Structure Building in Human Sentence Processing: Argument Structures and General Routines

1
Structure Building in Human Sentence
ProcessingArgument Structures and General
Routines

• Fernanda Ferreira
• Cognitive Science Program and
• Department of Psychology
• Michigan State University

2
Big Questions in Human Sentence Processing

• Projection of syntactic structure from arg.str
• How do arg.strs turn into phrase-markers?
• Do arg.strs verify structures created through
  independent tree-building routines?
• Are the answers to these questions different
  for
• Formal linguistic theory
• Sentence comprehension
• Sentence production

3
Big Questions in Human Sentence Processing

• Multiple argument structures for a lexical item
• If a lexical item (e.g., verb) has more than one
  argument structure, how are those represented
  (see Burning Question 3)?
• What are parsers preferences for handling
  multiple arg.strs? Does it weight more heavily
  the ones that are frequent? Simple? Maximal?
• How are arg.strs retrieved from memory? All
  available at the same time, or serially retrieved
  and evaluated?

4
Big Questions in Human Sentence Processing

• Null elements in arg.str, especially null objects
• How are they represented?
• How are the intran forms of English verbs like
  eat and bathe different?
• And how different from languages such as Chinese
  which have real null objects and which make
  liberal use of them?
• Does the parser wait for lexical constituents,
  or does it anticipate a null object (see 2b)?

5
Tree construction rules and /or argument
structures

• General routines build trees. Arg.strs
  separately evaluate those trees (e.g.,
  Garden-Path model)
• Arg.strs associated with tree bits (e.g., MPS)
• Arg.strs encoded in form of trees (TAG)

6
Psycholinguistic questions

• Is there evidence for general routines?
• Minimal structure
• Recency / late closure
• Evidence for independent arg.str effects?
• How are adjuncts processed?

7
Evidence for general routines? Ferreira
Henderson, 1990

• Ed realized the story was not true
• Ed recalled the story was not true
• Ed realized that the story was not true
• Ed recalled that the story was not true

8
How are adjuncts processed?

• Parser tries to integrate new material as
  argument rather than as adjunct
• Adjuncts done by rule rather than lexically
  (Boland)

9
Multiple argument structures for a lexical item

• If a lexical item (e.g., verb) has more than one
  argument structure, how represented (see Burning
  Question 3)?
• What are parsers preferences for handling
  multiple arg.strs? Does it weight more heavily
  the ones that are frequent? Simple? Maximal? Does
  it prefer lexical arguments over null?
• How are arg.strs retrieved from memory? All
  available at the same time, or serially retrieved
  and evaluated?

10
What type of arg.str for a lex.item does parser
prefer?

• Likes the one that is most frequent
• Likes one that is maximal (Pritchetts
  Theta-attachment)

11
Evidence that parser prefers maximal grid
(Schmitt, Munn, Ferreira, in progress)

• Even when transitive form less frequent, parser
  gets garden-pathed in structures such as While
  the man studied the papers blew off the desk.

N26
12
Even when verb has special argument structure
(Schmitt, Munn, Ferreira, in progress)

• Reflexive verbs Heather dressed, more often
  intransitive, and has an object in intran form
  (Heather dressed herself)
• But same amount of garden-pathing
• While Heather dressed the child sat on the couch

N26
13
Op.Tran Verbs and RAT Verbs Side-by-Side
So parser seems to prefer lexical arguments over
null ones
14
Retrieval of arg.strs from memory (Fodor
Ferreira, in progress)

• At least in reanalysis, retrieval is serial and
  time-consuming
• Evidence from grammaticality judgments made at
  various time points in sentence

15
Implications

• Structural analysis goes first
• Lexical look-up evaluates results of structural
  analysis and reanalysis
• Accessing the lexicon and searching for a
  particular arg.str is slow process

16
Null elements in arg.str, especially null objects

• In English, if pit lexical NP against null, Ss
  prefer lexical (saw with RAT verbs)
• Not true in Chinese
• But null object dramatically influences
  interpretations in English

17
Null Objects in English Interpretations
18
Null Objects in Chinese (Xiang, Munn, Schmitt,
Ferreira, in progress)

• Huangrong took out a little medicine bottle and
  Kuojing touched her wound because
• Huangrong took out a little medicine bottle and
  Kuojing touched because

Filled Gap Effect
19
What about language production?

• Argument structure much less studied
• Options used to defer heavy constituents
• Primed constituents tend to occur early
• Bock syntactic priming cares about number of
  arguments but not thematic roles (locative primes
  dative)
• But Pickering priming greater for same than for
  different verbs. So prod head-driven.
• Ferreira (1994) passives occur more frequently
  with th-exp verbs (frightened) than with normal
  verbs

20
General Conclusions

• Tree building routines exist
• In comprehension, arg.strs check trees built
  independently (unlike production)
• Parser likes trees that conform to frequent,
  maximal argument structures
• In English, prefer lexical NPs over null (but
  null objects affect interpretations)
• In Chinese, perhaps prefer null objects over
  lexical NPs
• Different type of null object, different
  frequencies

21
Burning Questions in Human Sentence Processing

• Are argument structures structural or semantic?
• Is thematic structure different from argument
  structure? What is the right vocabulary for
  argument structures
• What is the inventory of thematic roles? How do
  they map onto syntactic constituents?
• Is there a thematic hierarchy, and if so, what is
  its basis?

22
Burning Questions in Human Sentence Processing

• Arguments versus Adjuncts
• What is the basis of this distinction, both
  conceptually and syntactically?
• How are adjuncts made distinct in phrase-markers?
  How are they syntactically represented?
• How does the parser know (or when does it
  realize) that a constituent is an adjunct ?

23
Burning Questions in Human Sentence Processing

• Redundancy and representation of argument
  structures
• How are alternative argument structures for a
  single lexical item represented? An exhaustive
  list? Underspecification?
• If a verb allows sentential complements both with
  and without a complementizer, are these
  alternatives different argument structures?
• Is the answer to the above question different
  depending on whether we talk about linguistic
  representations, comprehension, or production?

24
Maybe Neither Big nor Burning

• How do changes in language use over long periods
  of time lead to changes in argument structure
  representations?
• How does this relate to the use of corpus
  statistics to get at issues of argument structure
  representations and their use in processing?