What causes "PARSER RETURNED NULL!!!" in stanfordnlp/CoreNLP?

Check/raise parser options (e.g., maxLength / sentence length limits) and ensure fallback scoring for unseen characters

stanfordnlp/CoreNLP · error · RuntimeException

PARSER RETURNED NULL!!!

Error message

PARSER RETURNED NULL!!!

What it means

During evaluation/testing in ChineseCharacterBasedLexiconTraining.main, if parsing is enabled and lp.parseTree(sentence) returns null — meaning the parser failed to produce any parse for the sentence — the code throws RuntimeException("PARSER RETURNED NULL!!!") instead of silently skipping.

Solutions

Check/raise parser options (e.g., maxLength / sentence length limits) and ensure fallback scoring for unseen characters
Log the failing sentence and either skip it or record a parse failure instead of throwing, so a full test set can be evaluated
Verify the lexicon was trained/loaded correctly (unseen-character backoff present)
Catch the RuntimeException around the parse loop and continue with the next sentence

Example fix

// before
tree = lp.parseTree(s);
if (tree == null) { throw new RuntimeException("PARSER RETURNED NULL!!!"); }
// after
tree = lp.parseTree(s);
if (tree == null) {
  log.warning("Failed to parse sentence: " + s);
  failedSents++;
  continue;
}

Defensive patterns

Strategy: try-catch

Validate before calling

// skip sentences the parser cannot handle
if (s.size() > op.testOptions.maxLength) { log.warning("Sentence too long, skipping"); continue; }

Try / catch

try { tree = lp.parseTree(s); if (tree == null) { failed++; continue; } } catch (RuntimeException e) { failed++; log.warning("Parse failed: " + s); continue; }

Prevention

Train the lexicon with broad character coverage or add unseen-character backoff
Set realistic maxLength and test options
Collect parse failures per sentence rather than aborting the whole evaluation run

When it happens

Trigger: Testing with a character-based lexicon/parser on a sentence that fails to parse (e.g., no reachable analysis under the grammar, extremely long or degenerate input, corrupted lexicon yielding zero-probability everything).

Common situations: Feeding test sentences containing characters unseen in training so all analyses score -Inf; misconfigured lexicon/parser options; gold-sentence handling where sentence length exceeds parser limits.

Related errors

AI-assisted analysis of stanfordnlp/CoreNLP@1b7edd19c4 (2026-09-10). Data as JSON: /api/errors/6fb6bcde3a6da3e2. Report an issue: GitHub.

Appendix: source

Thrown at src/edu/stanford/nlp/parser/lexparser/ChineseCharacterBasedLexiconTraining.java:336

          log.info("Processing sentence; length: " + goldSentence.size());
        }
        List<HasWord> s;
        if (segmentWords) {
          StringBuilder goldCharBuf = new StringBuilder();
          for (HasWord aGoldSentence : goldSentence) {
            StringLabel word = (StringLabel) aGoldSentence;
            goldCharBuf.append(word.value());
          }
          String goldChars = goldCharBuf.toString();
          s = seg.segment(goldChars);
        } else {
          s = goldSentence;
        }
        Tree tree;
        if (parse) {
          tree = lp.parseTree(s);
          if (tree == null) {
            throw new RuntimeException("PARSER RETURNED NULL!!!");
          }
        } else {
          tree = Trees.toFlatTree(s);
          tree = subcategoryStripper.transformTree(tree);
        }

        if (pw != null) {
          if (parse) {
            tree.pennPrint(pw);
          } else {
            Iterator sentIter = s.iterator();
            for (; ;) {
              Word word = (Word) sentIter.next();
              pw.print(word.word());
              if (sentIter.hasNext()) {
                pw.print(" ");
              } else {
                break;

View on GitHub (pinned to 1b7edd19c4)