1 · Stored phonological graph
Complete reachable subgraph. Every branch denotes InclPhon; shared nodes may appear on multiple branches.
MAKING LANGUAGE VISIBLE
From sound to meaning.
And back to words.
A spiking neural model of the mental lexicon
Explore how a neural network understands words and retrieves word forms for production. Eight test words in a 613-entry lexicon. Two perspectives.
INSIDE THE MENTAL LEXICON
How sounds become meaning
Similarity to a word representation Selected in associative memory
Choose a sound form and start the perception pathway.
LEXICAL NEIGHBOURHOOD · BEFORE ABLATION
Closest neighbours in the lexicon
Five nearest other entries among all 613 words. The highlighted row is the selected input. These are cosine similarities between undamaged word templates, not the neural activations above. They stay the same across tasks, playback times and ablation settings.
Loading lexical neighbours…
Cell colour: lower higher similarity (0–1). Bold names belong to the eight test words. Neighbours reflect the original WWT associations and phonological structure.
Stored word vectors, mathematical similarities and ablation. In this model, ablation leaves the stored word vectors and their mathematical similarities completely unchanged. It damages the neural processing of these vectors: activations may become weaker or distorted, and retrieval may fail or produce a different word. The matrices therefore always show the undamaged reference relationships. The bars above show what the neural model actually produces under the selected ablation settings.
TASK GOAL
Which concept does this sound form represent?
The perception pathway stops here. Word production starts only when you choose to continue.
CONCEPT-TO-PHONO PATHWAY → SYLLABIFICATION
From the retrieved word to ordered syllables
This stored phonological graph underlies the recursive construction of the word’s Semantic Pointer. Phonological similarities are calculated from the resulting vectors. The stored order of direct word links is treated as learned lexical syllable order.
Play production to inspect the retrieved word.
2 · Ordered lexical syllables
3 · Simplified syllable output
Stored graph readout after accepted neural word retrieval, not measured neural syllable activation. St = stressed; e = schwa. Model symbols are retained. Positions specify order, not durations or gesture phases. This is the endpoint of the production pathway. Graph and syllable readouts are stored with each recording; there is no automatic connection to the separate Phono-to-Gesture demonstration.
EXPLORE NEURAL DAMAGE
The mental lexicon
Choose one of five production cases below. The sliders display the recorded neuron percentages and cannot be moved. Then press Play production above.
Production · five comparison cases
Why these components? The concept buffer, concept-to-lemma associative memory and lemma buffer were selected following the dysfunction sites studied by Stille et al. (2020). Perception and word comprehension also play a role in the WWT framework, but the emphasis here is on word retrieval for production. These five configurations illustrate the current model; they do not reproduce the paper’s experimental conditions or results. Perception is shown as an undamaged reference.
One reference and four damage configurations. Values below are remaining neurons: Concept / Concept → Lemma memory / Lemma. Other production components remain at 100%.
Select a case, then press Play production above. Your selected word is retained.
Case explanations refer to Kanne. Activation values are mean dot products with its phonological pointer during 0.45–0.70 s, not probabilities. Other words may behave differently. All cases use the same fixed neuron selection.
All eight test words · Case 3: 6/8 fail · Case 4: 6/8 retrieved
At these fixed seeds, Case 3 fails for six of eight test words; Tasse and Topf are still retrieved. Case 4 retrieves six of eight; Stuhl and Teller remain below the retrieval criteria. Values are mean phonological target dot products over 0.45–0.70 s. These are demonstrations of word-dependent effects, not a general success rate for the 613-entry lexicon.
| Word | Case 3 · 50/55/50 | Case 4 · 50/55/75 |
|---|---|---|
| Tasse | 0.367 · retrieved | 0.621 · retrieved |
| Tasche | -0.001 · no retrieval | 0.474 · retrieved |
| Kanne | 0.001 · no retrieval | 0.325 · retrieved |
| Katze | 0.062 · no retrieval | 0.506 · retrieved |
| Teller | 0.001 · no retrieval | 0.001 · no retrieval |
| Topf | 0.508 · retrieved | 0.720 · retrieved |
| Stuhl | 0.000 · no retrieval | 0.201 · no retrieval |
| Schrank | 0.016 · no retrieval | 0.471 · retrieved |
State buffer Associative memory100% remaining = no ablation · 0% remaining = complete ablation
Settings apply to the next run. Each task simulates its own pathway; settings on the other side are retained for that task.
Randomly selected neurons are silenced throughout the run. The same reproducible selection is used when comparing damage levels. Ablation changes neuronal activity, not the lexical representations. This 613-entry model illustrates robustness and failure under damage; its ablation thresholds should not be generalized to a full human lexicon.
Architecture adapted from Kröger’s speech-processing model; ablation approach: Stille et al. (2020). Mental syllabary not shown.
About this visualisation +
A real neural model. The traces replay 48 precomputed Nengo runs: eight undamaged perception runs and five production cases for each of eight test words, with all 613 lexical entries participating in each run. Each model has 199,400 LIF neurons before ablation and covers 1.1 seconds. Only the selected recording is downloaded; no simulation runs on the web server. Vector seed: 10; neural seed: 11; ablation seed: 2026.
Similarity is not probability. The bars show dot products between the neural buffer state and the eight word templates. Related representations may also show elevated values. A separate small marker indicates the active associative-memory entry. All 613 entries participate in retrieval; only these eight are plotted. Concept and phonological similarities retain the original WWT network structure. Lemma vectors do not yet share word-class features.
Deliberately simplified. Sound forms are supplied as whole-word pointers. The IPA display uses a fixed broad transcription for each retrieved word. Acoustic analysis, neural syllable sequencing and a Dynart connection are not yet included.