Elsevier

Cognition

Volume 107, Issue 2, May 2008, Pages 763-774
Cognition

Brief article
Syntactic structure and artificial grammar learning: The learnability of embedded hierarchical structures

https://doi.org/10.1016/j.cognition.2007.09.002Get rights and content

Abstract

Embedded hierarchical structures, such as “the rat the cat ate was brown”, constitute a core generative property of a natural language theory. Several recent studies have reported learning of hierarchical embeddings in artificial grammar learning (AGL) tasks, and described the functional specificity of Broca’s area for processing such structures. In two experiments, we investigated whether alternative strategies can explain the learning success in these studies. We trained participants on hierarchical sequences, and found no evidence for the learning of hierarchical embeddings in test situations identical to those from other studies in the literature. Instead, participants appeared to solve the task by exploiting surface distinctions between legal and illegal sequences, and applying strategies such as counting or repetition detection. We suggest alternative interpretations for the observed activation of Broca’s area, in terms of the application of calculation rules or of a differential role of working memory. We claim that the learnability of hierarchical embeddings in AGL tasks remains to be demonstrated.

Introduction

A fundamental issue in language acquisition research concerns which rules children develop as a part of their grammatical knowledge and how these rules may be discovered (e.g., Chomsky, 1957, Reali and Christiansen, 2005). Artificial grammar learning (AGL) is a potentially valuable paradigm for determining processes of rule learning, both in terms of what structures are learnable (e.g., Fitch and Hauser, 2004, Gentner et al., 2006, Newport et al., 2004), and which properties of the language facilitate learning of these structures (e.g., Gomez and Gerken, 2000, Newport and Aslin, 2004, Onnis et al., 2005).

A natural language structure that has attracted interest in recent AGL studies is hierarchical centre embeddings (Fitch & Hauser, 2004, henceforth F&H; Friederici et al., 2006, Gentner et al., 2006, Perruchet and Rey, 2005). In English, structures exemplified by The rat [the cat ate] was brown illustrate such centre embeddings, with additional embeddings possible, e.g., The rat [the cat [the boy chased] ate] was brown. Critically, these centre-embedded structures establish dependencies between constituents. Thus, such sentences have the structure A3A2A1B1B2B3, where the index values indicate the dependency between Ai- and Bi-elements. Such hierarchical embeddings are notoriously difficult to process in natural language (Bach et al., 1986, Blaubergs and Braine, 1974, Foss and Cairns, 1970). Thus, demonstrating their learnability in AGL-experiments is a notable success.

Hierarchical embeddings have also been claimed to be of theoretical importance, as they require a context free grammar1 to generate them and have been the focus of studies of human-unique structures in artificial language learning (Fitch et al., 2005, Hauser et al., 2002, Premack, 2004). In this respect they have been classified as different from the structures generated by finite-state grammars, for which local transitional dependencies can generate the sequence. F&H observed that humans could discriminate AAABBB-syllable sequences from ABABAB-syllable sequences (finite-state grammar), where A-syllables were spoken by a male human voice and B-syllables were spoken by a female. In contrast, cotton-top tamarins were insensitive to this distinction (though see Perruchet & Rey, 2005 for an explanation in terms of biological relevance, rather than structural distinctions between species). F&H thus claimed that humans were sensitive to the distinction between context free and finite-state grammars,2 whereas nonhuman primates were not.

Friederici et al. (2006) and Bahlmann and Friederici (2006) also contrasted learning of hierarchical (A3A2A1B1B2B3)3 and finite-state grammar (A1B1A2B2A3B3) sequences. Ai- and Bi-syllables were distinguished in terms of phonological properties (see Section 3.1). They observed that processing of hierarchical embeddings selectively activated Broca’s area (BA44/45) – typically involved in syntactic processing (see Kaan & Swaab, 2002) – whereas processing finite-state grammars selectively engaged the left frontal operculum. Broca’s area is thought to be phylogenetically younger (Friederici, 2004) and, in these studies, was claimed to be functionally specific to processing hierarchical embeddings.

We argue here that the data from the studies reported above can be explained by alternative learning strategies which do not imply hierarchical embeddings, but, instead, involve counting and matching the number of A- and B-elements. The relevance of AGL to human language becomes obscure without explicitly testing learning of hierarchical embeddings, as otherwise these sequences may not probe linguistically relevant processing. Our arguments critically hinge on the materials used in the testing phases of AGL tasks. The illegal sequences during testing should differ only in terms of their hierarchical structure if this is the property being tested. We will show, however, that such violating sequences differ also in terms of surface features enabling alternative, non-linguistic strategies to be applied during learning. We present data indicating that participants do indeed use alternative strategies instead of learning the rules of hierarchical embeddings in AGL-tasks. As such strategies depend on information not present in natural language centre-embedding structures, we challenge the evidence provided for such processing using current AGL tasks.

Section snippets

Counting vs. hierarchical processing

In the AGL studies of context free grammars reported above, knowledge of the precise hierarchical connections between elements was not explicitly tested. In F&H, participants had to distinguish alternating male/female voices from male sequences followed by female sequences. Perruchet and Rey (2005) replicated this study, and found that participants were unable to distinguish A3A2A1B1B2B3 from A3A2A1B1B3B2 sequences, if in the latter the dependencies between hierarchical elements were broken but

Experiment 1

To tease apart different strategies, we tested participants’ learning of hierarchical sequences using an AGL, but varied the testing conditions to compare learning a counting strategy to hierarchical dependency learning. In this experiment, we replicated Friederici et al.’s (2006) study comparing hierarchical sequences to number-violating sequences. We tested whether the learning effect in this study was due to counting by removing the hierarchical dependencies in sequences. We also tested

Participants

Ten students (5 female), aged 19–27, from the University of Münster participated for payment or course credit. All were native German speakers, right handed, and had normal or corrected to normal vision. None had participated in Experiment 1.

Materials

Participants were trained on hierarchical sequences. A-syllables began with voiced plosives and ended with –e/–i, and B-syllables began with voiceless plosives and ended with –o/–u. The plosives were paired according to their place of articulation: b-p,

General discussion

What do participants learn from AGL-tasks when trained with hierarchical sequences? We found no evidence for the learning of hierarchical embeddings in Experiments 1 or 2. This is clear from the data for the Hier-Scram-group of Experiment 1: Participants could not discriminate structures without hierarchical dependencies (scrambled structure) from those requiring dependency-learning (hierarchical structure). Moreover, the same strategy appeared to be used to distinguish both hierarchical and

Acknowledgements

This research was supported by an EU Sixth Framework Marie Curie Research Training Network Program on Language and Brain: http://www.hull.ac.uk/RTN-LAB/. Meinou de Vries, Padraic Monaghan, and Stefan Knecht are members of this training network. Many thanks to Pierre Perruchet for helpful discussions and suggestions with respect to Experiment 2, to Isabel Ellerbrock, Christin Döpke, Anna-Victoria Schmidt, Julia Elen Beumler, and Anne Schürmann for their valuable assistance with the experiments,

References (28)

  • L. Onnis et al.

    Phonology impacts segmentation in speech processing

    Journal of Memory and Language

    (2005)
  • E. Bach et al.

    Crossed and nested dependencies in German and Dutch: A psycholinguistic study

    Language and Cognitive Processes

    (1986)
  • Bahlmann, J., & Friederici, A.D. (2006). FMRI investigation of the processing of simple linear and embedded...
  • M.S. Blaubergs et al.

    Short-term memory limitations on decoding self-embedded sentences

    Journal of Experimental Psychology

    (1974)
  • View full text