Why this wrong word?

Lampe et al. (2024)

Utku Turk

LING 449U, How words are built

2026-09-24

What did earlier studies ask?

Which words are hard to name? What kinds of errors occur?

• Frequency and age of acquisition: mixed findings on whether errors are more frequent or earlier learned than targets.
• Typicality: people with semantic dementia often substitute a more typical category member (Woollams et al., 2008; Woollams, 2012).
• Semantic features: words with more features can be named faster and more accurately (Taylor et al., 2012; Rabovsky et al., 2016).

But why say pelican, rather than peacock, for flamingo?

What can a wrong word tell us?

Someone sees flamingo but says pelican.

• The activation account: the representation of pelican became more strongly activated in the mental lexicon than the representation of flamingo.
• The question: which properties of the words, and of their relationship, could explain that activation pattern?
• The approach: compare targets and errors on psycholinguistic variables such as typicality, number of semantic features, and similarity.

Patterns in naming errors can help us infer how word knowledge is represented and activated during production.

Task: speeded picture naming

Show a picture → require a fast name → collect the mistakes.

• A short response deadline produces more naming errors in neurotypical speakers.
• Mirman (2011): errors often came from the target’s near semantic neighbours, more often than chance would predict.
• Lampe et al.’s next question: among possible wrong words, which semantic properties favour selection?

Being related to the target may get a word activated. What makes it win?

Earlier evidence: neighbourhood effects

More near neighbours: more errors. More distant neighbours: fewer errors overall.

The logic: compare each error with its target

Related words become active together. An error shows which alternative was selected.

Property	Comparison	Prediction
Typicality	How representative is the error of its category, relative to the target?	Error > target
Number of features	How many semantic properties does each word have?	Error > target
Semantic similarity	Is the error closer to the target than the target’s near neighbours are, on average?	Error similarity > neighbour average

Account for frequency, age of acquisition, familiarity, word length, imageability, and naming history.

Materials

• 78 participants analysed: native Australian English speakers.
• 297 colour photographs of objects, with high name agreement (>75%).
• McRae et al. (2005) norms: semantic features for calculating all three measures.
• Six presentation lists; at least two other-category pictures between pictures from the same category.

Compare real-word errors with known target names. Both words need the relevant norm data.

Procedure: start speaking within 600 ms

First: standard naming. Then standard and speeded naming, in counterbalanced order.

Practice: picture + beep at 600 ms. Try to beat the beep.

Experimental trial:

Fixation (500 to 1000 ms) → picture (2000 ms, no beep) → speed feedback (400 ms) → blank (1000 ms)

Say the first word that comes to mind. Prioritize speed over accuracy.

Pictures

Try it: name the pictures

Say the first name that comes to mind, out loud.

• Look at the + before each picture.
• Begin speaking before the picture disappears: 600 ms.
• Keep going if you make a mistake. There are eight pictures.

Loading pictures...

Fixation 1

+

Picture 1

Blank 1

 

Fixation 2

+

Picture 2

Blank 2

 

Fixation 3

+

Picture 3

Blank 3

 

Fixation 4

+

Picture 4

Blank 4

 

Fixation 5

+

Picture 5

Blank 5

 

Fixation 6

+

Picture 6

Blank 6

 

Fixation 7

+

Picture 7

Blank 7

 

Fixation 8

+

Picture 8

Blank 8

 

What happened under time pressure?

Picture exposure and response deadline

Can we know the answer from this experiment?

Results: Table 1

Broad category	Error type	Count	% of errors
Semantic errors	Coordinate	1,677	85.30
	Semantic other	92	4.68
	Part-whole relationship	23	1.17
	Associate	8	0.41
	Superordinate	3	0.15
Other errors	Unrelated	111	5.65
	Visual	29	1.48
	Phonologically related word	22	1.12
	Compound	1	0.05
	Total	1,966	100.00

What did they find?

1,966 errors in the primary analysis; 85.3% were same-category substitutions.

Prediction	Result
More typical than the target	Supported: adjusted difference +2.52, p < .01
More semantic features than the target	Supported: +0.61 features, p < .05
Closer than the near-neighbour average	Not supported: similarity difference −0.06, p < .001

Errors were often near neighbours, but not systematically closer than the average near neighbour.

The distributions: typicality and features

Similarity: what is the comparison?

Zero: the error is as similar to the target as the average near neighbour is.

Below zero: the error is less similar than that average.

Naming demo: image credits

Photographs from Wikimedia Commons, saved locally. Originals or Commons thumbnails; no content edits.

Picture and source	Creator	License
Banana	Paolo Neo	Public domain
Hammer	Evan-Amos	Public domain
Pear	MartinThoma	CC0 1.0
Scissors	Help e choose	CC BY-SA 4.0
Apple	Amada44	Public domain
Spoon	Ari Abitbol	Public domain
Orange	Evan-Amos	CC BY-SA 3.0
Lollipop	Evan-Amos	Public domain

Appendix: how are the measures calculated?

Start with features people listed for each word, such as has wings or flies (McRae et al., 2005).

• Typicality: for each feature, multiply the proportion of category members sharing it by the number of people who listed it for that word, then add the scores. Shared, often-mentioned features give higher typicality.
• Number of features: count the different listed properties, excluding category labels such as is a bird.
• Semantic similarity: compare two words’ feature profiles using cosine similarity. More overlap gives a higher score.