Is Humpback Whale Song a Language?

The humpback whale, Megaptera novaeangliae, produces the most complex vocalisations of all 77 cetacean species, which have been dubbed by Payne and MacVay (1971) as “songs”. These songs are hierarchical in nature, with rules seemingly governing their organisation and evolution over breeding seasons. No one hypothesis of the song’s function adequately explains its complexity and structure, except perhaps for the theory that it constitutes the first non-human language yet discovered. Buck and Suzuki (1999) have applied Information theory to analyse a sample of humpback song converted into a stream of symbols using a self-organising neural network (SONN). This theory can be used to determine the maximum amount of information contained within a coded sequence by the unpredictability of the next symbol. Different assumptions can be made about the nature of the sequence; the next symbol is randomly determined (thus no hierarchical structure is possible within the sequence), or the probability of the next symbol is dependent on the previous one, or two symbols (0th, 1st and 2nd Order Markov models respectively). It was found that a first-order assumption could not reasonably model humpback song, meaning that humpback song possesses a hierarchical structure suggestive of language. The low rate of information transmission, about 0.1 – 0.6 bits per second, may ensure reliable communication over long distances in noisy, unpredictable acoustic conditions.

Language, as opposed to simple communicative signals, has a deep-structure that is both hierarchical and recursive. This recursiveness means that words sometimes greatly separated in a sentence must agree (for example in gender or plurality); so called “long-distance dependency”. Words of the same category (for example adjectives or nouns) behave similarly with respect to syntactically correct positioning within a sentence, and thus language also contains a lexical category structure. Elman (1992) has demonstrated that a partially recursive artificial neural network trained on a set of English sentences can detect and internally represent both recursive and lexical category structure.

This study constitutes a novel approach to analysis of humpback song for signs of language. A network architecturally identical to Elman’s was used to test song for evidence of lexical category structure, detection of which would provide support for the language hypothesis.

Download full report as a pdf