// algorithm: treat sentence snippets as miniature documents
// we can intersect these with the postings lists via BreakIterator.preceding(offset),s
// score each sentence as norm(sentenceStartOffset) * sum(weight * tf(freq))
private Passage[] highlightDoc(
String field,
BytesRef terms[],
int contentLength,
BreakIterator bi,
int doc,
TermsEnum termsEnum,
DocsAndPositionsEnum[] postings,
int n)
throws IOException {
PassageScorer scorer = getScorer(field);
if (scorer == null) {
throw new NullPointerException("PassageScorer cannot be null");
}
PriorityQueue<OffsetsEnum> pq = new PriorityQueue<>();
float weights[] = new float[terms.length];
// initialize postings
for (int i = 0; i < terms.length; i++) {
DocsAndPositionsEnum de = postings[i];
int pDoc;
if (de == EMPTY) {
continue;
} else if (de == null) {
postings[i] = EMPTY; // initially
if (!termsEnum.seekExact(terms[i])) {
continue; // term not found
}
de =
postings[i] = termsEnum.docsAndPositions(null, null, DocsAndPositionsEnum.FLAG_OFFSETS);
if (de == null) {
// no positions available
throw new IllegalArgumentException(
"field '" + field + "' was indexed without offsets, cannot highlight");
}
pDoc = de.advance(doc);
} else {
pDoc = de.docID();
if (pDoc < doc) {
pDoc = de.advance(doc);
}
}
if (doc == pDoc) {
weights[i] = scorer.weight(contentLength, de.freq());
de.nextPosition();
pq.add(new OffsetsEnum(de, i));
}
}
pq.add(new OffsetsEnum(EMPTY, Integer.MAX_VALUE)); // a sentinel for termination
PriorityQueue<Passage> passageQueue =
new PriorityQueue<>(
n,
new Comparator<Passage>() {
@Override
public int compare(Passage left, Passage right) {
if (left.score < right.score) {
return -1;
} else if (left.score > right.score) {
return 1;
} else {
return left.startOffset - right.startOffset;
}
}
});
Passage current = new Passage();
OffsetsEnum off;
while ((off = pq.poll()) != null) {
final DocsAndPositionsEnum dp = off.dp;
int start = dp.startOffset();
if (start == -1) {
throw new IllegalArgumentException(
"field '" + field + "' was indexed without offsets, cannot highlight");
}
int end = dp.endOffset();
// LUCENE-5166: this hit would span the content limit... however more valid
// hits may exist (they are sorted by start). so we pretend like we never
// saw this term, it won't cause a passage to be added to passageQueue or anything.
assert EMPTY.startOffset() == Integer.MAX_VALUE;
if (start < contentLength && end > contentLength) {
continue;
}
if (start >= current.endOffset) {
if (current.startOffset >= 0) {
// finalize current
current.score *= scorer.norm(current.startOffset);
// new sentence: first add 'current' to queue
if (passageQueue.size() == n && current.score < passageQueue.peek().score) {
current.reset(); // can't compete, just reset it
} else {
passageQueue.offer(current);
if (passageQueue.size() > n) {
current = passageQueue.poll();
current.reset();
} else {
current = new Passage();
}
}
}
// if we exceed limit, we are done
if (start >= contentLength) {
Passage passages[] = new Passage[passageQueue.size()];
passageQueue.toArray(passages);
for (Passage p : passages) {
p.sort();
}
// sort in ascending order
Arrays.sort(
passages,
new Comparator<Passage>() {
@Override
public int compare(Passage left, Passage right) {
return left.startOffset - right.startOffset;
}
});
return passages;
}
// advance breakiterator
assert BreakIterator.DONE < 0;
current.startOffset = Math.max(bi.preceding(start + 1), 0);
current.endOffset = Math.min(bi.next(), contentLength);
}
int tf = 0;
while (true) {
tf++;
BytesRef term = terms[off.id];
if (term == null) {
// multitermquery match, pull from payload
term = off.dp.getPayload();
assert term != null;
}
current.addMatch(start, end, term);
if (off.pos == dp.freq()) {
break; // removed from pq
} else {
off.pos++;
dp.nextPosition();
start = dp.startOffset();
end = dp.endOffset();
}
if (start >= current.endOffset || end > contentLength) {
pq.offer(off);
break;
}
}
current.score += weights[off.id] * scorer.tf(tf, current.endOffset - current.startOffset);
}
// Dead code but compiler disagrees:
assert false;
return null;
}
