nucleation

LIT-815: Using Speakers' Referential Intentions to Model Early Cross-Situational Word Learning

2 min read

Properties4
StatusActive — read 2026-10-09 (NOTE-611); worth reading as the model in which a word learner infers the speaker's referential intention and the lexicon jointly: words are generated from an unobserved intended referent (a subset of the objects present, possibly empty) through the lexicon, or non-referentially, and the MAP lexicon is found by search. On two ten-minute CHILDES videos it learns a far more precise lexicon than association and translation baselines (F .55 against at most .22), and it reproduces mutual exclusivity, Xu's word-driven object individuation and Baldwin's intention-reading result. Read it as a word-learning paper: it has no model of an informative speaker; Bergen, Goodman and Levy (2012) cite a different 2009 paper, by Frank, Goodman, Lai and Tenenbaum, for that.
Tagscognition, linguistics, probabilistic-modeling
Cited as “paper” byNOTE-611 — Referential intentions in word learning
Cited as “source” byTHEORY-542 — A word learner that infers speakers' referential intentions jointly with the lexicon, and lets words be uttered non-referentially, learns object names from child-directed video more precisely than co-occurrence and translation models, and mutual exclusivity, word-guided object individuation and intention-guided mapping follow from it without principles built in for them

Graph View

Michael C. Frank, Noah D. Goodman and Joshua B. Tenenbaum (2009), Psychological Science 20(5):578–585 — DOI-10.1111/j.1467-9280.2009.02335.x

Key takeaways

  • Joint inference. The corpus is a set of situations, each with the objects present O, an unobserved intention I ⊆ O (possibly empty) and the words W. P(L|C) ∝ P(C|L)P(L), with P(L) ∝ e^(−α|L|) and P(W|I,L) = Π_w [γ Σ_{o∈I} P_R(w|o,L)/|I| + (1 − γ) P_NR(w|L)]. A word is referential (chosen uniformly among the lexicon’s words for an intended object) or not (drawn from the vocabulary, with lexicon words down-weighted by κ).
  • Corpus result. On two CHILDES Rollins videos, annotated by the authors, the MAP lexicon has precision .67, recall .47, F .55; the best baseline (IBM Model 1, word|object) reaches F .22. The intentions inferred have F .58 against at most .48. The advantage holds across the three free parameters and when two are set by empirical Bayes.
  • Where the precision comes from. Words can be uttered non-referentially, and an utterance can have an empty intention, so inconsistent co-occurrence is explained away rather than learned.
  • Behavioural coverage. Mutual exclusivity emerges as a soft preference for one-to-one lexicons (shared with simple baselines). Xu’s (2002) individuation result comes out as a crossover in surprisal: two words make two objects less surprising. Baldwin’s (1993) result follows once the speaker’s intention is given to the model; a salience model cannot get it.
  • Not an informative-speaker model. The speaker here picks uniformly among the words linked to a referent; nothing makes her choose words to be informative. Bergen, Goodman and Levy (LIT-809) cite Frank, Goodman, Lai and Tenenbaum’s 2009 CogSci paper, not this one, for that step; the Lai paper was not read.

Standing in the record

Filed on 2026-10-09 at the owner’s request, from the manuscript bibliography of 2026-10-09 (work what-survives-translation): one of the works the manuscript considered and dropped from its final reference list.

Read on 2026-10-09 (NOTE-611). If the bibliography meant the paper that founds the informative-speaker likelihood of the RSA line (LIT-821, LIT-816), it is the CogSci paper with Lai, which the record does not hold; this one is the word-learning model.

Backlinks