Emergent Language : Watching, then reading

Language models learn words from text alone, and need a great deal of it. A child learns words whilewatching what they refer to, from a few thousand sentences, and uses them to think about things nobodyhas shown it. We test whether a mind with no neural network can learn words the second way. LOCUSfirst acts in small worlds and builds its own runnable model of each. It then hears sentences while itwatches events, keeps a word only if how often the word is said tracks something it already has, andreads a new sentence by running it in that model. On 840 fresh sentences from 8 blind writers, sealedafter an independent audit, it judges sentences about events nobody made or described correctly 89.9%of the time, against 85.4% for the best worldless table reading the same words (+4.5 points, 4.2 se), 82.3%for the registered bar (+7.6 points, 6.4 se) and 77.3% for the cheapest rule. Small neural networks trainedon the same 1,200 lessons score 74.7% and 66.1%, below the 77.7% of a rule that always guesses thegadget’s usual colour. A world built from being told alone reaches 85.4%, 6.9 points above the same ruleapplied to what it was told. Two child effects were tested: fast mapping emerges, and mutual exclusivitydoes not. Relation words attach to an alliance concept the mind invented from behaviour; on fresh storiesevery stated alliance it read got the right sign, but the gain in prediction (2.5 se) fell short of all threeregistered 3 se tests. Real history prose and real children’s books are where it stops: on Thucydides itscores 43% against 60% for the majority guess. Handed perfect reading it places 53% of 33 states,against 61% for the majority side: real prose also needs a world that places allies of allies and wordmeanings that reach beyond the register taught, which is where Phase 3 begins