stopping halfway

Tip of the tongue. You have the word and you cannot say it.

What people can still know in a TOT state:

• its first letter or sound, its number of syllables, words it rhymes with
• whether it is count or mass
• its grammatical gender in
• which auxiliary the verb takes

Anomia after brain damage looks the same: gender preserved, form gone.

What do we need for TOT in our models?

Where the morphology is hiding

The chapter is about content words: nouns, mostly. Yet:

• Property 1. Morphemes substitute in speech errors, on their own, away from their stems. useless for useful.

• Property 4. Grammatical class constrains selection harder than meaning does. Class is a morphosyntactic fact.

• Property 8. In a TOT you can report gender and count or mass status without the form. You are holding morphosyntax with no morphology attached.

Even for a “single” words that are seemingly “single” nodes we need morphological information

Three questions to carry

1. Every one of those observations is about a morpheme, and none of them get a mechanism in this chapter. Why not?
2. Is a compound one selection or two?
3. If Distributed Morphology is right and words are built by syntax, what happens to a stage called “word selection”?

Section 1.3 starts to answer the first one. That is Thursday.

Key takeaways

1. Deciding what to say is its own step, and it is slow.
2. Meaning first, sounds second, and never a shortcut between them.
3. Selection is competitive, and the competition is over similar meanings inside the same grammatical class.
4. The word can succeed while its form fails. That is a TOT.
5. The models mostly agree on the facts. They differ on the mechanism.

Bring questions

• Property 1 says meaning based selection comes first. What would it look like if that was not the case ?
• Property 6: how would you tell “the word was never activated” apart from “the word was activated and then suppressed”?
• Name three things wrong with our experiment demo.