Lateral connections turn a static map into a settling process
BiLex represents a bilingual lexicon with three self-organizing maps: one phonetic map for Spanish words, one for English words, and one shared semantic map. In the original architecture, a map responds instantly with the unit closest to its input. The paper's extension adds short-range excitation and long-range inhibition between units, so a response settles over time instead of appearing at once.
That change matters because settling time gives the model a measurable quantity that can be compared with human behavior. The authors treat the number of settling steps as a condition analogous to lexical retrieval under time pressure. Increasing the maximum settling time, t_max, raised naming accuracy and reduced no-responses; at t_max = 400 the model balanced error matching and overall accuracy, and accuracy plateaued between 600 and 1000 steps in the L2-Spanish-early simulation.
Familiarity, typicality, and specificity are built into training
The model's semantic and phonetic maps were trained on 792 words: 128 superordinate words such as category labels and 664 subordinate words. Semantic feature values came from GPT-4o, and the authors report over 90% agreement between GPT-based semantics and earlier MTurk judgments for the same kind of yes/no feature questions. Phonetic representations were feature-based encodings of IPA transcriptions.
Three semantic variables were operationalized in specific ways. Familiarity was modeled with word frequency, drawn mainly from ClearPond and supplemented with COCA for English words; missing Spanish frequencies were filled from English counterparts, which the authors report as sufficiently predictive in the relevant range (R² = 0.584). Typicality was modeled through map organization and through associative connections scaled by a softmax weight over category members, so typical category words formed stronger connections. Specificity was modeled with superordinate-subordinate associative connections, using GPT-4o answers to whether a subordinate word is a kind of superordinate word.
The closest statistical match came from early L2-English learners
Only one of the four bilingual groups had enough participants for statistical testing: the L2-English-early group, with n = 12. In Spanish naming, the model's accuracy did not differ significantly from humans (68.3% vs. 62.9%), and no-response, superordinate, associative, coordinate, and unrelated errors also showed no significant differences. The exception was phonological errors, which the model produced more often (6.7% vs. 3.5%).
In English naming, the gap was larger. Model accuracy was significantly lower than human accuracy (55.0% vs. 74.7%), and no-response, associative, phonological, and unrelated errors all differed significantly. Coordinate errors (5.0% vs. 4.3%) and superordinate errors (1.7% vs. 0.7%) did not differ significantly. The authors attribute the weaker L2 performance to training structure: the L2 phonetic maps began training later than the semantic and L1 maps, leaving them less well formed.
Frequency slopes point the same way but are mostly not reliable
Human naming showed positive frequency-accuracy slopes in all four L2 naming conditions, meaning higher-frequency words were named more accurately. The model also produced positive slopes in every case, which the authors read as qualitative agreement. Quantitatively, however, only two model slopes were statistically reliable, both in the L2-English-early group: Spanish L1 naming (β = +1.84) and English L2 naming (β = +1.32). The remaining six slopes were smaller and not reliable.
The authors connect this weakness to map capacity. Across the four bilingual models, the relationship between word frequency and map-unit ownership was stronger in semantic maps (about r² = 0.6) than in phonetic maps (about r² = 0.1 to 0.3). They suggest larger phonetic maps and more training data would give frequency more representational room.
Error types trace back to different parts of the architecture
The paper assigns each error type a distinct proposed source. No-responses are retrieval failures; superordinate errors arise from specificity-modulated associative connections; semantic errors come from laterally connected semantic activations; phonological errors come from lateral interactions in phonetic space; and unrelated errors reflect broader breakdown of the naming process. These are the authors' mechanistic interpretations, not separately isolated experimental tests.
The pattern across groups supports a semantic-level account more than a phonological one. In nearly every group and naming condition, human phonological errors ranged from roughly 2% to 13%, and the model's errors fell inside that range, though the model tended to sit at the high end. Where model accuracy was weaker, as in late-AoA L2 naming, no-responses became the dominant error. The authors conclude that the model captures semantic-level processes in bilingual naming and that phonological-level processes would need a more refined map.
Overall accuracy remains the binding constraint
The main limitation the authors identify is that the model cannot reach high naming accuracy. Its accuracy is usually around 80% to 90%, while human accuracy often reaches 96% to 97%. In late-AoA L2 naming simulations, accuracy was extremely low, at 2% and 15%, with no-responses as the primary output.
Evaluation is also limited in scope. Three of the four bilingual groups had only one to three participants and were compared qualitatively rather than statistically. The authors note that running multiple seeds per model and testing multiple naming items per Boston Naming Test frequency would strengthen the evaluation, and that more participants across L1/L2 age-of-acquisition groups are needed for statistical comparisons.
The semantic fit, not raw accuracy, is the paper's real claim
The most defensible reading of this paper is that lateral connections buy the model a mechanism for producing errors and response timing, and that this mechanism reproduces semantic-level naming behavior across bilingual histories better than it reproduces overall accuracy. The statistically tested group supports that reading in L1 naming, where accuracy and most error categories matched, while L2 naming exposed the model's weaker phonetic maps.
That distinction sets the priority for anyone using this work. The model is presented as a foundation for individualized models of bilingual lexical-semantic processing, with future work aimed at incorporating language exposure, proficiency, and confidence from LUQ data and at modeling bilingual semantic dementia and aphasia. Before that, the gap to close is representational capacity in the phonetic maps and in late L2 learning, because those are where accuracy and frequency effects fall short.
Sources used in this answer
Modeling bilingual naming performance through laterally-connected SOMs
Adding excitatory and inhibitory lateral connections to BiLex let the model produce settling-based response times and human-like naming errors, approximating Spanish-English bilinguals' Boston Naming Test accuracy and error types, with the strongest statistical match in early L2-English learners and clear shortfalls in overall accuracy and late L2 naming.