Title of Paper: The segmentation of speech and the emergence of
language structure: survival of the fittest in a statistical environment

Abstract:
Empirical observations indicate that the ability to communicate  with structured 
language rather than an unstructured stream of words confers an evolutionary 
advantage. This paper looks at the process by which this ability could have 
developed, and shows how an intermediate  stage was likely to evolve, with 
its own advantage. Well known tools from Information Theory are used in a 
novel way to evaluate the effectiveness of communication when language is 
or is not structured.

The structure of language can be seen as a tertiary form. First there are 
relationships between adjacent words, which can partially be modelled by Markov 
processes. Secondly, words can be grouped together creating constituents, which 
can then be organised in a hierarchical syntactic structure. Thirdly, there are 
relationships between elements of constituents. In this paper we consider the 
grouping together of words into constituents. We will show that the transmission 
of information is more efficient when a stream of words is appropriately 
segmented, and that such segments are related to syntactic elements.

Human speech capabilities have evolved for speed, reliability and scope, but at 
a significant physiological cost. The ability to communicate effectively must 
have outweighed the concomitant disadvantages. Given this evidence of selection 
for efficient communication we assume that speech will have developed so that 
messages can be encoded and decoded as efficiently as possible.

To evaluate ease of decoding, the entropy  of sequences of spoken words are 
compared under varying conditions. Lower entropy means that successive words in 
a sequence are more easily predicted, and entropy declines as more of the 
context is taken into account. The  contribution we make is to show that 
entropy can also be reduced by explicitly representing appropriate segmentation 
markers. This appropriate segmentation is produced by observed discontinuities 
in spontaneous spoken English. There is a relationship between prosody and 
syntactic structure, and these prosodic segments can usually be seen as 
syntactic elements.

This investigation was carried out using the MARSEC corpus of spoken English, 
annotated with prosodic markers. Unscripted News broadcasts were used. Words 
were mapped onto part-of speech tags, then the entropy was calculated with and 
without major and minor discontinuities. Entropy declines when discontinuities 
are represented. If an arbitrary pattern of segmentation is imposed, the entropy 
rises sharply.

As language has evolved, we expect selection pressure to encourage methods of 
efficient communication, and our work indicates that this includes decomposing 
the speech stream into constituents with some syntactic cohesion. This process 
is a step on the way to the development of a hierarchical language structure, 
but the evolutionary advantage of reaching this stage can be explained without 
invoking the merits of a fully developed language structure.

References include Arnfield; Bell, Cleary and Witten; Charniak; 
Cover and Thomas; Fang and Huckvale; Jelinek; Lieberman; Lyon and Brown; 
Morgan and Demuth; Ostendorf and Vielleux; Shannon; Taylor and Black.