Skip to content

ayoungprogrammer/Lango

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

17 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Lango

Gitter

Lango is a natural language processing library for working with the building blocks of language. It includes tools for:

Need help? Ask me for help on Gitter

Installation

Install package with pip

pip install lango

Download Stanford CoreNLP

Make sure you have Java installed for the Stanford CoreNLP to work.

Download Stanford CoreNLP

Extract to any folder

Run the Stanford CoreNLP server

Run the following command in the folder where you extracted Stanford CoreNLP

java -mx4g -cp "*" edu.stanford.nlp.pipeline.StanfordCoreNLPServer

Docs

Matching

Matching is done by comparing a set rules and matching it with a parse tree. You can see parse trees for sentences from examples/parser_input.py.

The set of rules is recursive and can match multiple parts of the parse tree.

Rules can be broken down into smaller parts:

  • Tag
  • Token
  • Token Tree
  • Rules

Tag

A tag is a POS (part of speech) tag to match. A list of POS tags used by the Stanford Parser can be found here.

Format:
tag = string

Example:
'NP'
'VP'
'PP'

Token

A token is a string comprising of a tag and modifiers/labels for matching. We specify a match_label to match the tag to. We can specify opts for extracting the string from a tree. We can specify eq for matching the tree to a string.

Example string:
The red car

opts:
-o Get object by removing "a", "the", etc. (Ex. red car)
-r Get raw string (Ex. The red car)
Format: (only tag is required)
token = tag:match_label-opts=eq

Example: 
'VP'
'NP:subject-o'
'NP:np'
'VP=run'
'VP:action=run'

Token Tree

A token tree is a recursive tree of tokens. The tree matches the structure of a parse tree.

Format:
token_tree = ( token token_tree token_tree ... )

Examples: 
'( NP ( DT ) ( NP:subject-o ) )'
'( NP )'
'( PP ( TO=to ) ( NP:object-o ) )'

Rules

Rules are a dictionary of token trees to dictionaries of matching labels to a nested set of rules.

Format:
rules = {token_tree: {match_label: rules}}

Example: 
{
    '( S ( NP:np ) ( VP ( VBD:action-o ) ( PP:pp ) ) )': {
        'np': {
            '( NP:subject-o )': {}
        },
        'pp': {
            '( PP ( TO=to ) ( NP:to_object-o ) )': {},
            '( PP ( IN=from ) ( NP:from_object-o ) )': {},
        }
    },
}

When matching a rule to a parse tree, the token tree is first matched. Then, all matching tags are matched to nested rules corresponding to their matching label.

All nested match labels must have a subrule match or the rules will not match.

The first rule to match is returned so the order of match is based on key ordering (use OrderedDict if order matters). Once a rule is matched, it calls the callback function with the context as arguments.

Example

Suppose we have the sentence "Sam ran to his house" and we wanted to match the subject ("Sam"), the object ("his house") and the action ("ran").

Sample parse tree for "Sam ran to his house" from the Stanford Parser.

(S
  (NP 
    (NNP Sam)
    )
  (VP
    (VBD ran)
      (PP 
        (TO to)
        (NP
          (PRP$ his)
          (NN house)
          )
        )
    )
  )

Simplified image of tree:

tree

Matching:
Parse Tree: 
(S (NP (NNP Sam) ) (VP (VBD ran) (PP (TO to) (NP (PRP$ his) (NN house))))

Matched token tree: '( S ( NP:np ) ( VP ( VBD:action-o ) ( PP:pp ) ) )'
Matched context: 
  np: (NP (NNP Sam))
  action-o: 'ran'
  pp: (PP (TO to) (NP (PRP$ his) (NN house)))

Rule for '( S ( NP:np ) ( VP ( VBD:action-o ) ( PP:pp ) ) )':

tree

Matching 'NP' matches the whole NP tree and converts to a word:

Matched token tree for np: '( NP:subject-o )'
Matched context:
  subject-o: 'Sam'

Matching 'PP' requires matching the nested rules:

Match token tree for pp: '( PP ( TO=to ) ( NP:to_object-o ) )'
Match context:
  object-o: 'his house'

Match token tree for pp: '( PP ( IN=from ) ( NP:from_object-o ) )'
No match found

PP of the sample sentence:

tree

Nested PP rules:

tree tree

Only the first rule matches for 'PP'.

Now that we have a match for all nested rules, we can return the context:

Returned context:
  action: 'ran'
  subject: 'sam'
  to_object: 'his house'

Full code:

from lango.parser import StanfordServerParser
from lango.matcher import match_rules

parser = StanfordServerParser()

rules = {
  '( S ( NP:np ) ( VP ( VBD:action-o ) ( PP:pp ) ) )': {
    'np': {
        '( NP:subject-o )': {}
    },
    'pp': {
        '( PP ( TO=to ) ( NP:to_object-o ) )': {},
        '( PP ( IN=from ) ( NP:from_object-o ) )': {}
    }
  }
}

def fun(subject, action, to_object=None, from_object=None):
    print "%s,%s,%s,%s" % (subject, action, to_object, from_object)

tree = parser.parse('Sam ran to his house')
match_rules(tree, rules, fun)
# output should be: sam, ran, his house, None

tree = parser.parse('Billy walked from his apartment')
match_rules(tree, rules, fun)
# output should be: billy, walked, None, his apartment