SD-Rules Optimization Models and Procedures

• Optimize a set of SD-rules
  Keep the good rules and remove the bad rules from a known set of rules:
  • The original Lexical Tools - 97 SD-Rules, is used as the baseline
  • Duplicated SD-Rules are removed (none is found in these 97-set)
  • Evaluated all parent-child rules and find the optimized set
    Go through each parent-rule to choose parent-rule or child-rules by following steps:
    • All child-rules are temporary removed (keep the parent-rule)
    • Decompose a parent-rule to child-rules, grandchild-rules, etc.
      Only decompose recursively on a child rule if it is a potential good rule:
      • its accuracy rate is higher than root parent-rule
      • its coverage rate is higher than 40% of root parent-rule. 40% is a default number and can be adjusted.
    • Evaluate child-rules for one parent-rule at a time
      Child-rules should be evaluated only if
      • Child-rules have higher accuracy rate than its parent-rule.
        Otherwise, just the ignore child-rule and use its parent-rule because parent-rule will have better accuracy and coverage rate than the child-rule.
      • Child-rules have more than 35% of coverage of its parent-rule.
        All rules have to be a good rule to be in the optimized set. So, child-rules should have good coverage. 35% is a default number and can be adjusted.
      • Compare the system performance between parent-rule and child-rules, and choose the better one:
        • higher system performance
        • more rules if system performance is the same
  • Find the optimized SD-Rule set with best system performance by superposition those better parent-child rules

• Add a new rule to a set of SD-rules

  The same model and procedures as above can be used to evaluate new SD-rules. If a new SD-rule is suggested to add to a (optimized) set of SD-rules, the procedures are as follows:

  • Check if the new rules is a duplicated rule,
    if so, no need to evaluate the new rule.
  • Check if the new rule is a child-rule,
    if so, use the same parent-child evaluation procedures as above.
  • Check if the new rule is a parent-rule
    if so, use the same parent-child evaluation procedures as above.
  • Others:
    Evaluate it by comparing system performance

  The evaluation procedures need to get tagging stats of matching SD-pairs from Lexicon:

  • Get all raw SD-pairs matches the new SD-Rule from Lexicon
  • Tag the raw SD-pairs
  • These data can be derived from the existing set if the new rule is a child-rule.

The following figure shows the SD-Rules model and optimization procedures.