Recursive syntactic pattern learning by songbirds

Abstract

Humans regularly produce new utterances that are understood by other members of the same language community. Linguistic theories account for this ability through the use of syntactic rules (or generative grammars) that describe the acceptable structure of utterances. The recursive, hierarchical embedding of language units (for example, words or phrases within shorter sentences) that is part of the ability to construct new utterances minimally requires a 'context-free' grammar that is more complex than the 'finite-state' grammars thought sufficient to specify the structure of all non-human communication signals. Recent hypotheses make the central claim that the capacity for syntactic recursion forms the computational core of a uniquely human language faculty. Here we show that European starlings (Sturnus vulgaris) accurately recognize acoustic patterns defined by a recursive, self-embedding, context-free grammar. They are also able to classify new patterns defined by the grammar and reliably exclude agrammatical patterns. Thus, the capacity to classify sequences from recursive, centre-embedded grammars is not uniquely human. This finding opens a new range of complex syntactic processing mechanisms to physiological investigation.

Figure 1. Grammatical forms

a, Finite-state form (AB)ⁿ. b, Context-free form AⁿBⁿ. Both grammars describe patterned sequences of elements (lower-case letters) of the sets ‘A’ and ‘B’. Longer strings of the form (AB)ⁿ, where n gives the number of AB iterations, are produced by appending elements to the end of an n - 1 sequence. Longer strings with the form AⁿBⁿ are produced by embedding elements into the centre of an n - 1 sequence. Learning of and generalization to an AⁿBⁿ pattern implies the capacity to process syntactic structures generated through recursive centre-embedding. Black arrows denote insertion points for higher-order sequences. Brightly coloured squares mark the ‘AB’ phrase inserted at each order. Different hues denote different elements.

Figure 2. Classification of grammatical pattern stimuli

a, b, Sonograms (frequency range 0.2-10.0 kHz; scale bars, 1 s) showing four of the eight sequences constructed from the finite-state grammar (AB)ⁿ (a) and the context-free grammar AⁿBⁿ (b) used in the initial FSG versus CFG pattern classification training with n = 2. Similarly coloured boxes mark the same motifs in multiple sequences. The position of a motif within a sequence is arbitrary with respect to its subscript label. See Supplementary Information for complete stimulus patterns and sonograms. c, Acquisition curves for the baseline FSG/CFG classification, showing mean d’ over the first 250 blocks (100 trials per block) for birds that learned quickly and were subjected to further testing (green), birds that learned slowly (black) and birds that did not reach the accuracy criterion (red; see Methods). d, Mean d’ (±s.d.) on the baseline CFG versus FSG classification task at asymptote. Open circles show means from individual birds. Colours and groups as in c.

Figure 3. Generalization to new FSG and CFG sequences

a, Mean d’ (±s.e.m.) for transfer from the training to new FSG and CFG stimuli (turquoise, mean performance over the five blocks of trials preceding transfer; blue, performance in the first five blocks after transfer; 100 trials per block). Performance was stable across these post-transfer blocks (F_3,4 = 1.15, P = 0.35, repeated measures ANOVA), then increased gradually to pre-transfer levels (not shown). All mean d’ values shown are significantly greater than zero (see text). Acquisition of the transfer stimuli was much faster than for the original training sets (12.50 ± 3.11 blocks to criterion (mean ± s.e.m.), range 8-15 blocks; 100 trials per block), which can be attributed partially to generalization across the CFG and FSG classes. b, Mean d’ (±s.e.m.) during grammatical probe sessions. Birds correctly classified new AⁿBⁿ and (AB)ⁿ sequences when n = 2 (blue), n = 3 or n = 4 (purple). Classification accuracy was significantly above chance for all three types of probe sequences (mean d’ for n = 2, 1.63 ± 0.39; see text for n = 3, n = 4). Classification of the baseline training stimuli (turquoise) was well above chance for all three conditions (mean d’ ≥ 2.39, s.d. ≤ 0.25). The drop between training and probe stimulus classification was significant in only the n = 4 condition (P < 0.05, Mann-Whitney U-test), suggesting that these sequences were more difficult to classify correctly than the other grammatical test sequences (see Supplementary Information).

Figure 4. Agrammatical controls for alternative strategies

Mean d’ (±s.e.m.) values for comparisons among the AAAA, BBBB, ABBA and BAAB agrammatical stimuli, to rule out the use of alternate solution strategies. For primacy (see main text), AAAA and ABBA should be classified similarly to new n = 2 CFG and FSG patterns, respectively, presented during the same probe sessions (Methods). For recency (see main text), BBBB and BAAB should be classified similarly to new n = 2 CFG and FSG patterns, respectively. If starlings are listening for the presence of a B/A motif transition (see text), then the d’ value comparing BAAB and ABBA to AAAA and BBBB should be similar to that for new n = 2 CFG and FSG patterns. d’ for the new n = 2 CFG and FSG probe stimuli (dark blue) was significantly higher than that for all three control comparisons (light blue; asterisk indicates P < 0.05 for all cases, paired t-test).