Contribution
Cross-situational models learned words from co-occurrence and treated the speaker’s intentions as noise or as attentional salience; social accounts appealed to intentions but left the word–object mapping mechanism unspecified. This paper gives one generative model in which the intended referent is a latent variable between the scene and the words, so that the learner infers what the speaker meant and what the words mean together. It is the first, by the authors’ account, evaluated both on corpus learning and on behavioural phenomena.
Key insight
Make the speaker’s intention the hidden cause of the words, and let the model say that a word was not about anything present. Then a learner can discount the many utterances that do not refer to the scene, and phenomena that looked like special word-learning principles fall out of how a referentially used word is generated.
Assumptions
- Object names only. Contexts are sets of basic-level objects; intentions are subsets of the objects present; the lexicon is a set of word–object pairs.
- Uniform prior over intentions, P(I_s|O_s) ∝ 1 over subsets of O_s, including the empty set.
- Bag of words. Words in an utterance are independent given I and L; syntax is ignored.
- Two word sources. Referential with probability γ (uniform among the lexicon’s words for an intended object, averaged over |I|), otherwise non-referential (drawn from the corpus vocabulary, weight 1 for words outside the lexicon and κ for words in it).
- Parsimony prior P(L) ∝ e
- MAP inference by stochastic search with simulated tempering; no claim about cognitive mechanism (Marr’s computational level).
Key results
- Table 1 (lexicon). 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 both directions): precision .06–.15, F .10–.22.
- Table 2 (intentions). Intentional model F .58 (one-parameter .50); baselines F .36–.48.
- Table 3. The learned lexicon, with correct entries marked against a gold standard that includes plurals and baby talk.
- Mutual exclusivity (Fig. 2). On the corpus extended with “dax” uttered with a bird and a novel object, the lexicon mapping dax to the novel object has the highest posterior: mapping it to the bird lowers the likelihood of every earlier “bird” utterance; mapping it to nothing makes the experimental utterance non-referential. Baselines also get this.
- Object individuation (Fig. 3). For Xu’s (2002) two-word and one-word conditions, model surprisal shows the same crossover as infants’ looking times. Baselines cannot represent this.
- Intention reading. Given the speaker’s intended referent as an observation, the model maps Baldwin’s (1993) label to the hidden toy; a salience model would map it to the toy in hand.
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | Joint inference of intentions and lexicon learns a more precise lexicon from child-directed video than association and translation models | moderate | Tables 1–2; two ten-minute videos, one annotator’s gold standard |
| C2 | The advantage comes from non-referential words and empty intentions | weak | informal argument after Tables 1–2; no ablation reported in the article |
| C3 | Mutual exclusivity needs no dedicated principle | moderate | Fig. 2; also produced by baseline models |
| C4 | Words can guide object individuation through inferred intentions | moderate | Fig. 3; qualitative crossover against one experiment |
| C5 | Representing referential intention is needed for Baldwin’s result | weak | the intention is supplied to the model rather than inferred; the argument is that salience models cannot represent it |
Method
The likelihood factors over situations: P(C|L) = Π_s Σ_{I_s ⊆ O_s} P(W_s|I_s, L) P(I_s|O_s). The lexicon with the highest posterior is found by stochastic search with simulated tempering. Comparison models produce a word–object score, thresholded at the value maximizing F; their intended referents are the objects whose words in the lexicon were uttered. Surprisal (−log P) links model probabilities to looking times, after Levy (2007).
Concepts
- referential intention: the subset of present objects the speaker intends to refer to in an utterance; may be empty.
- referential / non-referential use: a word is generated to name an intended object, or drawn from the vocabulary independently of the intention (verbs, function words, names of absent objects).
- intentional model: the authors’ name for their model.
Connections
It takes the cross-situational problem from Siskind, Smith and Yu and Smith, and its main baseline (IBM Model 1) from Yu and Ballard (2007), whose corpus videos it re-annotates; the social side is Baldwin, Bloom and Tomasello; Quine’s chimney metaphor frames the joint problem. Xu and Tenenbaum (2007) is the Bayesian word-learning precedent, and its size principle reappears as the informative-speaker likelihood in Frank and Goodman (LIT-821).
Bearing on the record
- To the RSA line. The informative speaker of LIT-821 and LIT-816 is not in this paper. What it shares with them is the listener as a Bayesian inverter of a generative model of the speaker, and a latent intended referent. Bergen, Goodman and Levy (LIT-809) cite a different 2009 paper (Frank, Goodman, Lai and Tenenbaum, CogSci) for the informative utility; that paper is not in the record.
- No THEORY: its findings are specific to a small annotated corpus and to word learning, and the record holds nothing they correct.
- No instruction for machine-learning practice; nothing for the anthology.
Limitations
- About twenty minutes of video from two dyads, annotated by the authors; the baselines’ scores are much lower than Yu and Ballard reported, which the authors attribute to transcript, annotation and gold-standard differences.
- Three free parameters; robustness is asserted across a systematic variation the article does not tabulate.
- Object names only; no syntax, no aspect of objects, no non-object words.
- The behavioural fits are qualitative orderings, and the Baldwin case supplies the intention instead of inferring it.
Open questions
- Does the precision advantage survive larger corpora and independent annotation?
- What does an informative, rather than uniform, speaker add to the learner’s likelihood? Later RSA work answers this for interpretation, not for this corpus task.