Psycholinguistics James Myers April 16, 2004 Sentence comprehension OVERVIEW: 1. Memory for sentences 2. The basics of sentence parsing 3. Is syntactic processing modular? 4. Processing discontinuous dependencies 5. Interactionist models of sentence comprehension ============================================================= 1. Memory for sentences 1.1 Sometimes people remember sentences the way they remember words: they store both the meaning (semantics) and the form (syntax and lexical items). "To be or not to be...." (not: "Not to be or to be....") This happens most often when there is high "interactional content" -- that is, the sentences are meant to be performed, such as jokes, insults, or figures of speech (Keenan, MacWhinney & Mayhew 1977) 1.2 Usually, however, people seem to store just the meaning, and forget about the syntactic structure and the specific words. 1.2.1 Fillenbaum (1966): Subjects had to remember a long list of sentences, e.g.: "The window is not closed." Subjects were asked: Which of the following were in the list? (a) The window is not closed. (b) The window is closed. (c) The window is not open. (d) The window is open. Subjects usually chose (a) [so they did remember the form in this experiment] When they made an error, they were most likely to choose (d) [similar meaning, but different form] 1.2.2 Wanner (1974) Question: Will people remember the form of sentences if they don't know they will have to remember them? Subjects heard instructions for an experiment, but the experiment turned out to be about how they understood the instructions! The instructions included this sentence: "When you score the results, do nothing to [your correct] answer but [mark carefully] those answers which are wrong." Quiz: Subjects were asked (1) OR (2): (1) Did you hear "your correct" or "correct your"? (2) Did you hear "mark carefully" or "carefully mark"? Note: question (1) involves change in meaning; question (2) only changes form, not meaning Results: (1): perfect accuracy (2): accuracy at chance [just guessing!] Interpretation: subjects had forgotten the form, and only remembered the meaning. Criticism: the switch between "your correct" and "correct your" also involves a change in form, namely of the syntactic tree structure.... 1.2.3 Sachs (1967): Question: What is the time course for sentence memory? That is, when do people forget the form of sentences? Subjects were presented with a spoken passage. (Sachs 1974 used written language; Hanson and Bellugi (1982) studied ASL) When subjects are quizzed immediately after a sentence, they show very good memory for both the meaning and for the specific wording. When subjects are quizzed 40 syllables later, the memory for the wording fades (accuracy drops), and fades even more after 80 syllables. However, memory for the meaning does not fade. 1.3 If a sentence is not remembered in terms of syntactic structure, then how is it remembered? ...As a PROPOSITION. Proposition: a unit of meaning that contains a predicate ("function") and its arguments, e.g. f(x,y) = saw(Kim,Sandy) 1.3.1 Different sentences may involve the same proposition: "saw(Kim,Sandy)" "Kim saw Sandy" "Sandy was seen by Kim" "It was Sandy who was seen by Kim." "The one who saw Sandy was Kim." 1.3.2 A sentence may contain more than one proposition: "James tried to learn Chinese, and he had many problems." tried(James,learn(James,Chinese)) and had(James,many(problems)) 1.3.3 A sentence with unambiguous syntactic structure may nevertheless imply more than one proposition: "All that glitters is not gold." (a) for all x, if glitter(x) then not(gold(x)) (b) not for all x, if glitter(x) then gold(x) -- that is, there exists x such that glitter(x) and not(gold(x)) The distinction between syntactic "sentence" and semantic "proposition" is relevant below, and also for the next topic (discourse comprehension). 2. The basics of sentence parsing. People may forget syntactic structure, but syntax must be processed at some stage because it affects the meaning: "The dog bit the man" "The dog was bitten by the man" 2.1 PARSING: assigning elements of an incoming sentence into the appropriate syntactic structure. (see Carroll, p. 131) Parsing involves making decisions. Imagine you are hearing or reading the growing sentence below. At each point, how would you guess that the sentence might continue? Input Possible parses 那個... [S 那個 [VP ...]] ...? [S [NP 那個...] [VP ...]] ...? 很好看... [S 那個 [VP 很好看]] ...? 的人... [S [NP 那個 [S 很好看] 的人] [VP ..] ] ...? 要... ... [VP 要 NP] ...? ... [VP 要 [VP ...]] ...? 給你... ... [VP 要 [VP 給你 NP]] ...? ... [VP 要 [VP 給你 [VP ...]]] ...? 打... ...???? Comparing sentence-comprehension problems across languages shows that they are truly related to parsing the syntax, not due to meaning difficulties (examples from Wu 1989): "The clothes the dog the professor Johnson knows keeps wears are pretty." 「張三認識的教授養的狗穿的衣服很好看。」 2.2 What kind of information influences parsing decisions? 2.2.1 Prosody: The following sequence of words is parsed quite differently in the following sentences, and these different parses have prosodic effects that you can hear: ...狗不買馬... (1) 狗不買，馬他也不買。 (2) 狗不買馬，狗什麼都不買。 There is not much research on such prosodic cues, since most sentence experiments use visual stimuli! 2.2.2 Lexical semantics: Plausibility: 狗不賣... Unlikely to end with 馬, since we know that dogs cannot sell things. Thus the knowledge of word meanings can tell us that here, 狗 is probably a patient (object), not an agent (subject). Argument preferences: Some verbs prefer to be intransitive, and some prefer to be transitive. This lexical information can also help make a good guess about the correct syntactic parse. For example, in the word sequence "...the boy sneezed the doctor...", "the doctor" is unlikely to be the object, but is probably the subject of a following clause (e.g. "After the boy sneezed, the doctor gave him a pill.") [see Mitchell 1987] Even adjectives can have different argument preferences, with dramatic effects on the syntactic structure (example from Chomsky somewhere): (1) John is eager to please (2) John is easy to please In (1), John is the subject of "please" (he wants to please other people) In (2), John is the object of "please" (other people easily please him) 2.2.3 Sentential semantics: The context of a sentence can help determine how it should be parsed. [Tyler & Marslen-Wilson 1977] e.g. the phrase "landing planes" is ambiguous: (a) [NP [Adj landing] [N planes]] (i.e. planes that are landing) "If you walk too near the runway, landing planes can be dangerous." (b) [VP [V landing] [N planes]] (i.e. making planes land) "If you are a beginning pilot, landing planes can be dangerous." 2.3 Question: which way do people parse? (A) They wait until they know for sure before they decide on the syntactic structure for the whole sentence. (B) They make an immediate guess about where the current word goes, even if this turns out to be wrong. Answer: (B). This is called the IMMEDIACY PRINCIPLE. 2.3.1 The evidence comes from GARDEN PATH SENTENCES: grammatical sentences that are difficult to parse because at first they seem to have one structure, but later they turn out to have another (so they make you wander down a "garden path" and get lost!). Classic example: "The horse raced past the barn fell." ( = "The horse that was raced past the barn fell.") Other examples in English are seen in (2) and (3) below: (1) The glass dropped from my hand. [easy] (2) The glass dropped by the boy broke. [harder] (3) The glass dropped from my hand broke. [hardest] In each sentence, you expect that "dropped" is the main verb of the sentence, not part of a relative clause. So when you get to the word "broke" in (2,3), you are confused: at first, you parse the sentences incorrectly! This proves that you made an immediate parse before getting all the information. Chinese examples: [from Lee (to appear)] (1) 在首都機場已經關閉了。 (2) 我們多買一些蔬菜吃腸胃才會覺得舒服。 Sometimes people may not even notice that they have garden-pathed, but the effect is measurable (see below). 2.3.2 How do people make these immediate parsing decisions? (Frazier 1987) (1) Late Closure strategy: delay closing the current syntactic constituent as long as possible. That is, keep adding new words to the current constituent, rather than finding another place for them. "The spy saw the cop with binoculars" tends to be parsed as [S the spy saw [NP the cop with binoculars]] since then the NP obeys Late Closure; the alternative parse [S the spy saw [NP the cop] with binoculars] would violate Late Closure. (2) Minimal Attachment strategy: When fitting a new word into a syntactic tree, use the minimum number of syntactic nodes. This principle is necessary to handle cases that aren't explained by Late Closure. "The girl heard music is good for the soul" tends to be parsed as (a) [S the girl [VP heard music ] ... though it's really (b) [S the girl [VP heard [S music is good for the soul]]] Note that with either parse, "music" is still within the VP, so neither requires an early closure of the VP. The reason why (a) is preferred is that it contains fewer nodes (i.e. you don't add that extra S). 2.3.3 How are parsing strategies studied experimentally? Give sentences that should cause subjects to garden path ONLY IF subjects are using some hypothesized strategy, e.g. "After the customer ate the chicken crossed the road." preferred (but incorrect) initial parse: [S after the customer [VP ate the chicken] ...] Experimental measures of garden-pathing include: total reading time (a more "off-line" measure) other measures are more "on-line", e.g. number of regressions (using eye-tracking device) duration of eye fixations before vs. after disambiguating region (i.e. the word or phrase that makes the correct parse clear) self-paced reading task: sentence displayed on a computer screen one part at a time; subjects control speed of display word-by-word or phrase-by-phrase; compare duration of reading of word before vs. after disambiguating region 3. Is syntactic processing modular? If so, it must be automatic (it is); domain-specific (probably); neurologically distinct (we'll look at this later); and informationally encapsulated (...????). It's this fourth property that is the most controversial: Do people process syntax totally independently of semantics (whether lexical or sentential)? Easy answer: Of course not! Eventually, listeners and readers do use lexical and semantic information to help make parsing decisions. More sophisticated answer: Maybe syntactic parsing IS independent of lexical and semantic information when people make their initial parsing decisions ("on-line"). The other kinds of information may become available only later in processing ("off-line"). This second answer is the one given by modularists. In this view, syntactic parsing is both automatic, and informationally encapsulated, i.e. independent of lexical semantics & sentential semantics. 3.1 Does lexical semantics play a role in the initial parse? Rayner, Carlson and Frazier (1983) say NO: Issue: Are garden-path sentences hard because they violate syntactic parsing strategies, or because they violate semantic plausibility? "The florist sent the flowers was very pleased." (This has structure of "The horse raced past the barn fell.") But maybe people actually garden-path here because of lexical semantic knowledge (i.e. florists should SEND flowers to other people; it's less common for people to send flowers TO florists) Materials: Four kinds of sentences to test whether it's syntax or semantics that is more important in garden-pathing: SEMANTICS IMPLAUSIBLE PLAUSIBLE SYNTAX REDUCED a. The florist sent b. The performer the flowers was sent the flowers very pleased was very pleased UNREDUCED/ ACTIVE c. The florist who d. The performer was sent the flowers sent the flowers and was very pleased was very pleased Task: reading sentences on a computer screen, with eye movements tracked by an eye-tracking machine. Results: there was no significant effect of semantics (i.e. a vs b and c vs d), but there was a significant effect of syntax (i.e. a vs c and b vs d). By the way, I replicated this same result, using the same materials and American students in a psycholinguistics class (but measuring the garden-path effect with a self-paced reading task rather than an eye-tracker). It even worked with two non-native speakers in the class (one from Germany, one from Japan). Conclusions: syntax may play a more important role than semantics in causing garden-path effects. 3.2 Can sentential semantics affect the initial parse? Ferreira and Clifton (1986) say NO. Issue: can parsing strategies like Minimal Attachment be "turned off" if the semantic context points to the correct parse? Task: reading sentences; looking times were measured for the disambiguating region. Materials: "The editor played the tape agreed it was a big story." This violates Minimal Attachment, and so normally causes a garden-path, with longer looking times at the word "agreed". Context that biases the correct parse: "John worked as a reporter for a big city newspaper. He sensed that a major story was brewing over the city hall scandal, and he obtained some evidence that he believed pretty much established the mayor's guilt. He ran a tape for one of his editors, and he showed some photos to the other." (Note that this context makes it clear that in the key sentence, the phrase "played the tape" is a relative clause, and "the editor" is not the subject of "played". Think about this carefully if you don't see why this is!) Results: Subjects still garden-path. Thus context in this experiment did NOT turn off the Minimal Attachment strategy. 3.3 Not everyone agrees that syntax is processed so autonomously, though (e.g. part 5 below). 4. Processing discontinuous dependencies The FILLER-GAP effect: When people process sentences like the following, they must temporarily store information about the NP and then reactivate it when they reach the trace (t): [Which cat]i did the kid grab ti? [那個老師]i我很不喜歡ti This can be studied by using the cross-modal priming technique, similar to the one used to study semantic ambiguity (as in the "bug" experiment that we've discussed before). The question is: do you find priming of words semantically related to the "filler" even when you get to the "gap"? 4.1 Example: Nicol (1993) Issue: Take a sentence like the following: "That is the actress(i) that Susan admired t(i) very much." Previous research had shown that using a cross-modal priming technique, you can find that the semantics of "actress" are strong right after the word "actress": the word STAGE will be primed at that point. However, the semantics then drops off, so STAGE is not primed after "Susan". But if you test right after "admired", you find that STAGE is primed again! Two opposing explanations: The silent trace itself is linked to the semantics of the NP that it's coindexed with (i.e. "actress") The semantics is actually linked with the verb "admired", since the full understanding of this verb requires linking all of its arguments: subject AND object. To distinguish these two hypotheses, you must use an example of a discontinous dependency where there is no trace, such as the following ("is" goes with "actress", not "critics"): "The ACTRESS who really impressed the critics IS a failure with the audience." However, since there is no movement in this sentence, there is no trace. Using a cross-modal technique, STAGE was primed next to "is", even though there is no trace there. Conclusion: Reactivation can be caused by the verb, even if no trace is present. 4.2 Love and Swinney (1996), and related unpublished work: Issue: Can we show that syntax is independent of lexical semantics in the filler-gap effect? e.g. "bank" can mean both 銀行 (primary meaning) and 岸 (secondary meaning) HEAR: "Jeff had read about the problems with savings and loan institutions, so he went to [the bank]i, which his family always uses ti, to ask about the safety provided to CD investments." SEE: ... MONEY ... [target for primary meaning] ... RIVER ... [target for secondary meaning] Three probe points (points during the sentence where some subjects saw the target words MONEY or RIVER): ...the bank, A which his family always B uses C, ... Results: (* represents significant priming compared to a control) MONEY RIVER A: * * B: C: * Conclusions: A result shows that sentential semantics does not affect the access of lexical semantics (just as in "bug" experiment) C result shows that: when the NP semantics is activated at the gap only the contextually relevant meaning remains B result shows that activation at C is truly reactivation, not merely continuing activation from A 5. Interactionist models of sentence comprehension So far we have assumed that syntax and semantics are processed mostly separately (with perhaps some interaction between them), but there are also researchers who disagree. Competition Model (Bates and MacWhinney 1989): An extreme interactionist model of sentence processing. Bates et al. even believe that syntax and the lexicon are exactly the same thing! For a comprehensive summary of Bates's views, see Bates and Goodman (1997). Focus in this model is on the many different cues used by listeners to comprehend a sentence. Cues may be anything: semantics, lexical items, word order. These cues compete to different degrees, depending on their strength in a particular language or a particular sentence. E.g. Li, Bates and MacWhinney (1993): Issue: How do Chinese listeners decide which noun is the agent in a simple two-noun sentence? Task: Participants hear sentences and must explicitly decide which noun is the agent. Materials: Simple sentences that varied factors such as animacy of the two nouns, word order (NNV, NVN, VNN), and the use of 把 or 被 . For example, here are two sentence types (they warned subjects that sentences would vary a lot in how "good" they sounded): [animate-noun] [verb] [把] [inanimate-noun] 猴子吃掉把香蕉。 [verb] [animate-noun] [被] [inanimate-noun] 打破小狗被蘋果。 Results: 被 was the strongest cue: the noun following it was never interpreted as the patient. Animacy was a more important cue than word order, and word order was a more important cue than 把 . Conclusions: many factors interact and compete all at the same time during sentence comprehension My main objection: WHEN do they interact? The methodology used does not allow us to distinguish between immediate, on-line interaction (theoretically very interesting) and later, off-line interactions (theoretically much less interesting). REFERENCES Bates, E., & Goodman, J. C. (1997). On the inseperability of grammar and the lexicon: evidence from acquisition, aphasia and real-time processing. Language and Cognitive Process, 12, 5-7-584. Bates, E., & MacWhinney, B. (1989). Functionalism and the competition model. In B. MacWhinney and E. Bates (Eds.) The Crosslinguistic Study of Sentence Processing (pp. 3-76). Cambridge, UK: Cambridge University Press. Ferreira, F., & Clifton, C. (1986). The independence of syntactic processing. Journal of Memory and Language, 25, 348-368. Fillenbaum, S. (1966). Memory for gist: some relevant variables. Language and Speech, 9, 217-227. Frazier, L. (1987). Sentence processing: a tutorial review. In M. Coltheart (Ed.) Attention and Performance: Vol. XII: The Psychology of Reading (pp. 559-586). Erlbaum. Hanson, V. L., & Bellugi, U. (1982). On the role of sign order and morphological structure in memory for American Sign Language sentences. Journal of Verbal Learning & Verbal Behavior, 21, 621-633. Keenan, J. M., MacWhinney, B., & Mayhew, D. (1977). Pragmatics in memory: a study of natural conversation. Journal of Verbal Learning and Verbal Behavior, 16, 549-560. Lee, T. H.-T. (2001). Garden path sentences in Chinese. In M. Nakayama (Ed.) Sentence Processing in East Asian Languages. Stanford: CSLI. Li, P., Bates, E., & MacWhinney, B. (1993). Processing a language without inflections: a reaction time study of sentence interpretation in Chinese. Journal of Memory and Language, 32, 169-192. Love, T., & Swinney, D. (1996). Coreference processing and levels of analysis in object-relative constructions: demonstration of antecedent reactivation with the cross-modal priming paradigm. Journal of Psycholinguistic Research, 25, 5-24. Mitchell, D. C. (1987). Lexical guidance in human parsing: locus and processing characteristics. In M. Coltheart (Ed.) Attention and Performance: Vol. XII: The Psychology of Reading (pp. 601-618). Erlbaum. Nicol, J. (1993). Reconsidering reactivation. In G. Altmann & R. Shillcock (Eds.) Cognitive models of speech processing: The second Sperlonga meeting (pp. 321-347). Erlbaum. Rayner, K., Carlson, M., & Frazier, L. (1983). The interaction of syntax and semantics during sentence processing: eye movements in the analysis of semantically biased sentences. Journal of Verbal Learning and Verbal Behavior, 22, 358-374. Sachs, J. S. (1967). Recognition memory for syntactic and semantic aspects of connected discourse. Perception and Psychophysics, 2, 437-442. Sachs, J. S. (1974). Memory in reading and listening to discourse. Memory and Cognition, 2, 95-100. Tyler, L. K., & Marslen-Wilson, W. D. (1977). The on-line effects of semantic context on syntactic processing. Journal of Verbal Learning and Verbal Behavior, 16, 683-692. Wanner, E. (1974). On remembering, forgetting and understanding sentences. Mouton. Wu, A. (1989). "Jump in Difficulty" and the "Two-Tree Hypothesis" in Parsing. UCLA MA thesis.