Chunk Parsing

1
Chunk Parsing
2
Chunk Parsing

• Also called chunking, light parsing, or partial
  parsing.
• Method Assign some additional structure to
  input over tagging
• Used when full parsing not feasible or not
  desirable.
• Because of the expense of full-parsing, often
  treated as a stop-gap solution.

3
Chunk Parsing

• No rich hierarchy, as in parsing.
• Usually one layer above tagging.
• The process
• Tokenize
• Tag
• Chunk

4
Chunk Parsing

• Like tokenizing and tagging in a few respects
• Can skip over material in the input
• Often finite-state (or finite-state like) methods
  are used (applied over tags)
• Often application specific (i.e., the chunks
  tagged have uses for particular applications)

5
Chunk Parsing

• Chief Motivations to find data or to ignore
  data
• Example from Bird and Loper find the argument
  structures for the verb give.
• Can discover significant grammatical structures
  before developing a grammar
• gave NP
• gave up NP in NP
• gave NP up
• gave NP help
• gave NP to NP

6
Chunk Parsing

• Like parsing, except
• It is not exhaustive, and doesnt pretend to be.
• Structures and data can be skipped when not
  convenient or not desired
• Structures of fixed depth produced
• Nested structures typical in parsing
• SNP The cow PP in NP the barn ate
• Not in chunking
• NP The cow in NP the barn ate

7
Chunk Parsing

• Finds contiguous, non-overlapping spans of
  related text, and groups them into chunks.
• Because contiguity is given, finite state methods
  can be adapted to chunking

8
Longest Match

• Abney 1995 discusses longest match heuristic
• One automaton for each phrasal category
• Start automata at position i (where i0
  initially)
• Winner is the automaton with the longest match

9
Longest Match

• He took chunks from the PTB
• NP ? D N
• NP ? D Adj N
• VP ? V
• Encoded each rule as an automaton
• Stored longest matching pattern (the winner)
• If no match for a given word, skipped it (in
  other words, didnt chunk it)
• Results Precision .92, Recall .88

10
An Application

• Data-Driven Linguistics Ontology Development (NSF
  BCE-0411348)
• One focus locate linguistically annotated
  (read tagged) text and extract linguistically
  relevant terms from text
• Attempt to discover meaning of the terms
• Intended to build out content of the ontology
  (GOLD)
• Focus on Interlinear Glossed Text (IGT)

11
An Application

• Interlinear Glossed Text (IGT), some examples
• (1) Afisi a-na-ph-a nsomba
• hyenas SP-PST-kill-ASP fish
• The hyenas killed the fish.'
  (Baker 1988254)

12
An Application

• More examples
• (4) a. yerexa-n p'at'uhan-e bats-ets
  
  child-NOM window-ACC open-AOR.3SG
• The child opened the window. (Megerdoom
  ian ??)

13
An Application
(4) a. yerexa-n p'at'uhan-e bats-ets

child-NOM window-ACC open-AOR.3SG The
child opened the window. (Megerdoomian ??)

• Problem How do we discover the meaning of the
  linguistically salient terms, such as NOM, ACC,
  AOR, 3SG?
• Perhaps we can discover the meanings by examining
  the contexts in which the occur.
• POS can be a context.
• Problem POS tags rarely used in IGT
• How do you assign POS tags to a language you know
  nothing about?
• IGT gives us aligned text for free!!

14
An Application
(4) a. yerexa-n p'at'uhan-e bats-ets

child-NOM window-ACC open-AOR.3SG The
child opened the window. (Megerdoomian ??)
DT NN VBP DT NN

• IGT gives us aligned text for free!!
• POS tag the English translation
• Align with the glosses and language data
• That helps. We now know that NOM and ACC attach
  to nouns, not verbs (nominal inflections)
• And AOR and 3SG attach to verbs (verbal
  inflections)

15
An Application
(4) a. yerexa-n p'at'uhan-e bats-ets

child-NOM window-ACC open-AOR.3SG The
child opened the window. (Megerdoomian ??)
DT NN VBP DT NN

• In the LaPolla example, we know that NOM does not
  attach to nouns, but to verbs. Must be some
  other kind of NOM.

16
An Application
(4) a. yerexa-n p'at'uhan-e bats-ets

child-NOM window-ACC open-AOR.3SG The
child opened the window. (Megerdoomian ??)
DT NN VBP DT NN

• How we tagged
• Globally applied most frequent tags (stupid
  tagger)
• Repaired tags where context dictated a change
  (e.g., TO preceding race VB)
• Technique similar to Brill 1995

17
An Application
(4) a. yerexa-n p'at'uhan-e bats-ets

child-NOM window-ACC open-AOR.3SG The
child opened the window. (Megerdoomian ??)
DT NN VBP DT NN

• But can we get more information about NOM, ACC,
  etc.?
• Can chunking tell us something more about these
  terms?
• Yes!

18
An Application
(4) a. yerexa-n p'at'uhan-e bats-ets

child-NOM window-ACC open-AOR.3SG The
child opened the window. (Megerdoomian ??)
DT NN VBP DT NN

• Chunk phrases, mainly NPs
• Since relationship (in simple sentences) between
  NPs and verbs tells us something about the verbs
  arguments (Bird and Loper 2005)
• We can tap this information to discover more
  about the linguistic tags

19
An Application
(4) a. yerexa-n p'at'uhan-e bats-ets

child-NOM window-ACC open-AOR.3SG The
child opened the window. (Megerdoomian ??)
DT NN VBP DT NN
NP
NP
VP

• Apply Abney 1995s longest match heuristic to get
  as many chunks as possible (especially NP)
• Leverage English canonical SVO (NVN) order to
  identify simple argument structures
• Use these to discover more information about the
  terms
• Thus

20
An Application
(4) a. yerexa-n p'at'uhan-e bats-ets

child-NOM window-ACC open-AOR.3SG The
child opened the window. (Megerdoomian ??)
DT NN VBP DT NN
NP
NP
VP

• We know that
• NOM attaches to subject NPs may be a case
  marker indicating subject
• ACC attaches to object NPs may be a case marker
  indicating object

21
An Application

• What we do next look at co-occurrence relations
  (clustering) of
• Terms with terms
• Host categories with terms
• To determine more information about the terms
• Done by building feature vectors of the various
  linguistic grammatical terms (grams)
  representing their contexts
• And measuring relative distances between these
  vectors (in particular, for terms we know)

22
Linguistic Gram Space