Fix Laplace scorer to multiply by alpha (and not add) (#27125)

core/src/main/java/org/elasticsearch/search/suggest/phrase/CandidateScorer.java

@@ -93,7 +93,7 @@ final class CandidateScorer {
     private void updateTop(CandidateSet[] candidates, Candidate[] path, PriorityQueue<Correction> corrections, double cutoffScore, double score)
             throws IOException {
         score = Math.exp(score);
-        assert Math.abs(score - score(path, candidates)) < 0.00001;
+        assert Math.abs(score - score(path, candidates)) < 0.00001 : "cur_score=" + score + ", path_score=" + score(path,candidates);
         if (score > cutoffScore) {
             if (corrections.size() < maxNumCorrections) {
                 Candidate[] c = new Candidate[candidates.length];

core/src/main/java/org/elasticsearch/search/suggest/phrase/LaplaceScorer.java

@@ -38,10 +38,15 @@ final class LaplaceScorer extends WordScorer {
         return this.alpha;
     }
 
+    @Override
+    protected double scoreUnigram(Candidate word) throws IOException {
+        return (alpha + frequency(word.term)) / (vocabluarySize + alpha * numTerms);
+    }
+
     @Override
     protected double scoreBigram(Candidate word, Candidate w_1) throws IOException {
         join(separator, spare, w_1.term, word.term);
-        return (alpha + frequency(spare.get())) / (alpha +  w_1.frequency + vocabluarySize);
+        return (alpha + frequency(spare.get())) / (w_1.frequency + alpha * numTerms);
     }
 
     @Override
@@ -49,7 +54,7 @@ final class LaplaceScorer extends WordScorer {
         join(separator, spare, w_2.term, w_1.term, word.term);
         long trigramCount = frequency(spare.get());
         join(separator, spare, w_1.term, word.term);
-        return (alpha + trigramCount) / (alpha  +  frequency(spare.get()) + vocabluarySize);
+        return (alpha + trigramCount) / (frequency(spare.get()) + alpha * numTerms);
     }
 
 

core/src/main/java/org/elasticsearch/search/suggest/phrase/WordScorer.java

@@ -40,8 +40,8 @@ public abstract class WordScorer {
     protected final double realWordLikelyhood;
     protected final BytesRefBuilder spare = new BytesRefBuilder();
     protected final BytesRef separator;
+    protected final long numTerms;
     private final TermsEnum termsEnum;
-    private final long numTerms;
     private final boolean useTotalTermFreq;
 
     public WordScorer(IndexReader reader, String field, double realWordLikelyHood, BytesRef separator) throws IOException {
@@ -57,10 +57,11 @@ public abstract class WordScorer {
         final long vocSize = terms.getSumTotalTermFreq();
         this.vocabluarySize =  vocSize == -1 ? reader.maxDoc() : vocSize;
         this.useTotalTermFreq = vocSize != -1;
-        long numTerms = terms.size();
-        // -1 cannot be used as value, because scoreUnigram(...) can then divide by 0 if vocabluarySize is 1.
-        // -1 is returned when terms is a MultiTerms instance.
-        this.numTerms = vocabluarySize + numTerms > 1 ? numTerms : 0;
+        // terms.size() might be -1 if it's a MultiTerms instance. In that case,
+        // use reader.maxDoc() as an approximation. This also protects from
+        // division by zero, by scoreUnigram.
+        final long nTerms = terms.size();
+        this.numTerms = nTerms == -1 ? reader.maxDoc() : nTerms;
         this.termsEnum = new FreqTermsEnum(reader, field, !useTotalTermFreq, useTotalTermFreq, null, BigArrays.NON_RECYCLING_INSTANCE); // non recycling for now
         this.reader = reader;
         this.realWordLikelyhood = realWordLikelyHood;

docs/reference/search/suggesters/phrase-suggest.asciidoc

@@ -126,7 +126,7 @@ The response contains suggestions scored by the most likely spell correction fir
         "options" : [ {
           "text" : "nobel prize",
           "highlighted": "<em>nobel</em> prize",
-          "score" : 0.5962314
+          "score" : 0.48614594
         }]
       }
     ]