Psychology of Language

Transcription

1 PSYCH 155/LIG 155 UCI COGITIVE SCIECES syn lab Psychology of Language Prof. Jon Sprouse Lecture 18: Parsing 1

2 Grammar PP P A grammar is a set of equations/rules that generates all (and only) the sentences of a given language. V S V V PP V PP S V P Sentences have hierarchical structure: words combine to form phrases, which combine to form larger phrases, and so on. the boy ate the cookies after the party 2

3 The problem of sequential input The problem: Words are delivered in a specific temporal order, with no information about the hierarchical structure S This means that your brain needs to build the structure from the words alone. PP P V This string of words forms a sentence 3

4 Grammar can t do it alone A grammar is just a static list of rules. It doesn t specify the order that the rules should be a applied. PP P V V V PP V This is similar to the rules of arithmetic: Addition: x + y Subtraction: x - y Multiplication: x * y S ivision: Exponentiation: x / y x y Knowing the rules of arithmetic isn t enough to solve a complex problem, you need to know the order of operations : Please Excuse My ear Aunt Sally 4

5 What we need is a Parser PP P A grammar is a set of equations/rules that generates all (and only) the sentences of a given language. V V V PP V S Sentences have hierarchical structure: words combine to form phrases, which combine to form larger phrases, and so on. A parser is a cognitive system for building hierarchical structure from the sequential input of words based on the rules of the grammar. Just like there can be any number of orders of operations for arithmetic rules, there can be any number of parsers (that is, any number of ways to build hierarchical structure from sequential input). It is an empirical question which method of parsing the brain actually uses. 5

6 Extreme hypothesis 1: Hypothesis-riven Parsing Also known as top-down parsing, this is when the parser predicts a structure before the word is encountered, and then checks to see if it matches S PP P V This string of words forms a sentence 6

7 The problem with H-driven parsing We can already see that H-driven parsing may require several revisions if the hypothesis is incorrect. S PP P V This string of words forms a sentence 7

8 Extreme hypothesis 2: ata-riven Parsing Also known as bottom-up parsing, this is when the parser uses information about the word to generate the structure S PP P V This string of words forms a sentence 8

9 The problem with -driven parsing We can already see that -driven parsing will stall out if there is more than one option available (that is, given two choices, it must have a hypothesis about which is correct, but by definition, -driven parsing has no hypotheses!) S PP P V This string of words forms a sentence 9

10 Left-Corner Parsing What we need is a middle ground: a parser that uses some information from the words to choose a hypothesis (prediction) about what the structure is. Question: What information does the parser use to make its predictions? A left-corner parser combines both H-driven and -driven parsing by using syntactic categories and phrase structure rules to parse the sentence. Phrase Structure Rules S S -> -> V -> The first item after the arrow is the left corner of the rule. V -> hug -> boy -> girl V -> the The girl_ hugs the boy 10

11 Predictions of a left-corner parser If the human parser is a left-corner parser, then we would predict that the earliest stage of sentence processing would be driven by the syntactic category of the word, and the phrase structure rules that the word can appear in. We can use electroencephalography to test this prediction because it provides very good temporal resolution. egative Voltage egative Component Positive Voltage Positive Component 11

12 Predictions of a left-corner parser If the human parser is a left-corner parser, then we would predict that the earliest stage of sentence processing would be driven by the syntactic category of the word, and the phrase structure rules that the word can appear in. We can test this prediction by violating the phrase structure rules of a sentence, and looking to see how early the response to the violation appears: P The scientist criticized Max s proof of the theorem PP This sentence is grammatical, and serves as a control to show us what the normal response to of looks like P * The scientist criticized Max s of proof the theorem * P This sentence is ungrammatical because there is no phrase structure rule that allows a preposition to follow a determiner. 12

13 Predictions of a left-corner parser If the human parser is a left-corner parser, then we would predict that the earliest stage of sentence processing would be driven by the syntactic category of the word, and the phrase structure rules that the word can appear in. The scientist criticized Max s proof of the theorem * The scientist criticized Max s of proof the theorem The ungrammatical sentence causes a negative peak in amplitude 150ms - 250ms after the onset of the preposition. We call this the Early Left Anterior egativity because it is early, it appears over left anterior scalp locations, and it is negative. 13

14 How early is the ELA? The ELA occurs in response to phrase structure violations 150ms - 250ms after the onset of the critical word. But is this early enough satisfy the prediction of a left-corner parser? Is it the earliest stage of sentence processing? Well, to figure out if this is the earliest stage, we need to look at other stages of sentence processing to see if any are (or could be) earlier. Recall that shadowing experiments suggest that lexical access occurs around 150ms - 200ms after word onset. Speaker 1 Speaker 2: the shadower word1: 375ms 200ms lag word1: 375ms Because sentence processing requires words, the time course of lexical access provides a lower limit the speed of sentence processing. The ELA ( ms) occurs in approximately the same window as lexical access ( ms), so this is very early in sentence processing! 14

15 How early is the ELA? The ELA occurs in response to phrase structure violations 150ms - 250ms after the onset of the critical word. We can also look for other processes to see if any occur before the ELA. For example, we know that the brain must construct a meaning for the sentence. We can break that meaning to look for a time course of when meaning is constructed. He spread the warm bread with butter. * He spread the warm bread with socks. The incongruous sentence causes a negative peak in amplitude 3000ms - 500ms after the onset of the incongruous word. We call this the 400 because it is negative, and peaks around 400ms after the word. 15

16 How early is the ELA? Lexical Access occurs around 150ms - 200ms after the onset of the word. The ELA occurs in response to phrase structure violations 150ms - 250ms after the onset of the critical word. The 400 occurs in response to semantic violations 300ms - 500ms after the onset of the critical word. These facts suggest that the ELA is indexing an extremely early stage of sentence processing -- as predicted by a left-corner parsing architecture! 16

17 Syntax precedes Semantics A final prediction of a left-corner parser is that syntactic processing (building hierarchical structure according to phrase structure rules) will precede semantic processing (deriving a meaning from the sentence). The relative ordering of the ELA and the 400 suggest that this might be true: The ELA occurs in response to phrase structure violations 150ms - 250ms after the onset of the critical word. The 400 occurs in response to semantic violations 300ms - 500ms after the onset of the critical word. However, we can be more sophisticated than this. If syntactic structure building precedes semantic processing, then that means that semantic processing is dependent upon syntactic structure building. In other words, if syntactic structure building fails, then semantic processing is impossible. 17

18 Syntax precedes Semantics We can test this prediction by violating both syntax and semantic simultaneously. Syntax violation: er Strauch wurde trotz verpflantz von einem Gärtner, den wenige empfahlen The bush was despite replanted by a gardner, whom few recommended Semantics violation: as Buch wurde verpflantz von einem Verleger, den wenige empfahlen The book was replanted by a publisher, whom few recommended ouble violation: as Buch wurde trotz verpflantz von einem Verleger, den wenige empfahlen The book was despite replanted by a publisher, whom few recommended. Friederici et al

19 Syntax precedes Semantics If semantics proceeds independently, then we will see both an ELA and 400. If semantics is dependent on successful syntactic processing, then we should only see an ELA. The 400 will not show up because the system will not see the semantic violation (because semantic processing will not have taken place). Syntax violation: Semantics violation: ouble violation: Friederici et al