Tripsnode and Rule have the same structure:
- positionals: a list of elements
- type_word: the pair (type, word) as in
(:* ONT::PLANT W::GRASS) - kvpairs: the dictionary of {role: element}
If the (type word) pattern appears in the kvpairs, e.g. :role (:* type word),
currently it is recorded as {role: "type word"} in the TripsNode.
The design concerning type-word pairs will possibly need modification.
The matching function referenced Champin, Pierre-Antoine, and Christine Solnon. The basic idea of matching this kind of labeled (both vertices and edges), directed graph is:
- Define a scoring function that evaluates similarity between two graphs given a mapping.
- Use a greedy approach to find the best pair to add to the existing mapping, one at a time until all the nodes are mapped or the score cannot be improved (best pair means adding it to the mapping improves the score by the largest margin).
The goodness of fit between a Trips parse P and a rule set R' with respect to a mapping is evaluated as:
sum (intersection(p, r))/ sum (cardinality(r)) for each mapped pair (p, r),
where the intersection is the number of matches among positionals and kvpairs for each,
and the cardinality is kvpairs.length + positionals.length.
If we start with string-formatted logical forms, e.g.
((SPEECH-ACT ?x SA_TELL :CONTENT ?!theme)
(F ?!theme ?what ?type :AGENT ?f)
we need to load it into a set of rules using load_list_set().
We are also able to pass a json object into that same function.
Do the same to a TRIPS parse, e.g.
((ONT::SPEECHACT ONT::V1617693 SA_TELL :CONTENT ONT::V1617530 :LEX ONT::BUY)
(ONT::F ONT::V1617530 (:* ONT::PURCHASE W::BUY) :AGENT ONT::V1617694 :AFFECTED ONT::V1617576 :BENEFICIARY ONT::V1617534 :TENSE ONT::PRES :LEX ONT::BUY)
(ONT::IMPRO ONT::V1617694 REFERENTIAL-SEM :PROFORM ONT::SUBJ)
(ONT::PRO ONT::V1617534 (:* ONT::PERSON W::ME) :PROFORM ONT::ME :LEX ONT::ME :COREF ONT::USER)
(ONT::A ONT::V1617576 (:* ONT::COMPUTER-TYPE W::LAPTOP) :LEX ONT::LAPTOP)
)
Then passing the two loaded set to score() will give the score of the parse against the rule set.
We are also able to load in a list of rules from a file and test a given parse against that list of rules to find the best match.
The text file storing the list of rules must be formatted as follows:
---
# Description of pattern 1
/ Comment 1
/ Comment 2
/ ...
Pattern in logical form using Lisp-like syntax with variables in the format "?x" where x can be whatever you want.
---
# Description of pattern 2
...
We can parse a file in this format using the parse_rule_set() function which returns a list of tuples that hold (parsed rule, description of rule).
Then we can pass the output from the previous function and a parsed rule into the grade_rules() function which will return the description of the highest matching rule as well as the score of that match (1.0 for an exact match).
Champin, Pierre-Antoine, and Christine Solnon. "Measuring the similarity of labeled graphs." International Conference on Case-Based Reasoning. Springer, Berlin, Heidelberg, 2003.