How to separate words in a “sentence” with spaces?

问题:

Background

Looking to automate creating Domains in JasperServer. Domains are a "view" of data for creating ad hoc reports. The names of the columns must be presented to the user in a human readable fashion.

Problem

There are over 2,000 possible pieces of data from which the organization could theoretically want to include on a report. The data are sourced from non-human-friendly names such as:

payperiodmatchcode labordistributioncodedesc dependentrelationship actionendoption actionendoptiondesc addresstype addresstypedesc historytype psaddresstype rolename bankaccountstatus bankaccountstatusdesc bankaccounttype bankaccounttypedesc beneficiaryamount beneficiaryclass beneficiarypercent benefitsubclass beneficiaryclass beneficiaryclassdesc benefitactioncode benefitactioncodedesc benefitagecontrol benefitagecontroldesc ageconrolagelimit ageconrolnoticeperiod

Question

How would you automatically change such names to:

• pay period match code
• labor distribution code desc
• dependent relationship

Ideas

• Use Google's Did you mean engine, however I think it violates their TOS:

  lynx -dump «url» | grep "Did you mean" | awk ...

Languages

Any language is fine, but text parsers such as Perl would probably be well-suited. (The column names are English-only.)

Unnecessary Prefection

The goal is not 100% perfection in breaking words apart; the following outcome is acceptable:

• enrollmenteffectivedate -> Enrollment Effective Date
• enrollmentenddate -> Enroll Men Tend Date
• enrollmentrequirementset -> Enrollment Requirement Set

No matter what, a human will need to double-check the results and correct many. Whittling a set of 2,000 results down to 600 edits would be a dramatic time savings. To fixate on some cases having multiple possibilities (e.g., therapistname) is to miss the point altogether.

回答1:

Sometimes, bruteforcing is acceptable:

#!/usr/bin/perl

use strict; use warnings;
use File::Slurp;

my $dict_file = '/usr/share/dict/words';

my @identifiers = qw(
    payperiodmatchcode labordistributioncodedesc dependentrelationship
    actionendoption actionendoptiondesc addresstype addresstypedesc
    historytype psaddresstype rolename bankaccountstatus
    bankaccountstatusdesc bankaccounttype bankaccounttypedesc
    beneficiaryamount beneficiaryclass beneficiarypercent benefitsubclass
    beneficiaryclass beneficiaryclassdesc benefitactioncode
    benefitactioncodedesc benefitagecontrol benefitagecontroldesc
    ageconrolagelimit ageconrolnoticeperiod
);

my @mydict = qw( desc );

my $pat = join('|',
    map quotemeta,
    sort { length $b <=> length $a || $a cmp $b }
    grep { 2 < length }
    (@mydict, map { chomp; $_ } read_file $dict_file)
);

my $re = qr/$pat/;

for my $identifier ( @identifiers ) {
    my @stack;
    print "$identifier : ";
    while ( $identifier =~ s/($re)\z// ) {
        unshift @stack, $1;
    }
    # mark suspicious cases
    unshift @stack, '*', $identifier if length $identifier;
    print "@stack\n";
}

Output:

payperiodmatchcode : pay period match code
labordistributioncodedesc : labor distribution code desc
dependentrelationship : dependent relationship
actionendoption : action end option
actionendoptiondesc : action end option desc
addresstype : address type
addresstypedesc : address type desc
historytype : history type
psaddresstype : * ps address type
rolename : role name
bankaccountstatus : bank account status
bankaccountstatusdesc : bank account status desc
bankaccounttype : bank account type
bankaccounttypedesc : bank account type desc
beneficiaryamount : beneficiary amount
beneficiaryclass : beneficiary class
beneficiarypercent : beneficiary percent
benefitsubclass : benefit subclass
beneficiaryclass : beneficiary class
beneficiaryclassdesc : beneficiary class desc
benefitactioncode : benefit action code
benefitactioncodedesc : benefit action code desc
benefitagecontrol : benefit age control
benefitagecontroldesc : benefit age control desc
ageconrolagelimit : * ageconrol age limit
ageconrolnoticeperiod : * ageconrol notice period

See also A Spellchecker Used to Be a Major Feat of Software Engineering.

回答2:

I reduced your list to 32 atomic terms that I was concerned about and put them in longest-first arrangement in a regex:

use strict;
use warnings;

my $qr 
    = qr/ \G # right after last match
          ( distribution 
          | relationship 
          | beneficiary 
          | dependent 
          | subclass 
          | account
          | benefit 
          | address 
          | control 
          | history
          | percent 
          | action 
          | amount
          | conrol 
          | option 
          | period 
          | status 
          | class 
          | labor 
          | limit 
          | match 
          | notice
          | bank
          | code 
          | desc 
          | name 
          | role 
          | type 
          | age 
          | end 
          | pay
          | ps 
          )
    /x;

while ( <DATA> ) { 
    chomp;
    print;
    print ' -> ', join( ' ', m/$qr/g ), "\n";
}

__DATA__
payperiodmatchcode
labordistributioncodedesc
dependentrelationship
actionendoption
actionendoptiondesc
addresstype
addresstypedesc
historytype
psaddresstype
rolename
bankaccountstatus
bankaccountstatusdesc
bankaccounttype
bankaccounttypedesc
beneficiaryamount
beneficiaryclass
beneficiarypercent
benefitsubclass
beneficiaryclass
beneficiaryclassdesc
benefitactioncode
benefitactioncodedesc
benefitagecontrol
benefitagecontroldesc
ageconrolagelimit
ageconrolnoticeperiod

回答3:

Two things occur to me:

• this just isn't a task you can confidently attack programmatically, because ... English words don't work like that, they're often made of other words, so, is a given string "reportage" or "report age"? "Timepiece" or "time piece"?
• One way to do attack the problem would be to use anag which finds anagrams. After all, "time piece" is an anagram of "timepiece" ... now you just have to weed out the false positives.

回答4:

Here is a Lua program that tries longest matches from a dictionary:

local W={}
for w in io.lines("/usr/share/dict/words") do
    W[w]=true
end

function split(s)
    for n=#s,3,-1 do
        local w=s:sub(1,n)
        if W[w] then return w,split(s:sub(n+1)) end
    end
end

for s in io.lines() do
    print(s,"-->",split(s))
end

回答5:

Given that some words could be substrings of others, especially with multiple words smashed together, I think simple solutions like regexes are out. I'd go with a full-on parser, my experience being with ANTLR. If you want to stick with perl, I've had good luck using ANTLR parsers generated as Java through Inline::Java.

回答6:

Peter Norvig has a great python script that has a word segmentation function using unigram/bigram statistics. You want to take a look at the logic for the function segment2 in ngrams.py. Details are in the chapter Natural Language Corpus Data from the book Beautiful Data (Segaran and Hammerbacher, 2009). http://norvig.com/ngrams/