The Communicative Lexicon Hypothesis - eScholarship
Edward A. F. Gibson, Steven T. Piantadosi, Harry J. Tily · Proceedings of the Annual Meeting of the Cognitive Science Society · 2009
The Communicative Lexicon Hypothesis Steven T. Piantadosi†, Harry J. Tily‡, Edward Gibson† { [email protected], [email protected], [email protected] } † MIT Department of Brain and Cognitive Sciences ‡ Stanford University Department of Linguistics 43 Vassar Street, Building 46, Room 3037 450 Serra Mall Cambridge, MA 02139 Stanford, CA 94305 Abstract Recent work suggests that variation in online language produc- tion reflects the fact that speech is information-theoretically ef- ficient for communication. We apply this idea to studying the offline, structural properties of language, asking whether lexi- cal properties may similarly reflect communicative pressures. We present evidence for the Communicative Lexicon Hypoth- esis (CLH): human lexical systems are efficient solutions to the problem of communication for the human language processor. While the relationship between sounds and meanings may be arbitrary, pressure for concise and error-correcting communi- cation — within the constraints imposed by human articula- tory, perceptual and cognitive abilities — has influenced which sets of phonological forms have emerged in the lexicons of hu- man languages. We present two tests of the CLH: first, we show that word lengths are better predicted by a word’s av- erage predictability in context than its overall frequency. Sec- ond, we show that salient (lexically stressed) parts of words are more informative about a word’s identity, in English, German, Dutch, Hawai’ian, and Spanish. Keywords: Rational analysis; lexicon; word length; lexical stress; surprisal. Introduction Many human cognitive systems appear to implement good solutions to the problems they are required to solve (Oaks- ford & Chater, 1999). Decay patterns in human memory, for instance, can be interpreted as modeling the probability of needing to retrieve a given element from memory (Ander- son & Milson, 1989), and patterns in phonetic perception can be derived by considering a Bayes-optimal phoneme recog- nizer (Feldman & Griffiths, 2007). Along these lines, Hock- ett (1960) identified thirteen features of human language that make it well-designed for communication. For instance, lan- guage can be conveyed in the vocal-auditory channel, leav- ing the rest of the body free to simultaneously perform other tasks, and language productively allows users to create never- before-uttered sentences that are immediately comprehensi- ble to other speakers of the language. Another property de- sirable for communication is that more frequent words are shorter: Zipf (1935) argued that this means on average peo- ple can expend less effort and communicate more efficiently, since the most commonly uttered words take the least effort and time to articulate. A good communicative system must balance several mu- tually incompatible goals. For instance, a language should communicate meanings as concisely as possible, so words should be short. But a language should communicate mean- ings unambiguously and with as little confusion as possible, so words should sound as different as possible. Unfortu- nately, the limits of human ability make it impossible to sat- isfy both of these objectives simultaneously: the number of short wordforms that can be differentiated accurately by the human articulation and perception systems is very limited. At one extreme, a language could have only one wordform which is used for every meaning. Such a language would be highly ambiguous but very concise. At the other extreme, one could have long and distinct wordforms, which would be different from each other and thus not confusable even with considerable noise, but utterances would then become very long. A better solution would be a language with some inter- mediate number of wordforms that sound distinct enough to be identified accurately in context, but are not so distinct that utterances become overly long. Good communicative features can be found at the pho- netic level within the speech channel, such as shortening and reducing less informative parts of words (e.g. Aylett & Turk 2004; Bell et al., 2003; Jurafsky et al., 2001; Pluy- maekers, Ernestus & Baayen, 2005; van Son & Pols 2003). Here we ask whether—similar to known properties of speech production—characteristics of the lexicon itself may reflect communicative pressures. We present evidence for the com- municative lexicon hypothesis (CLH): human lexical systems are efficient solutions to the problem of communication for the human language processor. It is important to emphasize that the CLH makes predic- tions only relative to the limitations and capacities of the hu- man language processor; for instance, it would likely be pos- sible to design a better system for communication if humans had superior perceptual or cognitive abilities. The CLH holds that the lexicon is particularly well-structured for the specific communication mechanisms that humans use. Kuperman, Ernestus & Baayen (to appear) present a finding that exem- plifies the kind of property we would expect from language under the CLH. They show that for the four languages they study—English, Dutch, German, and Italian—speech units with medium duration are used more frequently than those with particularly high or low duration. Kuperman and col- leagues argue that particularly long sound speech units are inefficient because they are harder to produce and make ut- terances longer, but particularly short speech units are also inefficient since they may lead to more frequent mishearing. The studies we present here test two more predictions of the CLH: (i) word length should be better predicted by a word’s typical predictability in context than by its raw fre- quency, (ii) salient parts of a word should be more informa- tive about the word’s identity.