Many studies have found evidence that people really do understand sentences as a collection of phrases. In one study, subjects listened to auditory tapes of sentences, which were stopped at random places. When the tape stopped, the subject was asked to repeat as much of the sentence as possible from memory. Subjects were often able to repeat word-for-word the phrase that had been playing when the tape stopped, but their memory was much worse for words prior to the phrase boundary. It seems that as they worked on the current phrase, they needed to know it verbatim until they figured out its meaning. After they had gone on to the next phrase, they no longer needed to remember the last phrase word-for-word, so their verbatim memory for prior phrases was worse. For example, the following sentence can be divided into three phrases-- noun, verb, and object--that are each underlined: "The terribly spoiled girl should not have discarded her expensive sequined dress." If this sentence were stopped at "sequined," the subject listening would definitely recite "her expensive sequined" with no errors. However, the subject might make a few mistakes in the prior two phrases, perhaps saying something like, "The horrible spoiled girl shouldn't have thrown away her expensive sequined."

Further evidence comes from response-time studies. Subjects were shown a sentence on a computer screen one word at a time, and could press a key to advance to the next word when they were ready. So, they might spend 800 milliseconds looking at one word, and 1400 milliseconds looking at another. Researchers found that subjects tended to spend more time on the last word of a phrase, indicating that they were analyzing the phrase to take in its meaning before advancing to the next phrase. Using the example above, a subject might look longest at the words "girl," "discarded," and "dress," since they lie at the phrase boundaries.

In order to determine where phrase boundaries lie, people must analyze each word for its meaning and its grammatical placement. The process of figuring out each word's role in a sentence's syntax is called parsing the sentence. We can use semantics, the meanings behind words, to help parse a sentence. For example, many sentences are grammatically ambiguous, yet we interpret them in a straightforward way because of what we know about the words' meanings. In the sentence, "Time flies like an arrow," most people would agree that "time" is the subject phrase, "flies" is the verb phrase, and "like an arrow" is an adverb phrase. However, watch what happens to the sentence if I say: "Time flies like an arrow. But fruit flies, in contrast, like a banana." Now, we can reparse the sentence: "time flies" is the subject phrase, "like" is the verb, and "an arrow" is the object phrase. But, because this sentence is rather nonsensical, we normally parse it in the first way, not the second.

Some sentences contain ambiguities that are resolved in an unexpected way as the sentence continues. Imagine reading the following sentence one word at a time: "The secretary applauded for his efforts was soon promoted." At first, you think that the secretary is the one doing the applauding, but as you read the next phrase, you realize that the secretary is being applauded by someone else, making "applauded" an adjective phrase rather than a verb. The errors we make in parsing such sentences are called garden-path errors, because they lead us down the wrong path (the proverbial garden path). When subjects view garden-path sentences one word at a time, they tend to spend a long time looking at the word that reverses their first interpretation of the sentence. In the above example, they might get stuck on "was," since that word makes the previous interpretation of "applauded" as a verb impossible, and so requires the subject to figure out a new way to parse the sentence.

As we have seen, people tend to parse sentences as they go along, requiring them to make educated guesses about the role each word plays in the sentence. People tend to use several strategies to guide their decisions and help them parse efficiently. For one, they assume that a sentence will be active rather than passive; and indeed, English speakers do use active sentences more often than passive. Unfortunately, this strategy makes passive sentences a little more difficult to understand when we do encounter them. The garden- path sentence used above is written in the passive voice, so it works against this assumption. Morphemes can also guide our understanding of a sentence's structure; we know that words ending in the suffix "-ly" are usually adverbs, while words ending in "-ed" or "-ing" are often verbs. Furthermore, small function words can act as similar signals; if "the" or "a" precedes a word, we know that the word is probably a noun. The garden-path sentence above leaves out function words that might have been helpful, as in "the secretary who was applauded." The use of semantics to help us determine the syntax of an ambiguous sentence, as in the "time flies" example, is a top- down effect similar to those seen in other domains of perception where some interpretation is required (e.g. letters).