Ejemplo n.º 1
0
  public static void main(String[] args) {
    Options op = new Options(new EnglishTreebankParserParams());
    // op.tlpParams may be changed to something else later, so don't use it till
    // after options are parsed.

    System.out.println("Currently " + new Date());
    System.out.print("Invoked with arguments:");
    for (String arg : args) {
      System.out.print(" " + arg);
    }
    System.out.println();

    String path = "/u/nlp/stuff/corpora/Treebank3/parsed/mrg/wsj";
    int trainLow = 200, trainHigh = 2199, testLow = 2200, testHigh = 2219;
    String serializeFile = null;

    int i = 0;
    while (i < args.length && args[i].startsWith("-")) {
      if (args[i].equalsIgnoreCase("-path") && (i + 1 < args.length)) {
        path = args[i + 1];
        i += 2;
      } else if (args[i].equalsIgnoreCase("-train") && (i + 2 < args.length)) {
        trainLow = Integer.parseInt(args[i + 1]);
        trainHigh = Integer.parseInt(args[i + 2]);
        i += 3;
      } else if (args[i].equalsIgnoreCase("-test") && (i + 2 < args.length)) {
        testLow = Integer.parseInt(args[i + 1]);
        testHigh = Integer.parseInt(args[i + 2]);
        i += 3;
      } else if (args[i].equalsIgnoreCase("-serialize") && (i + 1 < args.length)) {
        serializeFile = args[i + 1];
        i += 2;
      } else if (args[i].equalsIgnoreCase("-tLPP") && (i + 1 < args.length)) {
        try {
          op.tlpParams = (TreebankLangParserParams) Class.forName(args[i + 1]).newInstance();
        } catch (ClassNotFoundException e) {
          System.err.println("Class not found: " + args[i + 1]);
        } catch (InstantiationException e) {
          System.err.println("Couldn't instantiate: " + args[i + 1] + ": " + e.toString());
        } catch (IllegalAccessException e) {
          System.err.println("illegal access" + e);
        }
        i += 2;
      } else if (args[i].equals("-encoding")) {
        // sets encoding for TreebankLangParserParams
        op.tlpParams.setInputEncoding(args[i + 1]);
        op.tlpParams.setOutputEncoding(args[i + 1]);
        i += 2;
      } else {
        i = op.setOptionOrWarn(args, i);
      }
    }
    // System.out.println(tlpParams.getClass());
    TreebankLanguagePack tlp = op.tlpParams.treebankLanguagePack();

    Train.sisterSplitters = new HashSet(Arrays.asList(op.tlpParams.sisterSplitters()));
    //    BinarizerFactory.TreeAnnotator.setTreebankLang(tlpParams);
    PrintWriter pw = op.tlpParams.pw();

    Test.display();
    Train.display();
    op.display();
    op.tlpParams.display();

    // setup tree transforms
    Treebank trainTreebank = op.tlpParams.memoryTreebank();
    MemoryTreebank testTreebank = op.tlpParams.testMemoryTreebank();
    // Treebank blippTreebank = ((EnglishTreebankParserParams) tlpParams).diskTreebank();
    // String blippPath = "/afs/ir.stanford.edu/data/linguistic-data/BLLIP-WSJ/";
    // blippTreebank.loadPath(blippPath, "", true);

    Timing.startTime();
    System.err.print("Reading trees...");
    testTreebank.loadPath(path, new NumberRangeFileFilter(testLow, testHigh, true));
    if (Test.increasingLength) {
      Collections.sort(testTreebank, new TreeLengthComparator());
    }

    trainTreebank.loadPath(path, new NumberRangeFileFilter(trainLow, trainHigh, true));
    Timing.tick("done.");
    System.err.print("Binarizing trees...");
    TreeAnnotatorAndBinarizer binarizer = null;
    if (!Train.leftToRight) {
      binarizer =
          new TreeAnnotatorAndBinarizer(op.tlpParams, op.forceCNF, !Train.outsideFactor(), true);
    } else {
      binarizer =
          new TreeAnnotatorAndBinarizer(
              op.tlpParams.headFinder(),
              new LeftHeadFinder(),
              op.tlpParams,
              op.forceCNF,
              !Train.outsideFactor(),
              true);
    }
    CollinsPuncTransformer collinsPuncTransformer = null;
    if (Train.collinsPunc) {
      collinsPuncTransformer = new CollinsPuncTransformer(tlp);
    }
    TreeTransformer debinarizer = new Debinarizer(op.forceCNF);
    List<Tree> binaryTrainTrees = new ArrayList<Tree>();

    if (Train.selectiveSplit) {
      Train.splitters =
          ParentAnnotationStats.getSplitCategories(
              trainTreebank,
              Train.tagSelectiveSplit,
              0,
              Train.selectiveSplitCutOff,
              Train.tagSelectiveSplitCutOff,
              op.tlpParams.treebankLanguagePack());
      if (Train.deleteSplitters != null) {
        List<String> deleted = new ArrayList<String>();
        for (String del : Train.deleteSplitters) {
          String baseDel = tlp.basicCategory(del);
          boolean checkBasic = del.equals(baseDel);
          for (Iterator<String> it = Train.splitters.iterator(); it.hasNext(); ) {
            String elem = it.next();
            String baseElem = tlp.basicCategory(elem);
            boolean delStr = checkBasic && baseElem.equals(baseDel) || elem.equals(del);
            if (delStr) {
              it.remove();
              deleted.add(elem);
            }
          }
        }
        System.err.println("Removed from vertical splitters: " + deleted);
      }
    }
    if (Train.selectivePostSplit) {
      TreeTransformer myTransformer = new TreeAnnotator(op.tlpParams.headFinder(), op.tlpParams);
      Treebank annotatedTB = trainTreebank.transform(myTransformer);
      Train.postSplitters =
          ParentAnnotationStats.getSplitCategories(
              annotatedTB,
              true,
              0,
              Train.selectivePostSplitCutOff,
              Train.tagSelectivePostSplitCutOff,
              op.tlpParams.treebankLanguagePack());
    }

    if (Train.hSelSplit) {
      binarizer.setDoSelectiveSplit(false);
      for (Tree tree : trainTreebank) {
        if (Train.collinsPunc) {
          tree = collinsPuncTransformer.transformTree(tree);
        }
        // tree.pennPrint(tlpParams.pw());
        tree = binarizer.transformTree(tree);
        // binaryTrainTrees.add(tree);
      }
      binarizer.setDoSelectiveSplit(true);
    }
    for (Tree tree : trainTreebank) {
      if (Train.collinsPunc) {
        tree = collinsPuncTransformer.transformTree(tree);
      }
      tree = binarizer.transformTree(tree);
      binaryTrainTrees.add(tree);
    }
    if (Test.verbose) {
      binarizer.dumpStats();
    }

    List<Tree> binaryTestTrees = new ArrayList<Tree>();
    for (Tree tree : testTreebank) {
      if (Train.collinsPunc) {
        tree = collinsPuncTransformer.transformTree(tree);
      }
      tree = binarizer.transformTree(tree);
      binaryTestTrees.add(tree);
    }
    Timing.tick("done."); // binarization
    BinaryGrammar bg = null;
    UnaryGrammar ug = null;
    DependencyGrammar dg = null;
    // DependencyGrammar dgBLIPP = null;
    Lexicon lex = null;
    // extract grammars
    Extractor bgExtractor = new BinaryGrammarExtractor();
    // Extractor bgExtractor = new SmoothedBinaryGrammarExtractor();//new BinaryGrammarExtractor();
    // Extractor lexExtractor = new LexiconExtractor();

    // Extractor dgExtractor = new DependencyMemGrammarExtractor();

    Extractor dgExtractor = new MLEDependencyGrammarExtractor(op);
    if (op.doPCFG) {
      System.err.print("Extracting PCFG...");
      Pair bgug = null;
      if (Train.cheatPCFG) {
        List allTrees = new ArrayList(binaryTrainTrees);
        allTrees.addAll(binaryTestTrees);
        bgug = (Pair) bgExtractor.extract(allTrees);
      } else {
        bgug = (Pair) bgExtractor.extract(binaryTrainTrees);
      }
      bg = (BinaryGrammar) bgug.second;
      bg.splitRules();
      ug = (UnaryGrammar) bgug.first;
      ug.purgeRules();
      Timing.tick("done.");
    }
    System.err.print("Extracting Lexicon...");
    lex = op.tlpParams.lex(op.lexOptions);
    lex.train(binaryTrainTrees);
    Timing.tick("done.");

    if (op.doDep) {
      System.err.print("Extracting Dependencies...");
      binaryTrainTrees.clear();
      // dgBLIPP = (DependencyGrammar) dgExtractor.extract(new
      // ConcatenationIterator(trainTreebank.iterator(),blippTreebank.iterator()),new
      // TransformTreeDependency(tlpParams,true));

      DependencyGrammar dg1 =
          (DependencyGrammar)
              dgExtractor.extract(
                  trainTreebank.iterator(), new TransformTreeDependency(op.tlpParams, true));
      // dgBLIPP=(DependencyGrammar)dgExtractor.extract(blippTreebank.iterator(),new
      // TransformTreeDependency(tlpParams));

      // dg = (DependencyGrammar) dgExtractor.extract(new
      // ConcatenationIterator(trainTreebank.iterator(),blippTreebank.iterator()),new
      // TransformTreeDependency(tlpParams));
      // dg=new DependencyGrammarCombination(dg1,dgBLIPP,2);
      // dg = (DependencyGrammar) dgExtractor.extract(binaryTrainTrees); //uses information whether
      // the words are known or not, discards unknown words
      Timing.tick("done.");
      // System.out.print("Extracting Unknown Word Model...");
      // UnknownWordModel uwm = (UnknownWordModel)uwmExtractor.extract(binaryTrainTrees);
      // Timing.tick("done.");
      System.out.print("Tuning Dependency Model...");
      dg.tune(binaryTestTrees);
      // System.out.println("TUNE DEPS: "+tuneDeps);
      Timing.tick("done.");
    }

    BinaryGrammar boundBG = bg;
    UnaryGrammar boundUG = ug;

    GrammarProjection gp = new NullGrammarProjection(bg, ug);

    // serialization
    if (serializeFile != null) {
      System.err.print("Serializing parser...");
      LexicalizedParser.saveParserDataToSerialized(
          new ParserData(lex, bg, ug, dg, Numberer.getNumberers(), op), serializeFile);
      Timing.tick("done.");
    }

    // test: pcfg-parse and output

    ExhaustivePCFGParser parser = null;
    if (op.doPCFG) {
      parser = new ExhaustivePCFGParser(boundBG, boundUG, lex, op);
    }

    ExhaustiveDependencyParser dparser =
        ((op.doDep && !Test.useFastFactored) ? new ExhaustiveDependencyParser(dg, lex, op) : null);

    Scorer scorer = (op.doPCFG ? new TwinScorer(new ProjectionScorer(parser, gp), dparser) : null);
    // Scorer scorer = parser;
    BiLexPCFGParser bparser = null;
    if (op.doPCFG && op.doDep) {
      bparser =
          (Test.useN5)
              ? new BiLexPCFGParser.N5BiLexPCFGParser(
                  scorer, parser, dparser, bg, ug, dg, lex, op, gp)
              : new BiLexPCFGParser(scorer, parser, dparser, bg, ug, dg, lex, op, gp);
    }

    LabeledConstituentEval pcfgPE = new LabeledConstituentEval("pcfg  PE", true, tlp);
    LabeledConstituentEval comboPE = new LabeledConstituentEval("combo PE", true, tlp);
    AbstractEval pcfgCB = new LabeledConstituentEval.CBEval("pcfg  CB", true, tlp);

    AbstractEval pcfgTE = new AbstractEval.TaggingEval("pcfg  TE");
    AbstractEval comboTE = new AbstractEval.TaggingEval("combo TE");
    AbstractEval pcfgTEnoPunct = new AbstractEval.TaggingEval("pcfg nopunct TE");
    AbstractEval comboTEnoPunct = new AbstractEval.TaggingEval("combo nopunct TE");
    AbstractEval depTE = new AbstractEval.TaggingEval("depnd TE");

    AbstractEval depDE =
        new AbstractEval.DependencyEval("depnd DE", true, tlp.punctuationWordAcceptFilter());
    AbstractEval comboDE =
        new AbstractEval.DependencyEval("combo DE", true, tlp.punctuationWordAcceptFilter());

    if (Test.evalb) {
      EvalB.initEVALBfiles(op.tlpParams);
    }

    // int[] countByLength = new int[Test.maxLength+1];

    // use a reflection ruse, so one can run this without needing the tagger
    // edu.stanford.nlp.process.SentenceTagger tagger = (Test.preTag ? new
    // edu.stanford.nlp.process.SentenceTagger("/u/nlp/data/tagger.params/wsj0-21.holder") : null);
    SentenceProcessor tagger = null;
    if (Test.preTag) {
      try {
        Class[] argsClass = new Class[] {String.class};
        Object[] arguments =
            new Object[] {"/u/nlp/data/pos-tagger/wsj3t0-18-bidirectional/train-wsj-0-18.holder"};
        tagger =
            (SentenceProcessor)
                Class.forName("edu.stanford.nlp.tagger.maxent.MaxentTagger")
                    .getConstructor(argsClass)
                    .newInstance(arguments);
      } catch (Exception e) {
        System.err.println(e);
        System.err.println("Warning: No pretagging of sentences will be done.");
      }
    }

    for (int tNum = 0, ttSize = testTreebank.size(); tNum < ttSize; tNum++) {
      Tree tree = testTreebank.get(tNum);
      int testTreeLen = tree.yield().size();
      if (testTreeLen > Test.maxLength) {
        continue;
      }
      Tree binaryTree = binaryTestTrees.get(tNum);
      // countByLength[testTreeLen]++;
      System.out.println("-------------------------------------");
      System.out.println("Number: " + (tNum + 1));
      System.out.println("Length: " + testTreeLen);

      // tree.pennPrint(pw);
      // System.out.println("XXXX The binary tree is");
      // binaryTree.pennPrint(pw);
      // System.out.println("Here are the tags in the lexicon:");
      // System.out.println(lex.showTags());
      // System.out.println("Here's the tagnumberer:");
      // System.out.println(Numberer.getGlobalNumberer("tags").toString());

      long timeMil1 = System.currentTimeMillis();
      Timing.tick("Starting parse.");
      if (op.doPCFG) {
        // System.err.println(Test.forceTags);
        if (Test.forceTags) {
          if (tagger != null) {
            // System.out.println("Using a tagger to set tags");
            // System.out.println("Tagged sentence as: " +
            // tagger.processSentence(cutLast(wordify(binaryTree.yield()))).toString(false));
            parser.parse(addLast(tagger.processSentence(cutLast(wordify(binaryTree.yield())))));
          } else {
            // System.out.println("Forcing tags to match input.");
            parser.parse(cleanTags(binaryTree.taggedYield(), tlp));
          }
        } else {
          // System.out.println("XXXX Parsing " + binaryTree.yield());
          parser.parse(binaryTree.yield());
        }
        // Timing.tick("Done with pcfg phase.");
      }
      if (op.doDep) {
        dparser.parse(binaryTree.yield());
        // Timing.tick("Done with dependency phase.");
      }
      boolean bothPassed = false;
      if (op.doPCFG && op.doDep) {
        bothPassed = bparser.parse(binaryTree.yield());
        // Timing.tick("Done with combination phase.");
      }
      long timeMil2 = System.currentTimeMillis();
      long elapsed = timeMil2 - timeMil1;
      System.err.println("Time: " + ((int) (elapsed / 100)) / 10.00 + " sec.");
      // System.out.println("PCFG Best Parse:");
      Tree tree2b = null;
      Tree tree2 = null;
      // System.out.println("Got full best parse...");
      if (op.doPCFG) {
        tree2b = parser.getBestParse();
        tree2 = debinarizer.transformTree(tree2b);
      }
      // System.out.println("Debinarized parse...");
      // tree2.pennPrint();
      // System.out.println("DepG Best Parse:");
      Tree tree3 = null;
      Tree tree3db = null;
      if (op.doDep) {
        tree3 = dparser.getBestParse();
        // was: but wrong Tree tree3db = debinarizer.transformTree(tree2);
        tree3db = debinarizer.transformTree(tree3);
        tree3.pennPrint(pw);
      }
      // tree.pennPrint();
      // ((Tree)binaryTrainTrees.get(tNum)).pennPrint();
      // System.out.println("Combo Best Parse:");
      Tree tree4 = null;
      if (op.doPCFG && op.doDep) {
        try {
          tree4 = bparser.getBestParse();
          if (tree4 == null) {
            tree4 = tree2b;
          }
        } catch (NullPointerException e) {
          System.err.println("Blocked, using PCFG parse!");
          tree4 = tree2b;
        }
      }
      if (op.doPCFG && !bothPassed) {
        tree4 = tree2b;
      }
      // tree4.pennPrint();
      if (op.doDep) {
        depDE.evaluate(tree3, binaryTree, pw);
        depTE.evaluate(tree3db, tree, pw);
      }
      TreeTransformer tc = op.tlpParams.collinizer();
      TreeTransformer tcEvalb = op.tlpParams.collinizerEvalb();
      Tree tree4b = null;
      if (op.doPCFG) {
        // System.out.println("XXXX Best PCFG was: ");
        // tree2.pennPrint();
        // System.out.println("XXXX Transformed best PCFG is: ");
        // tc.transformTree(tree2).pennPrint();
        // System.out.println("True Best Parse:");
        // tree.pennPrint();
        // tc.transformTree(tree).pennPrint();
        pcfgPE.evaluate(tc.transformTree(tree2), tc.transformTree(tree), pw);
        pcfgCB.evaluate(tc.transformTree(tree2), tc.transformTree(tree), pw);
        if (op.doDep) {
          comboDE.evaluate((bothPassed ? tree4 : tree3), binaryTree, pw);
          tree4b = tree4;
          tree4 = debinarizer.transformTree(tree4);
          if (op.nodePrune) {
            NodePruner np = new NodePruner(parser, debinarizer);
            tree4 = np.prune(tree4);
          }
          // tree4.pennPrint();
          comboPE.evaluate(tc.transformTree(tree4), tc.transformTree(tree), pw);
        }
        // pcfgTE.evaluate(tree2, tree);
        pcfgTE.evaluate(tcEvalb.transformTree(tree2), tcEvalb.transformTree(tree), pw);
        pcfgTEnoPunct.evaluate(tc.transformTree(tree2), tc.transformTree(tree), pw);

        if (op.doDep) {
          comboTE.evaluate(tcEvalb.transformTree(tree4), tcEvalb.transformTree(tree), pw);
          comboTEnoPunct.evaluate(tc.transformTree(tree4), tc.transformTree(tree), pw);
        }
        System.out.println("PCFG only: " + parser.scoreBinarizedTree(tree2b, 0));

        // tc.transformTree(tree2).pennPrint();
        tree2.pennPrint(pw);

        if (op.doDep) {
          System.out.println("Combo: " + parser.scoreBinarizedTree(tree4b, 0));
          // tc.transformTree(tree4).pennPrint(pw);
          tree4.pennPrint(pw);
        }
        System.out.println("Correct:" + parser.scoreBinarizedTree(binaryTree, 0));
        /*
        if (parser.scoreBinarizedTree(tree2b,true) < parser.scoreBinarizedTree(binaryTree,true)) {
          System.out.println("SCORE INVERSION");
          parser.validateBinarizedTree(binaryTree,0);
        }
        */
        tree.pennPrint(pw);
      } // end if doPCFG

      if (Test.evalb) {
        if (op.doPCFG && op.doDep) {
          EvalB.writeEVALBline(tcEvalb.transformTree(tree), tcEvalb.transformTree(tree4));
        } else if (op.doPCFG) {
          EvalB.writeEVALBline(tcEvalb.transformTree(tree), tcEvalb.transformTree(tree2));
        } else if (op.doDep) {
          EvalB.writeEVALBline(tcEvalb.transformTree(tree), tcEvalb.transformTree(tree3db));
        }
      }
    } // end for each tree in test treebank

    if (Test.evalb) {
      EvalB.closeEVALBfiles();
    }

    // Test.display();
    if (op.doPCFG) {
      pcfgPE.display(false, pw);
      System.out.println("Grammar size: " + Numberer.getGlobalNumberer("states").total());
      pcfgCB.display(false, pw);
      if (op.doDep) {
        comboPE.display(false, pw);
      }
      pcfgTE.display(false, pw);
      pcfgTEnoPunct.display(false, pw);
      if (op.doDep) {
        comboTE.display(false, pw);
        comboTEnoPunct.display(false, pw);
      }
    }
    if (op.doDep) {
      depTE.display(false, pw);
      depDE.display(false, pw);
    }
    if (op.doPCFG && op.doDep) {
      comboDE.display(false, pw);
    }
    // pcfgPE.printGoodBad();
  }