edu.stanford.nlp.trees.EnglishPTBTreebankCorrector Maven / Gradle / Ivy
Go to download
Show more of this group Show more artifacts with this name
Show all versions of stanford-corenlp Show documentation
Show all versions of stanford-corenlp Show documentation
Stanford CoreNLP provides a set of natural language analysis tools which can take raw English language text input and give the base forms of words, their parts of speech, whether they are names of companies, people, etc., normalize dates, times, and numeric quantities, mark up the structure of sentences in terms of phrases and word dependencies, and indicate which noun phrases refer to the same entities. It provides the foundational building blocks for higher level text understanding applications.
package edu.stanford.nlp.trees;
import edu.stanford.nlp.util.logging.Redwood;
import java.io.BufferedReader;
import java.io.StringReader;
import java.io.IOException;
import java.util.List;
import java.util.ArrayList;
import edu.stanford.nlp.trees.tregex.Macros;
import edu.stanford.nlp.trees.tregex.TregexPattern;
import edu.stanford.nlp.trees.tregex.TregexPatternCompiler;
import edu.stanford.nlp.trees.tregex.tsurgeon.Tsurgeon;
import edu.stanford.nlp.trees.tregex.tsurgeon.TsurgeonPattern;
import edu.stanford.nlp.util.Pair;
public class EnglishPTBTreebankCorrector implements TreebankTransformer {
/** A logger for this class */
private static Redwood.RedwoodChannels log = Redwood.channels(EnglishPTBTreebankCorrector.class);
private static final boolean DEBUG = false;
// "ai" is from "ai n't"; occasionally there is a bare "s" from "is"
private static final String BE =
"/^(?i:am|is|are|was|were|be|being|been|'s|'m|'re|s|ai)$/";
private static final String DO =
"/^(?i:do|did|does|doing|done)$/";
/** List of be, have, and get auxiliary forms: things that you would expect
* to have a VBN complement if anything verbal.
* A few times the apostrophe is missing on "'s", so we have just "s".
* Could also add become, feel, seem, remain for some PTB cases, but
* maybe should reparse them?
*/
private static final String BE_HAVE_GET =
"/^(?i:has|have|had|having|am|is|are|was|were|be|being|been|'s|'ve|'d|'m|'re|s|ai|get|gets|getting|got|gotten)$/";
private static final String MODAL_WORD =
"/^(?i:should|would|wo|could|may|might|ca|can|dare|will|'ll|must|shall|sha|'d)$/";
private static final String MODAL =
"[ < (__ < /^(?i:should|would|wo|could|may|might|ca|can|dare|will|'ll|must|shall|sha)$/) | < (MD < /^(?i:'d)$/) ]";
/** Contexts that take VB complement. This is a tregex subexpression
* that includes the tag. Something to specify next to VP:
* "(@VP" + MODAL_DO_TO + ")". na is gonna. sha is shan't (though doesn't occur).
* Not modals that take TO: need, ought. Or mighta.
* Note that "'d" can also be "had", so one needs to be careful on that!
*/
private static final String MODAL_DO_TO =
"[ < (__ < /^(?i:do|did|does|doing|done|to|na|should|would|wo|could|may|might|ca|can|dare|will|'ll|must|shall|sha)$/) | < (MD < /^(?i:'d)$/) ]";
/** Verbs that take bare VB complements.
* 'say' is sort of special, these are semi-direct speech cases where there
* are no inverted quotes but imperatives follow. Check for
* overgeneralization.
*/
private static final String BARE_VP_VERB =
" < (__ < /^(?i:help|helps|helping|helped|make|makes|making|made|see|sees|saw|seen|seeing|hear|heard|hears|hearing|let|lets|letting)$/)";
private static final String SAY_VERB =
" < (__ < /^(?i:say|says|said|saying)$/)";
private final List> ops;
public EnglishPTBTreebankCorrector() {
// initialize the transformations to be done
ops = new ArrayList<>();
TreebankLanguagePack tlp = new PennTreebankLanguagePack();
TregexPatternCompiler tpc = new TregexPatternCompiler(tlp.headFinder(), tlp.getBasicCategoryFunction());
Macros.addAllMacros(tpc, getBufferedReader(macroStr));
try {
BufferedReader br = getBufferedReader(editStr);
List tsp = new ArrayList<>();
for (String line; (line = br.readLine()) != null; ) {
TregexPattern matchPattern = tpc.compile(line);
tsp.clear();
if (DEBUG) log.info("Pattern is " + line + " [" + matchPattern + ']');
while (continuing(line = br.readLine())) {
TsurgeonPattern p = Tsurgeon.parseOperation(line);
if (DEBUG) log.info("Operation is " + line + " [" + p + ']');
tsp.add(p);
}
if ( ! tsp.isEmpty()) {
TsurgeonPattern tp = Tsurgeon.collectOperations(tsp);
ops.add(new Pair<>(matchPattern, tp));
}
} // while not at end of file
} catch (IOException ioe) {
ioe.printStackTrace();
}
}
/** Fix all the English Penn Treebank errors, or at least some of them (!).
*/
@Override
public MemoryTreebank transformTrees(Treebank tb) {
MemoryTreebank mtb = new MemoryTreebank(tb.treeReaderFactory(),
tb.encoding());
for (Tree t : tb) {
mtb.add(Tsurgeon.processPatternsOnTree(ops, t));
}
return mtb;
}
private static boolean continuing(String str) {
return str != null && ! str.matches("\\s*");
}
private static BufferedReader getBufferedReader(String source) {
return new BufferedReader(new StringReader(source));
}
// Note that backslashes and dollar signs need to be escaped.
// Maybe we should have an easy syntax for invoking Matcher.quoteReplacement?
private static final String macroStr =
"HAVE_AUX_WORD\t/^(?i:has|have|had|having|'s|'ve|'d)\\$/\n";
private static final String editStr =
// 1. Bung tree fixing
// NOTE: if you add more of these, make sure to group the string concatenations
// into chunks using parentheses, or the compiler will choke with a
// StackOverflowError (!)
(
// Fix a bad parse in wsj_0415.mrg
("@VP=adj < (NP < (NP=ex < (NN < growth)) < CC=bad < (NP=bd < (VB < service)))\n" +
"excise ex ex\n" +
"delete bad\n" +
"delete bd\n" +
"adjoinF (VP VP@ (CC and) (VP (VB service) (NP (NN debt)))) adj\n" +
'\n') +
// sec 24
("@SBAR=home <1 /^-NONE-$/=emp <2 (@S < (@NP <1 (DT=bad < that|That) <-1 NNS))\n" +
"delete emp\n" +
"relabel bad IN\n" +
"move bad >1 home\n" +
'\n') +
// sec 22 bad parse!
("@NP < (@NP=gone < (NN < authority)) < (@PP=bad < (TO < to) < (NP=vp < (NN=newv < block) < (NNS=newnp < mergers)))\n" +
"excise gone gone\n" +
"adjoin (S (NP-SBJ (-NONE- *)) VP@) bad\n" +
"relabel vp VP\n" +
"relabel newv VB\n" +
"adjoin (NP NN@) newnp\n" +
'\n') +
// Fix some cases of 'as well as' not made into a CONJP unit
// There are a few other wierd cases that should also be reviewed with the tregex
// well|Well|WELL , as|AS|As . as|AS|As !>(__ > @CONJP)
// but note that there are also non-CONJP uses as adverbial form of 'as good as'
// This bleeds retagging of 'well' inside NP below
("@NP < (__=bad < well|Well|WELL $, (__=before < as|AS|As) $. (__=after < as|AS|As))\n" +
"adjoinH (CONJP RB@) bad\n" +
"relabel bad CONJP\n" +
"move before >1 bad\n" +
"move after >-1 bad\n" +
'\n' )
) +
// 2. POS tag fixing
// 2.a. Ones specific to a phrasal category
// 2.a.i NP
(
("@NP < (/^``$/ < /^`$/) < (POS=bad < /^'$/)\n" +
"relabel bad /''/\n" +
'\n') +
("@NP < (IN|WDT=bad < /^(?:a|that|That)$/)\n" +
"relabel bad DT\n" +
'\n') +
("@NP < (IN=bad < /^(?:so|about)$/)\n" +
"relabel bad RB\n" +
'\n') +
("@NP < (IN=bad < /^(?:fiscal|next)$/)\n" +
"relabel bad JJ\n" +
'\n') +
("@NP < (RB=bad < /^(?:a|that|Some)$/)\n" +
"relabel bad DT\n" +
'\n') +
("@NP < (RB=bad < most $. DT)\n" +
"relabel bad PDT\n" +
'\n') +
("@NP < (RB=bad < /^(?:MORE)$/)\n" +
"relabel bad JJR\n" +
'\n') +
("@NP < (NN=bad < the)\n" +
"relabel bad DT\n" +
'\n') +
("@NP < (/^VB/=bad < won)\n" +
"relabel bad NN\n" +
'\n') +
// "well". Other rules for under INTJ further below.
("@NP < (/^RB/=bad < well|WELL|Well)\n" +
"relabel bad NN\n" +
'\n') +
// treat like "sheep": if it would be "dollars", make it plural
("@NP < (NN=bad < yen|won [ $- (CD !< one|One|1) | $- (@QP !< (CD < one|One|1)) ] )\n" +
"relabel bad NNS\n" +
'\n') +
// it's 69 NNPS, 17 NNS, 3 NNP for Democrats; 46, 17, 3 for Republicans; all political party mentions
("@NP < (NNP|NNS=bad < Democrats|Republicans)\n" +
"relabel bad NNPS\n" +
'\n') +
("@NP < (CD=bad < the)\n" +
"relabel bad DT\n" +
'\n') +
("@NP < (JJ=bad < the)\n" +
"relabel bad DT\n" +
'\n') +
("@NP|NX < (NNP=bad < the)\n" +
"relabel bad DT\n" +
'\n') +
("@NP=bad < (NNP=badder < Technically|Historically)\n" +
"relabel bad ADVP\n" +
"relabel badder RB\n" +
'\n') +
("@NP < (RB=bad < /^(?:McNally)$/)\n" +
"relabel bad NNP\n" +
'\n') +
("@NP < (RB=bad < /^(?:vice|night|multifamily|hand|fist)$/)\n" +
"relabel bad NN\n" +
'\n') +
("@NP < (RP=bad < /^(?:whole)$/)\n" +
"relabel bad JJ\n" +
'\n') +
("@NP < (RP=bad < Howard) < (NN=badder < /^A\\.$/)\n" +
"relabel bad NNP\n" +
"relabel badder NNP\n" +
'\n') +
("@NP < (JJ=bad < (First , (__ !> /^``$/ !> /^-LRB-$/ !> /^PRP\\$$/)) $. NNP)\n" +
"relabel bad NNP\n" +
'\n') +
("@NP < (JJ=bad < /^(?:U\\.S\\.|Sept\\.)$/)\n" +
"relabel bad NNP\n" +
'\n') +
("@NP < (JJ=bad <1 Sharp) !<2 __\n" +
"relabel bad NNP\n" +
'\n') +
("@NP < (JJ=bad < /^(?:mine)$/)\n" +
"relabel bad NN\n" + // noun usages (mining)
'\n') +
("@NP <-1 (JJ=bad < firm)\n" +
"relabel bad NN\n" + // noun usages
'\n') +
("@NP < (JJ=bad < /^(?:ours)$/)\n" +
"relabel bad PRP\n" +
'\n') +
// don't regard aluminum as an adjective (color)
// all uses of plastic are also for the noun, not adjectivally for things showing plasticity
("@NP < (JJ=bad < aluminum|plastic|textile)\n" +
"relabel bad NN\n" +
'\n') +
// gold and silver can validly be colours but are usually metals in WSJ. Use noun to decide
("@NP < (JJ=bad < gold|silver|bronze . stocks|bat|standard|reserves|prices|market|price|prices|fund|funds|consumers|use|commemorative|medal|bullion|exploration|producer|producers|trader|traders)\n" +
"relabel bad NN\n" +
'\n') +
// gold and silver can validly be colours but are usually metals in WSJ. Use noun to decide
("@NP <: (JJ=bad < gold|silver|bronze)\n" +
"relabel bad NN\n" +
'\n') +
// make all uses of 'the House' (Congress) NNP, like most, and
// like all uses of 'the Senate'
("@NP <2 (NN=bad < House $- (DT < /^[Tt]he$/))\n" +
"relabel bad NNP\n" +
'\n') +
("@NP < (NNP=bad < Democrats|Republicans)\n" +
"relabel bad NNPS\n" +
'\n') +
("@NP < (NNS=bad < Democrats|Republicans , __)\n" +
"relabel bad NNPS\n" +
'\n') +
("@NP < (NN=bad < /^(?:Chapman|Ok|Oslo|Boeing|Jan\\.|Sept\\.|Oct\\.|Nov\\.|Dec\\.|Treasury|Esso)$/)\n" +
"relabel bad NNP\n" +
'\n') +
("@NP < (NN=bad < /^(?:members|bureaus|days|outfits|institutes|innings|write-offs|wines|trade-offs|tie-ins|thrips|1980s|1920s|receivables|earnings)$/)\n" +
"relabel bad NNS\n" +
'\n') +
("@NP < (NNP=bad < Dutch $. NN|NNS|JJ|VBG)\n" +
"relabel bad JJ\n" +
'\n') +
("@NP < (NN=bad < /^(?:this)$/)\n" +
"relabel bad DT\n" +
'\n') +
("@NP < (/^:/=bad < /^(?:')$/)\n" +
"relabel bad /''/\n" +
'\n') +
("@NP < (NNS=bad < /^(?:start-up|ground-handling|word-processing|T-shirt|co-pilot|sell-off)$/)\n" +
"relabel bad NN\n" +
'\n') +
// not clear why Sens isn't NNPS
("@NP < (NNS=bad < /^(?:Sens\\.|Aichi|Asahi|Cincinnati|Hawaii|Pepsi)$/)\n" +
"relabel bad NNP\n" +
'\n') +
// VBZ under @NP. Strong rule ordering!
// move up misplaced possessives
("@NP <1 (@NP=dest !< @NP . (__=wrong < /^\u0027s$/))\n" +
"move wrong >-1 dest\n" +
'\n') +
// fix bung syntax in wsj_0295 and wsj_1142
("@S < (@NP=bad < PRP < (VBZ=bottom < /^'s$/)) < (@VP=adj < VBN|VBG)\n" +
"adjoin (VP (VBZ 's) VP@) adj\n" +
"delete bottom\n" +
'\n') +
("@S < (NP-SBJ < (PRP < I)) < (VP < (VB=bad < think) < SBAR)\n" +
"relabel bad VBP\n" +
'\n') +
// and an extra weird NP in wsj_0446 and wsj_1101
("@VP < (@NP=bad < (VBZ < kills|blames) < @NP)\n" +
"excise bad bad\n" +
'\n') +
// then turn all 's VBZ ino a POS
// except for Everything's a Dollar Inc. !
("@NP < (VBZ=bad < /^'s$/) !< (NNP < Everything)\n" +
"relabel bad POS\n" +
'\n') +
// and then turn all other VBZ that aren't the 's into NNS.
// 100% precision now!
// CHECK IF CAN DELETE kills HERE NOW!
("@NP|NX < (VBZ=bad !< /^(?:'s|kills)/)\n" +
"relabel bad NNS\n" +
'\n') +
// but turn POS under NP to PRP if it is the 's of let's
("@NP < (POS=bad < /^'s$/) > (@S > (@VP < (VB < let)))\n" +
"relabel bad PRP\n" +
'\n') +
// fix VB at root level
(
// these ones are VBD
("@S < (@VP < (VB=bad < /...ed$/ )) < (/^NP-SBJ/ !< /^-NONE-$/) > (__ !> __)\n" +
"relabel bad VBD\n" +
'\n') +
// these ones are also VBD in practice (though in principle ambiguous)
("@S < (@VP < (VB < let|cut|bid| )) < (/^NP-SBJ/ !< /^-NONE-$/) > (__ !> __)\n" +
"relabel bad VBD\n" +
'\n') +
// Then, except in a few weird cases (treebank err, missing letter, reduced "better not" they should be VBP
("@S < (@VP < (VB=bad !, Tait|MD|not )) < (/^NP-SBJ/ !< /^-NONE-$/) > (__ !> __)\n" +
"relabel bad VBP\n" +
'\n')
) +
// VBP under NP. First fix one that should be a verb! NP is wrong
("@NP=bad < (VBP < are) > (@VP > (@S < NP-SBJ))\n" +
"excise bad bad\n" +
'\n') +
("@NP < (VBP=bad < charge)\n" +
"relabel bad NN\n" +
'\n') +
("@NP < (VBP=bad < the)\n" +
"relabel bad DT\n" +
'\n') +
("@NP < (VBP=bad < we)\n" +
"relabel bad PRP\n" +
'\n') +
("@NP < (VBP=bad < /^[A-Z]/)\n" +
"relabel bad NNP\n" +
'\n') +
// VBN under NP
("@NP < (VBN=bad < Applied !$ __)\n" +
"relabel bad NNP\n" +
'\n') +
("@NP < (VBG=bad < preferred)\n" +
"relabel bad VBN\n" +
'\n') +
("@NP < (VB=bad < The)\n" +
"relabel bad DT\n" +
'\n') +
("@NP < (VB=bad < allowed)\n" +
"relabel bad VBD\n" +
'\n') +
("@NP <-1 (JJR=bad < cleaner)\n" +
"relabel bad NN\n" +
'\n') +
("@NP < (VB=bad < /^(?:Nov\\.|Jan\\.|Dec\\.|Tandy|Release|Orkem|McDonald|Citicorp|Anne)$/)\n" +
"relabel bad NNP\n" +
'\n') +
("@NP < (VB=bad < /^(?:short|key|many|last|further)$/)\n" +
"relabel bad JJ\n" +
'\n') +
("@NP < (VB=bad < lower)\n" +
"relabel bad JJR\n" +
'\n') +
("@NP < (VB=bad < /^(?:spill|watch|review|risk|realestate|love|experience|control|Transport|mind|term|program|gender|audit|blame|stock|run|group|affect|rent|show|accord|change|finish|work|schedule|influence|school|freight|growth|travel|call|autograph|demand|abuse|return|defeat|pressure|bank|notice|tax|ooze|network|concern|pit|contract|cash|help|lunch|combat|pot|care|date|Streetspeak|face|effect|worry)$/)\n" +
"relabel bad NN\n" +
'\n') +
("@NP <1 (NNP=bad < Officials|Cartoonists|Prices)\n" +
"relabel bad NNS\n" +
'\n') +
("@NP=badder < (NNP=bad < Currently)\n" +
"relabel bad RB\n" +
"relabel badder ADVP-TMP\n" +
'\n') +
// nth-quarter as a modifier: may as well treat it as JJ as manual says and majority of instances are
("@NP < (NN=bad < /^(?i:first|second|third|fourth)-quarter$/ $+ __)\n" +
"relabel bad JJ\n" +
'\n') +
// may as well take all of "K mart" as a proper noun
("@NP < (NNP < K $+ (NN=bad < mart))\n" +
"relabel bad NNP\n" +
'\n') +
("@NP < (PRP=bad < US & $. __)\n" +
"relabel bad NNP\n" +
'\n') +
("@NP < (PRP=bad < her & $. __)\n" +
"relabel bad /PRP$/\n" +
'\n') +
("@NP <1 (PRP=bad < his) !<2 __\n" +
"relabel bad /PRP$/\n" +
'\n') +
("VBD=bad [ > @NP | > (@ADJP < CC|CONJP > @NP) ]\n" +
"relabel bad VBN\n" +
'\n') +
("@NP < (NN=bad < Time) < (NNP < Warner)\n" +
"relabel bad NNP\n" +
'\n') +
("@NP < (MD=bad < Can|May)\n" +
"relabel bad NNP\n" +
'\n') +
// there are a number of cases of 'the the' -- errors in sources or
// mistakes in Perl scripts?
// and one case of a/DT half/DT
("@NP <1 (DT=bad !< the $. (DT !< half))\n" +
"relabel bad PDT\n" +
'\n') +
// filling out NP with PP modifier. This rule should precede
("@NP=place < (@NP <1 DT !<2 __ $. (JJ=bad $. (NN=badder $. PP)))\n" +
"move bad >-1 place\n" +
"move badder >-1 place\n" +
'\n') +
// filling out NP with PP modifier. This rule should follow
("@NP=place < (@NP <1 DT|JJ !<2 __ $. (NN=bad $. PP))\n" +
"move bad >-1 place\n" +
'\n') +
// NEWSPAPERS
("@NP < (NNPS=bad < NEWSPAPERS ! $ /^NN/)\n" +
"relabel bad NNS\n" +
'\n') +
("@NP < (@NP < (NNPS=bad < CERTIFICATES)) < (PP < (IN < OF) < (NP < (__ < DEPOSIT)))\n" +
"relabel bad NNS\n" +
'\n') +
("@NP < (@NP < (__ < CERTIFICATES)) < (PP < (IN < OF) < (NP < (NNPS=bad < DEPOSIT)))\n" +
"relabel bad NN\n" +
'\n') +
("@NP < (NNP=bad < DISCOUNT) < (NNP=badder < RATE)\n" +
"relabel bad NN\n" +
"relabel bad NN\n" +
'\n') +
("@NP < (JJ=bad < DISCOUNT) < (NN < RATE)\n" +
"relabel bad NN\n" +
"relabel bad NN\n" +
'\n') +
("@NP < (__ < chief $. (NN=bad < executive $. (NN < officer)))\n" +
"relabel bad JJ\n" +
'\n') +
("@NP < (NN=bad < chief $. (__ < executive $. (NN < officer)))\n" +
"relabel bad JJ\n" +
'\n') +
("@NP <: (NNP=bad < /^'s$/)\n" +
"relabel bad PRP\n" +
'\n') +
"NP-TMP=bad < (NN < Leisure)\n" +
"relabel bad NP\n" +
'\n' +
"@NP < (RB=bad < well !$, (RB < as) $,, (DT < The|the))\n" +
"relabel bad NN\n" +
'\n' +
// chief executive officer
("@NP < (NN < officer $- (NN=bad < executive $- (__ < chief)))\n" +
"relabel bad JJ\n" +
'\n') +
// chief executive officer
("@NP < (NN < officer $- (JJ < executive $- (NN=bad < chief)))\n" +
"relabel bad JJ\n" +
'\n') +
// the infamous "Ad Notes"
("@NP < (NNP=bad < Ad) < (/^NN/ < Notes)\n" +
"relabel bad NN\n" +
'\n') +
// the infamous "Ad Notes"
("@NP < (/^NN/ < Ad) < (NNPS=bad < Notes)\n" +
"relabel bad NNS\n" +
'\n') +
// the infamous "Ad Notes"
("NP=bad < (/^NN/ < Ad) < (/^NN/ < Notes)\n" +
"relabel bad NP-HLN\n" +
'\n') +
("@NP < (NN=bad < nonperforming)\n" +
"relabel bad JJ\n" +
'\n') +
("@NP < (IN=bad < next|Next) < NN|NNP\n" +
"relabel bad JJ\n" +
'\n') +
("NP-PRD=x < (JJ=y $+ (NN=z < simple))\n" +
"relabel x ADJP-PRD\n" +
"relabel y RB\n" +
"relabel z JJ\n" +
'\n') +
("NP-PRD=x <: (NN=z < simple)\n" +
"relabel x ADJP-PRD\n" +
"relabel z JJ\n" +
'\n') +
// this needs to be after the above two "simple" ones
("NN=z < simple\n" +
"relabel z JJ\n" +
'\n') +
"") +
("@PP < (IN < behind) < (@NP < PRP=bad < (RB=no < back))\n" +
"relabel bad /PRP$/\n" +
"relabel no NN\n" +
'\n') +
// equally infamous "Who's News:"
("@SBAR=bad < (@WHNP < (WP < WHO)) < (S=badder < (VP < (@NP < (NN < NEWS))))\n" +
"relabel bad /SBARQ-HLN/\n" +
"relabel badder SQ\n" +
'\n') +
// equally infamous "Who's News:"
("@SBARQ < (@WHNP < (WP < WHO)) < (SQ < (VP < (@NP < (NNP=bad < NEWS))))\n" +
"relabel bad NN\n" +
'\n') +
// equally infamous "Who's News:"
("@SBARQ < (@WHNP < (WP < WHO)) < (S=bad < (VP < (@NP < (NN < NEWS))))\n" +
"relabel bad SQ\n" +
'\n') +
("@WHNP < @WHNP=bad < @WHPP\n" +
"relabel bad NP\n" +
'\n') +
// How much always has 'much' as a JJ
("/^WH/ < (WRB < /^(?i:how)$/) < (__=bad < (much !> JJ))\n" +
"relabel bad JJ\n" +
'\n') +
// the VP is for extraposed relatives....
("@WHNP|WHADVP < (VBP|DT|IN=bad < /^(?i:that)$/) [ > (@SBAR > @NP|VP) | > (@SBAR > (@SBAR < /^(?:CC|CONJP|,)$/ > @NP|VP)) ]\n" +
"relabel bad WDT\n" +
'\n') +
("@UCP < (RB=bad < multifamily)\n" +
"relabel bad NN\n" +
'\n') +
("@PRT < (RBR=bad < in)\n" +
"relabel bad RP\n" +
'\n') +
("@PRT < (NNP=bad < up)\n" +
"relabel bad RP\n" +
'\n') +
// PP parent
("@PP < (RP=bad < through) < @NP\n" +
"relabel bad IN\n" +
'\n') +
("@PP < (RP|NN=bad < in) < @NP\n" +
"relabel bad IN\n" +
'\n') +
("@PP < (RB=bad < for|For|after|After|past|Past|under|Under)\n" +
"relabel bad IN\n" +
'\n') +
("@PP < (JJ=bad < if)\n" +
"relabel bad IN\n" +
'\n') +
// VP parent
("@VP=bad < (IN < past) < @NP\n" +
"relabel bad PP\n" +
'\n') +
("@VP < (RB=bad < back $. (PRT < (RP < down))) > (@SINV|SQ|VP < MD)\n" +
"relabel bad VB\n" +
'\n') +
("@VP < (IN=bad < complicated) > @S\n" +
"relabel bad VBD\n" +
'\n') +
("@VP < (IN=bad < near) > @VP\n" +
"relabel bad VB\n" +
'\n') +
("@VP < (IN=bad < like|post) > (@SQ|VP < /^(?:VB|MD)/)\n" +
"relabel bad VB\n" +
'\n') +
("@VP < (IN=bad < like|post) [ > @S | > (@VP < CC|CONJP > @S) ]\n" +
"relabel bad VBP\n" +
'\n') +
("@VP < (/^VBD?$/=ins < take|sold) < (IN=bad < off)\n" +
"delete bad\n" +
"insert (PRT (RP off)) $- ins\n" +
'\n') +
// NNS under VP. Easy.
("@VP < NNS=bad\n" +
"relabel bad VBZ\n" +
'\n') +
// ordered rules for NN under VP
// first fix the ones ending in -ing
("@VP < (NN=bad < /.{2}ing$/)\n" +
"relabel bad VBG\n" +
'\n') +
("@VP < (NN=bad [ < set|beat|bid|redone|reset|hurt|underwritten|overrun | < /.{2}[^e]ed$/ ]) [ > (@VP < (/^VB/ < " + BE_HAVE_GET + ")) | > (@NP < @NP)]\n" +
"relabel bad VBN\n" +
'\n') +
("@VP < (NN=bad < agreed|set|rebounded|fell) [ > @S | > (@VP < @CC|CONJP > @S) ]\n" +
"relabel bad VBD\n" +
'\n') +
// contexts where an untensed verb should occur
// an imperative is VB! The 'say' cases are direct speech imperatives
("@VP !< /^VB/ < (NN|NNP|JJ=bad !< /...(?i:ing)$/) [ > (@VP|SINV|SQ " + MODAL_DO_TO + ") | > (@VP|SINV|UCP|SQ < CC|CONJP > (@VP|SINV|UCP|SQ " + MODAL_DO_TO + ")) | > (@S > (@VP " + BARE_VP_VERB + ")) | > (@VP " + BARE_VP_VERB + ") | > (@S < (/^NP-SBJ/ < /^-NONE-$/) > (@VP " + SAY_VERB + ")) ]\n" +
"relabel bad VB\n" +
'\n') +
// file should maybe be VB - subjunctive (look at higher verb? wsj_0979)
("@VP !< /^VB/ < (NN=bad [ [ !< /s$/ & !< /e[dn]$/ & !< /ing$/ ] | < stress ] ) [ > @S | > (@VP < @CONJP|CC > @S) ]\n" +
"relabel bad VBP\n" +
'\n') +
// not just anything ending in s. Definitely not if -ss, but also 'focus'
("@VP < (NN=bad < institutes) > @S\n" +
"relabel bad VBZ\n" +
'\n') +
("@VP < (VBP=bad !< /...(?i:ing)$/) [ > (@VP|SQ|SINV " + MODAL_DO_TO + ") | > (@VP|UCP|SQ|SINV < CC|CONJP > (@VP|UCP|SQ|SINV " + MODAL_DO_TO + ")) | > (@S > (@VP " + BARE_VP_VERB + ")) | > (@VP " + BARE_VP_VERB + ") | > (@S < (/^NP-SBJ/ < /^-NONE-$/) > (@VP " + SAY_VERB + ")) ]\n" +
"relabel bad VB\n" +
'\n') +
// Don't allow SINV for VBN as get participial clauses
// unless the SINV has a modal under it
// Allow VBN in complement of SEE or HEAR (but not LET or MAKE)
("@VP < (VBN=bad !< /...(?i:ing)$/ !< /...(?i:ed)$/) [ > (@VP|SQ " + MODAL_DO_TO + ") | > (@VP|SQ < CC|CONJP !< /^VB/ > (@VP|SQ " + MODAL_DO_TO + ")) | > (@SINV " + MODAL + ") | > (@SINV < CC|CONJP !< /^VB/ > (@SINV " + MODAL + ")) ]\n" +
"relabel bad VB\n" +
'\n') +
// for transitive, it can't be VBN under see/hear
("@VP < (VBN=bad !< /...(?i:ing)$/) < (NP !< /^-NONE-$/) [ > (@S > (@VP " + BARE_VP_VERB + ")) | > (@VP " + BARE_VP_VERB + ") | > (@S < (/^NP-SBJ/ < /^-NONE-$/) > (@VP " + SAY_VERB + ")) ]\n" +
"relabel bad VB\n" +
'\n') +
("@VP < (NN=bad < relocate) > (@VP < @CONJP > (@VP < MD))\n" +
"relabel bad VB\n" +
'\n') +
("@VP < (NN=bad < might|will)\n" +
"relabel bad MD\n" +
'\n') +
// ordered rules for NNP under VP
// it's VBD for SINV
("@VP < (NNP=bad < /...(?i:ed)$/) [ > (SINV !> /TTL/) | > (@VP < @CONJP|CC > (SINV !> /TTL/)) ]\n" +
"relabel bad VBD\n" +
'\n') +
// it's VBN not VBD because the one or two there are headline-speak
("@VP < (NNP=bad < /...(?i:ed)$/) [ > (S|S-HLN|S-ADV !> /TTL/) | > (@VP < @CONJP|CC > (S|S-HLN|S-ADV !> /TTL/)) | > (@VP < (__ < " + BE_HAVE_GET + ")) ]\n" +
"relabel bad VBN\n" +
'\n') +
// it's VBN not VBD because the one or two there are headline-speak
("@VP < (NNP=bad < Got|Gotten) [ > (S|S-HLN|S-ADV !> /TTL/) | > (@VP < @CONJP|CC > (S|S-HLN|S-ADV !> /TTL/)) | > (@VP < (__ < " + BE_HAVE_GET + ")) ]\n" +
"relabel bad VBN\n" +
'\n') +
// ordered rules for NNP under VP (2 dots as want to get "Adds")
("@VP < (NNP=bad < /..[^Ss](?i:s)$/) [ > (S|S-HLN|SINV !> /TTL/) | > (@VP < @CONJP|CC > (S|SINV|S-HLN !> /TTL/)) ]\n" +
"relabel bad VBZ\n" +
'\n') +
// ordered rules for NNP under VP
("@VP !< /^VB/ < (NNP=bad < /(?i:ing)$/)\n" +
"relabel bad VBG\n" +
'\n') +
// ordered rules for NNP under VP
("@VP < NNP=bad [ > (S|S-HLN !> /TTL/) | > (@VP < @CONJP|CC > (S|S-HLN !> /TTL/)) ]\n" +
"relabel bad VBP\n" +
'\n') +
// this one is in a title....
("@VP < (NNP=bad < are) [ > @S | > (@VP < @CONJP|CC > @S) ]\n" +
"relabel bad VBP\n" +
'\n') +
/* ----------------------------
tregex3 -f -w '@VP < NN' | & less
// unresolved: some bid, ask cases
// so far only done 1-1000 inclusive
} else if (word.equals("reconfirm")) {
cat = changeBaseCat(cat, "VB");
} else if (word.equals("rebounded")) {
cat = changeBaseCat(cat, "VBD");
}
// still need to do NNP > VP rules as well!
--------------------- */
/*
} else if (baseCat.equals("NNP")) {
if (word.equals("GRAB")) {
cat = changeBaseCat(cat, "VBP");
} else if (word.equals("mature")) {
cat = changeBaseCat(cat, "VB");
} else if (word.equals("Face")) {
cat = changeBaseCat(cat, "VBP");
} else if (word.equals("are")) {
cat = changeBaseCat(cat, "VBP");
} else if (word.equals("say")) {
cat = changeBaseCat(cat, "VBP");
} else if (word.equals("Added")) {
cat = changeBaseCat(cat, "VBD");
} else if (word.equals("Adds")) {
cat = changeBaseCat(cat, "VBZ");
} else if (word.equals("BRACED")) {
cat = changeBaseCat(cat, "VBD");
} else if (word.equals("REQUIRED")) {
cat = changeBaseCat(cat, "VBN");
} else if (word.equals("REVIEW")) {
cat = changeBaseCat(cat, "VB");
} else if (word.equals("code-named")) {
cat = changeBaseCat(cat, "VBN");
} else if (word.equals("Printed")) {
cat = changeBaseCat(cat, "VBN");
} else if (word.equals("Rated")) {
cat = changeBaseCat(cat, "VBN");
} else if (word.equals("FALTERS")) {
cat = changeBaseCat(cat, "VBZ");
} else if (word.equals("Got")) {
cat = changeBaseCat(cat, "VBN");
} else if (word.equals("Adds")) {
cat = changeBaseCat(cat, "VBZ");
}
*/
// Modal fixes: of 'd tagged as VBD and others tagged VBP
// first fix one where the complement verb tag is wrong
("@VP < (VBD < /^(?i:'d)$/) < (@VP < (VB=badder < seen))\n" +
"relabel badder VBN\n" +
'\n') +
("@VP < (VBP=bad < " + MODAL_WORD + ") < (@VP < VB)\n" +
"relabel bad MD\n" +
'\n') +
("@VP < (VBP=bad < /^(?i:'d)$/) < (@VP < VB)\n" +
"relabel bad MD\n" +
'\n') +
// this must be ordered before the have/be auxiliary complement fix!
("@VP < POS=bad\n" +
"relabel bad VBZ\n" +
'\n') +
("@VP < (VBD=bad < heaves)\n" +
"relabel bad VBZ\n" +
'\n') +
("@VP < (VB=bad < /.{2}[^e]ed$/) > @S\n" +
"relabel bad VBD\n" +
'\n') +
("@VP < (VB=bad < /^(?i:.{2,}[^e]ed|reset|run|become|hit|remade|gone|rid|put|hurt|become)$/) [ > (@VP < (/^VB/ < " + BE_HAVE_GET + ")) | > (@NP < @NP) ]\n" +
"relabel bad VBN\n" +
'\n') +
("@VP < (VBN=bad < has)\n" +
"relabel bad VBZ\n" +
'\n') +
("@VP < (VBN=bad < grew|fell|had) [ > @S | > (@VP < CONJP|CC > @S)]\n" +
"relabel bad VBD\n" +
'\n') +
// Doing the VBN to VBD is too general because of auxiliary deletion in
// contracted constructions, headlines, etc.
// Nevertheless, there are quite a few cases....
// Try ones with objects??
// "@VP < VBN=bad [ > @S | > (@VP < CONJP|CC > @S) ]\n" +
// "relabel bad VBN\n" +
// "\n" +
// Below are three patterns that doesn't overgeneralize and still gets a
// lot of the cases: note that they don't use @ so as to not match NP-TMP
// or S-HLN, and checks for non-NULL constituents
// must be finite in clause with overt NPsubj and NPobj
("@VP < VBN=bad < (NP !< /^-NONE-$/) [ > (S < (@NP !< /^-NONE-$/)) | > (@VP < CONJP|CC > (S < (@NP !< /^-NONE-$/))) ]\n" +
"relabel bad VBD\n" +
'\n') +
// root main clauses with overt subj must be finite
("@VP < VBN=bad [ > (S < (@NP !< /^-NONE-$/) > (__ !> __)) | > (@VP < CONJP|CC > (S < (@NP !< /^-NONE-$/) > (__ !> __))) ]\n" +
"relabel bad VBD\n" +
'\n') +
// that clauses (or that-less that clauses) with overt subjects are finite
("@SBAR [ < (/^-NONE-$/ < /^0$/) | < (IN < that) ] < (@S < (NP-SBJ !< /^-NONE-$/) < (@VP < VBN=bad))\n" +
"relabel bad VBD\n" +
'\n') +
// similar but more limited corrections of VB to VBP
// have to beware getting VB in imperatives, subjunctives, etc.
// root main clauses with overt subj must be finite
("@VP < VB=bad [ > (S < (NP-SBJ !< /^-NONE-$/) > (__ !> __)) | > (@VP < CONJP|CC > (S < (NP-SBJ !< /^-NONE-$/) > (__ !> __))) ]\n" +
"relabel bad VBP\n" +
'\n') +
// Should have a finite VBP not a VB in a finite relative clause
("@NP < @NP < (@SBAR < @WHNP < (@S < (VP < VB=bad)))\n" +
"relabel bad VBP\n" +
'\n') +
("@VP < CONJP|CC <1 (VBP $.. VB=bad)\n" +
"relabel bad VBP\n" +
'\n') +
("@VP < (VBP=bad < has)\n" +
"relabel bad VBZ\n" +
'\n') +
// JJ under VP rewrites (see also above for BARE_VP_VERB generalizations)
("@VP < (JJ=bad < own|elaborate) [ > @S | > (@VP < CONJP|CC > @S)]\n" +
"relabel bad VBP\n" +
'\n') +
("@VP < (JJ=bad < /..ing$/) < (@S < (@NP !< /^-NONE-$/) < (VP < TO))\n" +
"relabel bad VBG\n" +
'\n') +
("@VP < (JJ=bad < /...ed$/) [ > @S | > (@VP < CONJP|CC > @S)]\n" +
"relabel bad VBD\n" +
'\n') +
("@VP < (JJ=bad < pressured|known) [ > (@VP < (__ < " + BE + ")) | > (@VP < CONJP|CC > (@VP < (__ < " + BE + "))) ]\n" +
"relabel bad VBN\n" +
'\n') +
("@VP < (JJ=bad < /(?i:..e[dn])$/) > (@VP < (__ < HAVE_AUX_WORD))\n" +
"relabel bad VBN\n" +
'\n') +
("@VP < (JJ=bad < /.{2}ing$/) [ > @S | > (@VP < CONJP|CC > @S)]\n" +
"relabel bad VBG\n" +
'\n') +
("@VP < (JJ=bad < to)\n" +
"relabel bad TO\n" +
'\n') +
("@VP|S < (JJ=bad < all|ALL|All)\n" +
"relabel bad RB\n" +
'\n') +
// VBN after have/be auxiliaries
("@VP < VBD=bad [ > (@VP < (/^VB/ < " + BE_HAVE_GET + ")) | > (@VP < CONJP|CC > (@VP < (/^VB/ < " + BE_HAVE_GET + "))) | > (@NP < @NP) ]\n" +
"relabel bad VBN\n" +
'\n') +
// Asked, Warned, etc. at start of sentence
("@VP < (VBD=bad < /[A-Z]/) > S-ADV\n" +
"relabel bad VBN\n" +
'\n') +
("@VP < (/^VB/ $. (IN=bad < up|off))\n" +
"adjoin (PRT (RP@)) bad\n" +
'\n') +
// remove the two (!) unsure phrasal categories: ADVP|PRT, PRT|ADVP
("@VP < /^VB/ < (/^(?:ADVP|PRT)\\|(?:ADVP|PRT)$/=bad < (RB|NN=badder < back))\n" +
"relabel bad PRT\n" +
"relabel badder RP\n" +
'\n') +
// make other instances of 'win back' particle verbs
("@VP < (/^VB/ < win|wins|winning|won) < (@ADVP=bad < (RB=badder < back))\n" +
"relabel bad PRT\n" +
"relabel badder RP\n" +
'\n') +
// TODO: do other instances of put back. Worth looking at Treebank II -PUT
("@VP < (PDT=bad < all)\n" +
"relabel bad RB\n" +
'\n') +
("@VP < (PRT < (VBP=bad < down))\n" +
"relabel bad RP\n" +
'\n') +
("@VP < (PRT=bad < (RBS < best))\n" +
"relabel bad ADVP\n" +
'\n') +
("@VP <1 (VB=bad < plea) <2 (NN=badder < bargain) > (@VP" + MODAL_DO_TO + ")\n" +
"relabel bad NN\n" +
"relabel badder VB\n" +
'\n') +
// ADJP rules
(
("@ADJP < UH=bad\n" +
"relabel bad JJ\n" +
'\n') +
// "alive and well"
("@ADJP < (JJ < alive) < CC < (RB=bad < well)\n" +
"relabel bad JJ\n" +
'\n') +
("@ADJP < (JJ=bad < more)\n" +
"relabel bad JJR\n" +
'\n') +
("Korean > (NNP=x $- (NNP=y < South))\n" +
"relabel x JJ\n" +
"relabel y JJ\n" +
'\n') +
("@ADJP <1 (NNP=x < /^(?:New|San|Los|Des|St\\.|Washington|Hong)$/) <2 (__ < /^(?:York|Francisco|Angeles|London|Orleans|Zealand|Diego|Moines|Louis|D\\.C\\.|Kong)-based$/) !<3 __\n" +
"relabel x JJ\n" +
'\n') +
("@ADJP <1 (__ < New|San|Los) <2 (NNP=y < /^(?:York|Francisco|Angeles)-based$/) !<3 __\n" +
"relabel y JJ\n" +
'\n') +
("@ADJP < (NN=bad < firm|due|permissible)\n" +
"relabel bad JJ\n" +
'\n') +
("@ADJP < (NNS=bad < due)\n" +
"relabel bad JJ\n" +
'\n') +
("@ADJP < (NNP=bad < READY)\n" +
"relabel bad JJ\n" +
'\n') +
("@ADJP < (RB=bad < free|clear|tight|sure|particular|due)\n" +
"relabel bad JJ\n" +
'\n') +
// consider also "hard" -- many but not all cases should be JJ
("@ADJP < (RB=bad < likely) > @VP\n" +
"relabel bad JJ\n" +
'\n') +
// reconsider also cases not under VP, but some looked more complex....
("@ADJP < (VB=bad < /^(?i:stock|No\\.)$/)\n" +
"relabel bad NN\n" +
'\n') +
("@ADJP < (VBP=bad < fit|close)\n" +
"relabel bad JJ\n" +
'\n') +
("@ADJP < (VB=bad < secure|keen|quiet)\n" +
"relabel bad JJ\n" +
'\n') +
("@ADJP < JJ < (IN=bad < that)\n" +
"relabel bad RB\n" +
'\n') +
"") +
("@QP < (IN|JJ|RBR|RP=bad < about)\n" +
"relabel bad RB\n" +
'\n') +
("@QP < (JJ=bad < as)\n" +
"relabel bad RB\n" +
'\n') +
// look at 'as' examples, but don't want to maul 'as much as X' idiom
("@QP < (JJ|JJS=bad < more|less)\n" +
"relabel bad JJR\n" +
'\n') +
// need to work out whether to relabel all the RBR ones similarly....
("@QP < (RP=bad < up $. (TO 1 dest\n" +
'\n') +
// tree in wsj_2155
("@SBAR=sbar < (/^WP\\$$/=wrong $. (S=ins < NP-SBJ=fix < VP))\n" +
"move wrong >1 fix\n" +
"relabel fix WHNP\n" +
"move fix >1 sbar\n" +
"insert (NP-SBJ (-NONE- *T*)) >1 ins\n" +
'\n') +
// tree in wsj_1457
("@SBAR <1 @WHNP <2 @S=loc <3 @VP=bad !<4 __\n" +
"move bad >-1 loc\n" +
'\n') +
("@WHNP < (@WHADVP=bad < (WRB < /^(?i:how)$/) < (JJ < many|much)) < NN|NNS|NNP|NNPS\n" +
"relabel bad WHADJP\n" +
'\n') +
("@WHNP < (@NP=bad < /^(?:WP\\$|WDT|WRB)$/ $.. @PP|PRN|NP)\n" +
"relabel bad WHNP\n" +
'\n') +
("@WHNP < (@WHADVP=bad < (WRB < /^(?i:how)$/)) < (JJ < much)\n" +
"excise bad bad\n" +
'\n') +
("@WHNP < @NP < (@PP=bad < (@WHNP < WDT|WP))\n" +
"relabel bad WHPP\n" +
'\n') +
// Make a "What" before a noun a WDT (Santorini p.22)
// Except not when there's a DT, since I'm unsure of the "What a good feeling" case....
("@WHNP < (WP=bad $.. NN|NNS !$.. DT)\n" +
"relabel bad WDT\n" +
'\n') +
("@WHNP < @NP < (@PP=bad < (@NP=badder < WDT))\n" +
"relabel bad WHPP\n" +
"relabel badder WHNP\n" +
'\n') +
// ordered PP with noun directly in it - fix wrong size NP
("@PP <1 (IN|TO $. (@NP=place < ADJP)) <-1 NN|NNS|NNP|NNPS=word\n" +
"move word >-1 place\n" +
'\n') +
// ordered PP with noun directly in it - fix missing NP
("@PP=head <1 IN|TO=prep <-1 NN|NNS|NNP|NNPS\n" +
"adjoinH (PP NP@) head\n" +
"move prep >1 head\n" +
'\n') +
("@S < (@SBAR < (SBAR < SINV $. (CC $. (SBAR=adj < VBD < S))))\n" +
"adjoin (SBAR SINV@) adj\n" +
'\n') +
("@SINV < (NP-SBJ=subj $. (@VP=base $.. (@VP=say < (/^VB/ < say|says|said))))\n" +
"adjoinH (S VP@) base\n" +
"move subj >1 base\n" +
"relabel base S-1\n" +
"insert (S (-NONE- *T*-1)) >-1 say\n" +
'\n') +
("NP-SBJ=bad < NP-TMP !< NP < PP-TMP\n" +
"excise bad bad\n" +
'\n') +
("@S < NP-SBJ < (NP=bad < (/^NN/ < Tuesday|yesterday|Yesterday))\n" +
"relabel bad NP-TMP\n" +
'\n') +
("@S=bad < (VBP $. (NP-SBJ $. VP))\n" +
"relabel bad SINV\n" +
'\n') +
// these two rules show why an "insert above" operation would be nice....
("@S < (/^NP-SBJ/ $. (VBP|VBZ=aux $. @VP=adj))\n" +
"adjoinH (VP VP@) adj\n" +
"move aux >1 adj\n" +
'\n') +
("@S < (/^NP-SBJ/ $. (VBP|VBZ=aux $. @ADJP=adj))\n" +
"adjoinH (VP ADJP@) adj\n" +
"move aux >1 adj\n" +
'\n') +
("@PP=adj < (IN=prep $. JJ)\n" +
"adjoinH (PP NP@) adj\n" +
"move prep >1 adj\n" +
'\n') +
("@SBARQ <1 @WHADVP <2 VBZ=bad <3 RB=badder <4 @SQ=loc\n" +
"move badder >1 loc\n" +
"move bad >1 loc\n" +
'\n') +
("@SBARQ <2 @WHADVP <3 MD=bad <4 @SQ=loc\n" +
"move bad >1 loc\n" +
'\n') +
("@SBARQ <2 @WHNP <3 VBZ=bad <4 @SQ=loc\n" +
"move bad >1 loc\n" +
'\n') +
// tree in wsj_0755
("@VP=adj < (VBN=bad < been) < (JJ < unable) !< CC|CONJP\n" +
"adjoin (VP (VBN been) ADJP@) adj\n" +
"delete bad\n" +
'\n') +
// fix a bad tree in wsj_1623
("@S < (@NP < (NNS=bad < runs)) < (VP=home !< /^VB/ < (IN=badder < up))\n" +
"relabel bad VBZ\n" +
"relabel badder RP\n" +
"move bad >1 home\n" +
"adjoin (PRT RP@) badder\n" +
'\n') +
("@VP=top < (MD=bottom < will $. /^VB/)\n" +
"adjoin (VP (MD will) VP@) top\n" +
"delete bottom\n" +
'\n') +
("@VP < (MD=bad < /^'d$/)\n" +
"relabel bad VBD\n" +
'\n') +
("@S < (TO=bottom < to $. (@VP=top < /^VB/))\n" +
"adjoin (VP (TO to) VP@) top\n" +
"delete bottom\n" +
'\n') +
("@S < /^NP-SBJ/ < (VP < (VBD < " + DO + ") < (@NP=bad < (NN=badder < work)))\n" +
"relabel bad VP\n" +
"relabel badder VB\n" +
'\n') +
("@SBARQ !< @SQ !< /^-NONE-$/ !< @SBARQ < (@S|SINV=bad < VBP|VBZ|MD|VBD)\n" +
"relabel bad SQ\n" +
'\n') +
("@SBARQ < (@SINV=bad < (__ < would))\n" +
"relabel bad SQ\n" +
'\n') +
// NP over NP with nothing else.
("@NP=top < @NP=bottom !<2 __\n" +
"excise bottom bottom\n" +
'\n') +
("NP-SBJ=bad < (RB < Earlier)\n" +
"relabel bad NP-TMP\n" +
'\n') +
("@NP < (RB=bad < late|early $+ /^NN/)\n" +
"relabel bad JJ\n" +
'\n') +
// clear cases of NN(P) tagged as JJ after an indefinite article; unclear cases are left JJ
("@NP <1 (DT < a|an) <2 (JJ=bad < Hungary) !<3 __\n" +
"relabel bad NNP\n" +
'\n') +
// spelling of "deterrant" is how it appears in wrong tree!
("@NP <1 (DT < a|an) <2 (JJ=bad < /^(?:official|deterrant|bible|academic|fine|buy-out|perk|installment)$/) !<3 __\n" +
"relabel bad NN\n" +
'\n') +
// most complements of help take an S complement (ECM analysis),
// but some with null subjects don't, just a VP. Fix them so they do.
("@VP < (/^VB/ < help|helps|helped|helping|start|started|starts|starting|begin|begins|began|beginning) < (@VP=site < VB)\n" +
"adjoin (S (NP-SBJ (-NONE- *)) VP@) site\n" +
'\n') +
// This is in a NX-TTL, but just NNP is wrong on both counts so make NNS
("@NP < CD < (NNP < Drugs=bad)\n" +
"relabel bad NNS\n" +
'\n') +
// "well" that should be interjection, but done as (ADVP (RB well))
("well|Well|WELL [ , /^[:,]$/ | !, __ ] . /^[:,]$/ > (RB > ADVP)\n" +
"relabel bad UH\n" +
"relabel badder INTJ\n" +
'\n') +
// last minute; check more carefully
(
("NNP=bad < Securities|Manufacturers|Enterprises|Securities|Resources|Corporations|Sports|Merchants|Industries|Holdings|Brothers|Airlines|Systems|Motors|Industries|Parks|Communications|Facilities|Technologies|Sons|Publications|Products|Nations|Monopolies|Mergers|Machines|INDUSTRIES|Giants|Firearms|Associates|ASSOCIATES\n" +
"relabel bad NNPS\n" +
'\n') +
("@NP < (NN=bad < chief $++ /^NN/)\n" +
"relabel bad JJ\n" +
'\n') +
("@NP < (VBG=bad < operating|Operating $++ /^NN/)\n" +
"relabel bad NN\n" +
'\n') +
("@NP <- (DT=bad < half)\n" +
"relabel bad NN\n" +
'\n') +
("@ADVP < (RB=bad < easier|harder|earlier)\n" +
"relabel bad RBR\n" +
'\n') +
("@ADJP < (RB|JJ=bad < easier|harder|earlier)\n" +
"relabel bad JJR\n" +
'\n') +
("NN=bad < salespeople\n" +
"relabel bad NNS\n" +
'\n') +
("NN=bad < Chevrolet\n" +
"relabel bad NNP\n" +
'\n') +
("NNS=bad < Caltrans\n" +
"relabel bad NNP\n" +
'\n') +
("NNS=bad < Dirks\n" +
"relabel bad NNP\n" +
'\n') +
("NN=bad < /^U\\.K\\.$/\n" +
"relabel bad NNP\n" +
'\n') +
("@PP < (JJR=bad < more|less)\n" +
"relabel bad RBR\n" +
'\n') +
("@PP < (@ADVP < (JJR=bad < more|less))\n" +
"relabel bad RBR\n" +
'\n') +
("RB=bad < earlier\n" +
"relabel bad RBR\n" +
'\n') +
("@NP <- (NN < month|year) < (@QP < (RBR=bad < less|more) < (IN < than))\n" +
"relabel bad JJR\n" +
'\n') +
("@NP < (NNP=bad < Afghan $+ /^NN/)\n" +
"relabel bad JJ\n" +
'\n') +
("NNS=bad < headquarters\n" +
"relabel bad NN\n" +
'\n') +
("@NP < (NN=bad < managing)\n" +
"relabel bad VBG\n" +
'\n') +
("@NP < (NNS=x < East) < (NNS=y < Germans)\n" +
"relabel x NNPS\n" +
"relabel y NNPS\n" +
'\n') +
("@NP < DT < (NNS=y < Germans)\n" +
"relabel y NNPS\n" +
'\n') +
("@NP < (NN=y < /^S&P$/)\n" +
"relabel y NNP\n" +
'\n') +
("JJ=y < benchmark\n" +
"relabel y NN\n" +
'\n') +
("JJ=y < Aeroflot\n" +
"relabel y NNP\n" +
'\n') +
("NNPS=y < ADRs\n" +
"relabel y NNS\n" +
'\n') +
("NN=y < telecommunications\n" +
"relabel y NNS\n" +
'\n') +
("@NP < (NN=y < executive $+ (NN|NNS < officer|officers|vice))\n" +
"relabel y JJ\n" +
'\n') +
("@NP < (CD|NN=y < /^(?:'[0-9]0s|1[1-9][0-9]0s)$/)\n" +
"relabel y NNS\n" +
'\n') +
("IN|CC|NN|JJ=bad < /^(vs\\.|versus)$/\n" +
"relabel bad FW\n" +
'\n') +
("NN=bad < /^U\\.S\\.A\\.$/\n" +
"relabel bad NNP\n" +
'\n') +
"") +
"";
// not yet done
// 94 NP < (RB < much) -- many are determinatives (JJ?)
// lots of weird stuff in NP < RB !!
// Lots of PP < RP that should be consistentized somehow!
// Don Blaheta
// Markus Dickinson (does he cite Blaheta??)
// Ratnaparkhi tagger paper
// Singer et al. POS tagging paper
// Huddleston and Pullum analyze 'ago' as IN. Check Treebank guidelines.
// turn off markCC in definition of goodFactored. It just doesn't help.
// Naiwen Xue et al. for CTB discuss similar style of rules and tgrep error
// detection and correction phase.
}
// Salomon Brothers: is Brothers NNPS or NNP?
// Securities (146 NNP vs. 160 NNPS).