Using Lucene's highlighting, getting too much highlighted, is there a workaround for this?

Question

I am using the highlighting feature of Lucene to isolate matching terms for my query, but some of the matched terms are excessive.

I have some simple test cases which are delivered in an Ant project (download details below).

Materials

You can download the test case here: mydemo_with_libs.zip

That archive includes the Lucene 8.6.3 libraries which my test uses; if you prefer a copy without the JAR files you can download that from here: mydemo_without_libs.zip

The necessary libraries are: core, analyzers, queries, queryparser, highlighter, and memory.

You can run the test case by unzipping the archive into an empty directory and running the Ant command ant synsearch

Input

I have provided a short synonym list which is used for indexing and analysing in the highlighting methods:

cope,manage
jobs,tasks
simultaneously,at once

and there is one document being indexed:

Queues are a useful way of grouping jobs together in order to manage a number of them at once. You can:
hold or release multiple jobs at the same time;
group multiple tasks (for the same event);
control the priority of jobs in the queue;
Eventually log all events that take place in a queue.
Use either job.queue or task.queue in specifications.

Process

When building the index I am storing the text field, and using a custom analyzer. This is because (in the real world) the content I am indexing is technical documentation, so stripping out punctuation is inappropriate because so much of it may be significant in technical expressions. My analyzer uses a TechTokenFilter which breaks the stream up into tokens consisting of strings of words or digits, or individual characters which don't match the previous pattern.

Here's the relevant code for the analyzer:

public class MyAnalyzer extends Analyzer {
   public MyAnalyzer(String synlist) {
      if (synlist != "") {
         this.synlist = synlist;
         this.useSynonyms = true;
      }
   }
   public MyAnalyzer() {
      this.useSynonyms = false;
   }

   @Override
   protected TokenStreamComponents createComponents(String fieldName) {
      WhitespaceTokenizer src = new WhitespaceTokenizer();
      TokenStream result = new TechTokenFilter(new LowerCaseFilter(src));
      if (useSynonyms) {
         result = new SynonymGraphFilter(result, getSynonyms(synlist), Boolean.TRUE);
         result = new FlattenGraphFilter(result);
      }
      return new TokenStreamComponents(src, result);
   }

and here's my filter:

public class TechTokenFilter extends TokenFilter {
   private final CharTermAttribute termAttr;
   private final PositionIncrementAttribute posIncAttr;
   private final ArrayList<String> termStack;
   private AttributeSource.State current;
   private final TypeAttribute typeAttr;

   public TechTokenFilter(TokenStream tokenStream) {
      super(tokenStream);
      termStack = new ArrayList<>();
      termAttr = addAttribute(CharTermAttribute.class);
      posIncAttr = addAttribute(PositionIncrementAttribute.class);
      typeAttr = addAttribute(TypeAttribute.class);
   }

   @Override
   public boolean incrementToken() throws IOException {

      if (this.termStack.isEmpty() && input.incrementToken()) {
         final String currentTerm = termAttr.toString();
         final int bufferLen = termAttr.length();

         if (bufferLen > 0) {
            if (termStack.isEmpty()) {
               termStack.addAll(Arrays.asList(techTokens(currentTerm)));
               current = captureState();
            }
         }
      }

      if (!this.termStack.isEmpty()) {

         String part = termStack.remove(0);
         restoreState(current);
         termAttr.setEmpty().append(part);
         posIncAttr.setPositionIncrement(1);

         return true;
      } else {
         return false;
      }
   }

   public static String[] techTokens(String t) {
      List<String> tokenlist = new ArrayList<String>();
      String[] tokens;
      StringBuilder next = new StringBuilder();
      String token;
      char minus = '-';
      char underscore = '_';
      char c, prec, subc;
      // Boolean inWord = false;

      for (int i = 0; i < t.length(); i++) {
         prec = i > 0 ? t.charAt(i - 1) : 0;
         c = t.charAt(i);
         subc = i < (t.length() - 1) ? t.charAt(i + 1) : 0;
         if (Character.isLetterOrDigit(c) || c == underscore) {
            next.append(c);
            // inWord = true;
         }
         else if (c == minus && Character.isLetterOrDigit(prec) && Character.isLetterOrDigit(subc)) {
            next.append(c);
         } else {
            if (next.length() > 0) {
               token = next.toString();
               tokenlist.add(token);
               next.setLength(0);
            }
            if (Character.isWhitespace(c)) {
               // shouldn't be possible because the input stream has been tokenized on
               // whitespace
            } else {
               tokenlist.add(String.valueOf(c));
            }
            // inWord = false;
         }
      }
      if (next.length() > 0) {
         token = next.toString();
         tokenlist.add(token);
         // next.setLength(0);
      }
      tokens = tokenlist.toArray(new String[0]);
      return tokens;
   }
}

Examining the index I can see that the index contains the separate terms I expect, including the synonym values. For example the text at the end of the first line has produced the terms

of
them
at , simultaneously
once
.
You
can
:

and the text at the end of the third line has produced the terms

same
event
)
;

When the application performs a search it analyzes the query without using the synonym list (because the synonyms are already in the index), but I have discovered that I need to include the synonym list when analyzing the stored text to identify the matching fragments.

Searches match the correct documents, but the code I have added to identify the matching terms over-performs. I won't show all the search method here, but will focus on the code which lists matched terms:

public static void doSearch(IndexReader reader, IndexSearcher searcher,
                               Query query, int max, String synList) throws IOException {

   SimpleHTMLFormatter htmlFormatter = new SimpleHTMLFormatter("\001", "\002");
   Highlighter highlighter = new Highlighter(htmlFormatter, new QueryScorer(query));
   Analyzer analyzer;
   if (synList != null) {
      analyzer = new MyAnalyzer(synList);
   } else {
      analyzer = new MyAnalyzer();
   }

   // Collect all the docs
   TopDocs results = searcher.search(query, max);
   ScoreDoc[] hits = results.scoreDocs;

   int numTotalHits = Math.toIntExact(results.totalHits.value);
   System.out.println("\nQuery: " + query.toString());
   System.out.println("Matches: " + numTotalHits);

   // Collect matching terms
   HashSet<String> matchedWords = new HashSet<String>();

   int start = 0;
   int end = Math.min(numTotalHits, max);

   for (int i = start; i < end; i++) {
      int id = hits[i].doc;
      float score = hits[i].score;
      Document doc = searcher.doc(id);
      String docpath = doc.get("path");
      String doctext = doc.get("text");
      try {
         TokenStream tokens = TokenSources.getTokenStream("text", null, doctext, analyzer, -1);
         TextFragment[] frag = highlighter.getBestTextFragments(tokens, doctext, false, 100);
         for (int j = 0; j < frag.length; j++) {
            if ((frag[j] != null) && (frag[j].getScore() > 0)) {
               String match = frag[j].toString();
               addMatchedWord(matchedWords, match);
            }
         }
      } catch (InvalidTokenOffsetsException e) {
         System.err.println(e.getMessage());
      }
      System.out.println("matched file: " + docpath);
   }

   if (matchedWords.size() > 0) {
      System.out.println("matched terms:");
      for (String word : matchedWords) {
         System.out.println(word);
      }
   }
}

Problem

While the correct documents are selected by these queries, and the fragments chosen for highlighting do contain the query terms, the highlighted pieces in some of the selected fragments extend over too much of the input.

For example, if the query is

+text:event +text:manage

(the first example in the test case) then I would expect to see 'event' and 'manage' in the highlighted list. But what I actually see is

event);
manage

Despite the highlighting process using an analyzer which breaks terms apart and treats punctuation characters as single terms, the highlight code is "hungry" and breaks on whitespace alone.

Similarly if the query is

+text:queeu~1

(my final test case) I would expect to only see 'queue' in the list. But I get

queue.
job.queue
task.queue
queue;

It is so nearly there... but I don't understand why the highlighted pieces are inconsistent with the index, and I don't think I should have to parse the list of matches through yet another filter to produce the correct list of matches.

I would really appreciate any pointers to what I am doing wrong or how I could improve my code to deliver exactly what I need.

Thanks for reading this far!

Answer 1

I managed to get this working by replacing the WhitespaceTokenizer and TechTokenFilter in my analyzer with a PatternTokenizer; the regular expression took a bit of work but once I had it all the matching terms were extracted with pinpoint accuracy.

The replacement analyzer:

public class MyAnalyzer extends Analyzer {
   public MyAnalyzer(String synlist) {
      if (synlist != "") {
         this.synlist = synlist;
         this.useSynonyms = true;
      }
   }
   public MyAnalyzer() {
      this.useSynonyms = false;
   }

   private static final String tokenRegex = "(([\\w]+-)*[\\w]+)|[^\\w\\s]";

   @Override
   protected TokenStreamComponents createComponents(String fieldName) {
      PatternTokenizer src = new PatternTokenizer(Pattern.compile(tokenRegex), 0);
      TokenStream result = new LowerCaseFilter(src);
      if (useSynonyms) {
         result = new SynonymGraphFilter(result, getSynonyms(synlist), Boolean.TRUE);
         result = new FlattenGraphFilter(result);
      }
      return new TokenStreamComponents(src, result);
   }