Measuring context dependency in birdsong using artificial neural networks

Abstract

The production of grammatically and semantically appropriate human language requires reference to non-trivially long history of past utterance, which is referred to as the context dependency of human language. Similarly, it is of particular interest to biologists how much effect past behavioral records of individual animals have on their future behavioral decisions. In particular, birdsong serves a representative case to study context dependency in sequential signals produced by animals. Previous studies have suggested that the songs of Bengalese finches (Lonchura striata var. domestica) exhibited a long dependency on previous outputs, while their estimates were upper-bounded by methodological limitations at that time. This study newly estimated the context dependency in Bengalese finch's song in a more scalable manner using a neural network-based language model, Transformer, whose accessible context length reaches 900 tokens and is thus nearly free from model limitations, unlike the methods adopted in previous studies. A quantitative comparison with a parallel analysis of English sentences revealed that context dependency in Bengalese finch song is much shorter than that in human language but is comparable to human language syntax that excludes semantic factors of dependency. Our findings are in accordance with the previous generalization reported in related studies that birdsong is more homologous to human language syntax than the entire human language, including semantics. Thus, this study supports the hypothesis that human language modules, such as syntax and semantics, evolved from different precursors that are shared with other animals. Significance Statement• We investigated context dependency in over 10-hour recordings of Bengalese finch songs using a neural network-based language model.• We proposed an end-to-end unsupervised clustering method of song elements (syllables) into statistically optimal categories.• Context dependency in the birdsong is shorter than that in English sentences but comparable with the dependency in English syntax excluding semantic factors.Making behavioral decisions based on past information is a crucial to how humans and animals live (Friston, 2003(Friston, , 2 2010 Friston and Stephan, 2007). This underscores the biological inquiry into how much effect past events have on 3 animal behaviors. These past records are not only limited to observations of the environment and other individuals 4 but also include each individual's behavioral history. A typical example is human language production where the 5 appropriate choice of words to utter depends on previously uttered words/sentences. For example, we can tell whether 6 was or were is the grammatical option after The photographs that were taken in the cafe and sent to Mary only 7 if we keep track of the previous words for a sufficient length, at least up to photographs, and successfully recognize 8 the two close nouns (cafe and station) as modifiers rather than the main subject. Similarly, semantically plausible 9 words are...

Long-range sequential dependencies precede complex syntactic production in language acquisition

“…Beyond human language, numerous other human behaviors [46][47][48][49][50][51], animal behaviors [52][53][54][55][56][57], animal vocalizations [37,[58][59][60][61][62][63][64][65][66], and other biologically-generated processes [25-27, 31, 67-70] have been described as being hierarchically organized or display long-timescale organization. Such behaviors range from the seemingly non-complex patterns of behavior exhibited by fruit flies [52,56] to tool usage in great apes [53,54].…”

Effects of short-term isolation on social animals' behavior: An experimental case study of Japanese macaque

“…It is to be noted that time-series data, such as human speech and text data, are often analyzed by a recurrent neural network (RNN) [41][42][43] whose input can have variable lengths in the time dimension (see also [44,45] for a more recent model of time-series processing and [25,46] for applications of RNNs to biological studies). In contrast, CNNs require their inputs to have the same length (when they return a single output [47][48][49][50] instead of converting the input to another time-series [51,52]).…”

Effects of short-term isolation on social animals’ behavior: an experimental case study of Japanese macaque

“…It is to be noted that time-series data, such as human speech and text data, are often analyzed by a recurrent neural network (RNN) [41–43] whose input can have variable lengths in the time dimension (see also [44,45] for a more recent model of time-series processing and [25,46] for applications of RNNs to biological studies). In contrast, CNNs require their inputs to have the same length (when they return a single output [47–50] instead of converting the input to another time-series [51,52]).…”