Symbolic Systems Forum - T. Florian Jaeger, Linguistics Department

T. Florian Jaeger, Linguistics Department, "No Altruism Needed: Dissociating Production and Comprehension in Language Processing"

ABSTRACT:

Proposals to explain aspects of linguistic structure in terms of

processing have generally focused on comprehension, arguing that some

grammatical feature of a language facilitates efficient parsing. Far less

attention has been given the role of production in shaping language.

However, speakers and hearers are subject to different performance

pressures when processing language (e.g., while a hearer can at best guess

what words are coming up while processing a sentence, a speaker usually

has at least some knowledge of what he is going to say before it

is put into words). So comprehension-based explanations require that

speakers are altruistic, i.e. they pay attention to the needs of the

hearer in formulating their utterances (e.g., by avoiding ambiguities).

One phenomenon that has been investigated making use of such an altruism

assumption is the optional absence of a relativizer in restrictive English

non-subject relative clauses (NSRCs), e.g.:

(1) Yesterday I ran into the guy (who) we met in Seattle.

(2) This is precisely the kind of talk (that) you may or may not attend.

An altruistic model predicts that the speaker inserts relativizers to

facilitate comprehension. Based on a study of 8,500 NSRCs from spoken and

written language (Penn Treebank III; Mitchell et al., 1999), we present

results that argue against such altruistic models. Instead, our results

favor a new Production Difficulty Alleviation (PDA) hypothesis: Using

relativizers provides time to alleviate production difficulties. In line

with the PDA, we also show that, if uttering a relativizer alleviates

production difficulties, speakers use a relativizer even if it does not

improve comprehension. The resulting dissociation between production and

comprehension complexity casts doubts on connectionist models of language

processing (e.g. MacDonald, 1999) that aim at reducing comprehension

complexity to frequencies in the input (and thereby to production

complexity).