Apple Pie Parser Satoshi Sekine, July 1996

1
Apple Pie ParserSatoshi Sekine, July 1996

• Discussion and Demonstration Session
• Kenneth Wilson
• The College of New Jersey
• Spring 2003

2
Apple Pie Parser

• Developed by Satoshi Sekine at NYU Spring 1995
• Bottom-up probabilistic char parser
• Uses a best-first search algorithm
• Grammar
• Semi-context sensitive grammar
• Two nonterminals
• Extracted from Penn Tree Bank

3
Apple Pie Parser

• Fully automatic acquisition of grammar from a
  tagged corpus
• Parser generates a syntactic tree as its output
• Goal is to make a parse tree as accurate as
  possible for reasonable sentences (newspapers,
  well written documents, etc.)
• Excludes conversation and ill-formed sentences

4
Obtaining the source
Via FTP _at_ cs.nyu.edu
Via WWW http//cs.nyu.edu/cs/projects/proteus/sek
ine http//cs.nyu.edu/cs/projects/proteus/app
5
Directories and files

• After unzipping and untaring

Commands for unzipping and untaring gzip d
APP5.9.tar.gz tar xvf APP5.9.tar
6
How to make

• Changes to be made to src/config.h file
• Change relative pathname to absolute pathname
• define DEFAULT_PARAMFILE /local2/users/wilson15/
  APP5.9/bin/app.prm
• Change memory allocation
• define ALLOCATE_ANODE 600000 / number of Anode
  /
• Changes to be made to bin/app.prm
• Change relative pathnames to absolute pathnames
• DICTIONARY_FILE /local2/users/wilson15/APP5.9/data
  /WS-A-001.dic
• GRAMMAR_FILE /local2/users/wilson15/APP5.9/data/WS
  -A-001.grm
• NICKNAME_FILE /local2/users/wilson15/APP5.9/data/W
  S-A-001.nic

7
How to make
Type make at ./src directory (rm .o first if
necessary)
8
How to run

• Type app at ./bin directory or from any directory
  (after modifying the path in shell)

9
Example of a session

• Example of an APP session

10
Parsing algorithm

• Techniques which enable the parser to handle
  large grammars
• Grammar rules are factored with common prefixes
• Best-first search
• Because the grammar is a probabilistic grammar,
  it is enough to find only one parse tree with
  highest possibility.
• Parser uses integer values for indicating POS or
  a grammatical node. Relationships between the
  integer values and corresponding POS are stored
  in the nickname file.

11
Out of vocabulary

• Treatment of words not found in the vocabulary
• Uses a list of part-of-speech d

12
Fitted parsing

• Treatment of very long sentences.
• Parser cant create a complete tree for some long
  sentences.
• Parser prepares a post process to make a complete
  tree from several partial trees in the chart.

13
List of parameters

• Default parameter file ./bin/app.prm

System Cut off START_SYMBOL Out of
Vocabulary OOV_NUM DICTIONARY_FILE TOO_LONG NP_LAB
EL FLAG_OOV OOV_NNP DICT_SCORETYPE CUT_GSCORE T
okenization OOV Others GRAMMAR_FILE CUT_DPROB TN_
SYMBOL OOV_NUMHYPN TRANS_POS NICKNAME_FILE PRUNE_R
ATE TN_TWO_SYMBOL OOV_CAPITAL INV_TRANS_POS DEBUG
Adjustment TN_TWO1_SYMBOL OOV_LY SUP_WORD PRINT
_STYLE WEIGHT_GRAM SPECIAL_HEAD OOV_Y PRINT_TOKEN
NO_WARNING WEIGHT_DICT DEL_TAIL OOV_ED PRINT_NT
Fitted parsing CAP_MINFREQ DEL_TAIL_SYMBOL OOV_D
FITTED_PARSE ATTACH DEL_TAIL_STRING OOV_S FITTED_
ROOT NO_LETTERCASE TAIL_NODE_CAT OOV_ION FITTED_LA
BEL DOUBLE_QUOTE OOV_ING FITTED_CCOST
14
Parameter setting by file

• Can set parameters by file (parameter file)
• ASCII file
• Each line contains a parameter and its value
  separated by at least one space.
• Default parameter file is app.prm (used when no
  other parameter file is specified.
• Specify parameter files using the p option.

15
Change parameter value

• A few parameters can also be changed during a
  session
• Type param and specify parameter name and value
  on the next line
• PRINT_STYLE
• TOO_LONG
• ATTACH
• NO_LETTERCASE
• DOUBLE_QUOTE
• OOV

16
Grammar file

• ltGRAMMAR RULEgt ltGRAMMAR RULEgt
• ltGRAMMAR RULEgt ltLHS NODEgt ltRHS NODESgt
• ltSCORE INFORMATIONgt
• ltSTRUCTURE INFORMATIONgt
• ltRHS NODESgt ltRHS NODEgt
• ltLHS NODEgt integer
• ltRHS NODEgt integer
• ltSCORE INFORMATIONgt score ltSCOREgt
• ltSCOREgt integer
• ltSTRUCTURE INFORMATIONgt string to be print out
  for the node
• ltnumgt correspond to the string of the
  num-th node in RHS
• ---------- Example--------------------------------
  -------------------------------------------------
• 1 53 54 2 103 205 1 57 55
• score 113
• struct (S lt1gtlt2gtlt3gt (VP lt4gt (SBAR lt5gt
  lt6gt)) lt7gt lt8gt)

17
Dictionary file

• ltDICTIONARY FILEgt ltWORD INFORMATIONgt
• ltWORD INFORMATIONgt ltSTRINGgt ltPOS
  INFORMATIONgt
• ltstringgt string
• ltPOS INFORMATIONgt ltPOSgtltSCOREgt
• ltPOSgt integer
• ltSCOREgt integer
• --------Example-----------------------------------
  ------------------------
• base 6515, 70119, 852, 892
• Base-price 701
• Base-rate 651
• Baseball 7054

18
Nickname file

• ltNICKNAME FILEgt ltNICKNAME INFORMATIONgt
• ltNICKNAME INFORMATIONgt ltPOSgt ltNICKNAMEgt
• ltPOSgt integer
• ltNICKNAMEgt string
• ------------Example-------------------------------
  ------------------
• 1 S
• 2 NP
• 5 S1
• 6 NP0
• 51
• 52
• 53
• 54 -LRB-
• 55 -RRB-

19
List of APP functions

• APP_init_param() Initialize parameter variables
• APP_read_param_file() Read parameter
• APP_init_global() Initialize global variables
• APP_parse() Parsing
• APP_set_param() Set parameter
• APP_current_param() Show current parameter values
• APP_debug_routine () Get into debug routine

20
Interface

• Initialization
• Parameter setting
• Must be done first
• APP_init_param() function sets default parameters
• Set your own parameter values using
  APP_set_param()
• Initialize internal data
• Load dictionary, grammar, and nickname
  information and store them
• APP_init_global()

21
Interface

• Parsing
• Initialized by APP_parse()
• Two arguments
• Sentence
• Address of return value structure
• Returns a 1 if parsing is completed, 0 if
  otherwise

22
Interface

• Parameter setting and look-up
• Can set limited kinds of parameters during the
  session using APP_set_param()
• Lookup parameters using APP_current_param()

23
Interface

• Debug routine
• Use to view
• internal structure
• Information about previous input sentences.
• Dictionary entries
• Part of grammar rules
• Sentence and parse tree data on the heap
• Type resume to quit the debug routine and
  resume parsing session

24
Grammar

• WS-A-001.grm is a grammar extracted from the Penn
  Tree Bank.
• Consists of all occurrences of two non-terminal
  grammar rules (S and NP)
• WS-A-002.grm also extracted from Penn Tree Bank
• contains all the rules whose frequencies are more
  than one
• Smaller grammar than WS-A-001.grm
• Tradeoffs
• WS-A-001.grm gives greater accuracy with slower
  parse time
• WS-A-002.grm gives lower accuracy with faster
  parse time

25
Dictionary

• WS-A-001.dic
• Extracted from Penn Tree Bank
• Supplemented by part-of-speech information from
  COMLEX syntax dictionary

26
Accuracy, time, size

• Statistics based on tests runs using WS-A-001.grm
  and WS-A-001.dic
• Tests performed on a SPARCstation 5 with 160 MB
  of memory and set ANODE_TOP to 5,000,000
• Number of sentences 1989
• Average length of sentences 23.28
• Number of parsed sentences 1788
• Number of fitted parse 203
• Average parsing time for all 18
• Average Parsing time excluding fitted parse 9.76
• Precision 71.04
• Recall 70.33
• Average Crossing 3.03