Human Parsing
How do people parse? Do we have evidence that people use any of the models of grammar and parsing developed over the last four chapters? Do people use probabilities to parse? The study of human parsing (often called human sentence processing) is a relatively new one, and we don't yet have complete answers to these questions. But in the last 20 years we have learned a lot about human parsing; this section will give a brief overview of some recent theories. These results are relatively recent, however, and there is still disagreement over the correct way to model human parsing, so the reader should take some of this with a grain of salt.
An important component of human parsing is ambiguity resolution. How can we find out how people choose between two ambiguous parses of a sentence? While almost every sentence is ambiguous in some way, people rarely notice these ambiguities. Instead, they only seem to see one interpretation for a sentence. Following a suggestion by Fodor (1978), Ford et al. (1982) used this fact to show that the human sentence processor is sensitive to lexical subcategorization preferences. They presented subjects with ambiguous sentences like examples in which the preposition phrase on the beach could attach either to a noun phrase (the dogs) or a verb phrase. They asked the subjects to read the sentence and check off a box indicating which of the two interpretations they got first. The results are shown after each sentence.
The women kept the dogs on the beach
The women kept the dogs which were on the beach. 5%
The women kept them (the dogs) on the beach. 95%
The women discussed the dogs on the beach
The women discussed the dogs which were on the beach. 90%
The women discussed them (the dogs) while on the beach. 10%
The results were that subjects preferred VP-attachment with keep and NP-attachment with discuss. This suggests that keep has a subcategorization preference for a VP with three constituents: (VP –> V NP PP) while discuss has a subcategorization preference for a VP with two constituents: (VP –> V NP), although both verbs still allow both subcategorizations.
Much of the more recent ambiguity-resolution research relies on a specific class of temporarily ambiguous sentences called garden-path sentences. These sentences, first described by Bever (1970), are sentences which are cleverly constructed to have three properties that combine to make them very difficult for people to parse:
1. They are temporarily ambiguous: The sentence is unambiguous, but its initial portion is ambiguous.
2. One of the two or more parses in the initial portion is somehow preferable to the parsing mechanism.
3. But the dispreferred parse is the correct one for the sentence.
The result of these three properties is that people are "led down the garden path" toward the incorrect parse, and then are confused when they realize it's the wrong one. Sometimes this confusion is quite conscious, as in Bever's example; in fact this sentence is so hard to parse that readers often need to be shown the correct structure… Other times the confusion caused by a garden-path sentence is so subtle that it can only be measured by a slight increase in reading time…
The garden-path and other methodologies have been employed to study many kinds of preferences besides subcategorization preferences... Sometimes these preferences have to do with part-of-speech preferences (e.g., whether houses is more likely to be a verb or a noun). Many of these preferences have been shown to be probabilistic and to be related to the kinds of probabilities we have been describing in this chapter... The human-processor is sensitive to whether a noun is more likely to be a head or a non-head of a constituent, and also to word-word collocation frequencies... Syntactic phrase-structure frequencies (such as the frequency of the relative clause construction) play a role in human processing... The human processor is sensitive to a combination of lexical and phrase-structure frequency.
Besides grammatical knowledge, human parsing is affected by many other factors which we will describe later, including resource constraints (such as memory limitations), thematic structure (such as whether a verb expects semantic agents or patients) and semantic, discourse, and other contextual constraints. While there is general agreement about the knowledge sources used by the human sentence processor, there is less agreement about the time course of knowledge use. Frazier and colleagues argue that an initial interpretation is built using purely syntactic knowledge, and that semantic, thematic, and discourse knowledge only becomes available later. This view is often called a modularist perspective; researchers holding this position generally argue that human syntactic knowledge is a distinct module of the human mind. Many other researchers hold an interactionist perspective, arguing that people use multiple kinds of information incrementally. For this latter group, human parsing is an interactive process, in which different knowledge sources interactively constrain the process of interpretation.
Some researchers argue that, whatever the time-course of the use of linguistic knowledge, these constraints must be fundamentally probabilistic. For example, the Bayesian sentence processing model uses probabilities to explain the difficulty of examples above and similar garden-path sentences. In this model, the human language processor takes an ambiguous input sentence and computes multiple parallel interpretations. Each of the interpretations is assigned a probability, by combining PCFG probabilities, syntactic and thematic subcategorization probabilities, and other contextual probabilities using a Bayesian belief network…
The model assumes that people are unable to maintain very many interpretations at one time. Whether because of memory limitations, or just because they have a strong desire to come up with a single interpretation, they prune away low-ranking interpretations.
Excerpt from Ch. 12.5 of Speech and Language Processing: An Introduction to Natural Language Processing, Computational Linguistics, and Speech Recognition, a textbook written c. 2000 by Computer Scientists Daniel Jurafsky and James H. Martin, author of A Computational Model of Metaphor Interpretation.
