Source
Frank, Goodman & Tenenbaum (2009), “Using Speakers’ Referential Intentions to Model Early Cross-Situational Word Learning”, LIT-815, read in NOTE-611.
NOTE-611 said “No THEORY: its findings are specific to a small annotated
corpus and to word learning, and the record holds nothing they correct.”
The second reason concerns existing THEORYs, and does not bar a new one.
The first is why this account is Proposed and says its scale in the
title.
What was actually shown
The model. The intended referent is a latent variable between the scene and the words: a subset of the objects present, possibly empty, with a uniform prior. A word is referential with probability γ, naming an intended object, or otherwise drawn from the vocabulary independently of the intention. The lexicon has a parsimony prior P(L) ∝ e^(−α|L|), and the MAP lexicon is found by stochastic search with simulated tempering. It is a computational-level model with no claim about mechanism.
The lexicon (Table 1). Intentional model: precision .67, recall .47, F .55; with κ and γ at their joint empirical-Bayes values, .57/.38/.46. Baselines (co-occurrence frequency, conditional probabilities, mutual information, IBM Model 1 in both directions): precision .06–.15, F .10–.22. For intended referents (Table 2) the model’s F is .58 (one-parameter .50), against .36–.48.
Behaviour. Mutual exclusivity: with “dax” uttered over a bird and a novel object, mapping dax to the novel object has the highest posterior (Fig. 2). Object individuation: model surprisal shows the same crossover as infants’ looking times in Xu’s (2002) two-word and one-word conditions (Fig. 3). Intention reading: given the speaker’s intended referent, the model maps Baldwin’s (1993) label to the hidden toy, where a salience model would map it to the toy in hand.
What could have come out otherwise. The baselines could have matched the model on precision, and the individuation crossover could have failed. Neither happened.
What this does not say
- That the advantage comes from non-referential words and empty intentions. That is argued informally after Tables 1–2. No ablation is reported (NOTE-611, C2, weak).
- That mutual exclusivity is evidence for intentions. The baselines also produce it (C3). It shows only that no dedicated principle is needed.
- That learners infer intentions in Baldwin’s case. The intention is supplied to the model as an observation (C5, weak).
- That it scales. About twenty minutes of video from two dyads, annotated by the authors, with three free parameters. The baselines score far below what Yu and Ballard reported on the same videos, which the authors attribute to transcript, annotation and gold-standard differences.
- Anything beyond object names. No syntax, no aspects of objects, no verbs or function words as targets.
Connections
- The informative speaker. The speaker here chooses uniformly among the lexicon’s words. Frank, Goodman, Lai and Tenenbaum (THEORY-472, LIT-824) make the speaker informative, which yields the size principle, and the rational-speech-act listeners of THEORY-172 invert such a speaker. What this account shares with them is a listener who inverts a generative model of the speaker, with a latent intended referent.
- Social and associative accounts. It takes the cross-situational problem from Siskind, Smith and Yu, and the social side from Baldwin, Bloom and Tomasello, and joins them in one generative model.