Search results ranking

Ranking overview

Ranking (aka weighting) of the search results can be defined as a process of computing a so-called relevance (aka weight) for every given matched document with regards to a given query that matched it. So relevance is in the end just a number attached to every document that estimates how relevant the document is to the query. Search results can then be sorted based on this number and/or some additional parameters, so that the most sought after results would come up higher on the results page.

There is no single standard one-size-fits-all way to rank any document in any scenario. Moreover, there can not ever be such a way, because relevance is subjective. As in, what seems relevant to you might not seem relevant to me. Hence, in general case it’s not just hard to compute, it’s theoretically impossible.

So ranking in Manticore is configurable. It has a notion of a so-called ranker. A ranker can formally be defined as a function that takes document and query as its input and produces a relevance value as output. In layman’s terms, a ranker controls exactly how (using which specific algorithm) will Manticore assign weights to the document.

Previously, this ranking function was rigidly bound to the matching mode. So in the legacy matching modes (that is, SPH_MATCH_ALL, SPH_MATCH_ANY, SPH_MATCH_PHRASE, and SPH_MATCH_BOOLEAN) you can not choose the ranker. You can only do that in the SPH_MATCH_EXTENDED mode. (Which is the only mode in SphinxQL and the suggested mode in SphinxAPI anyway.) To choose a non-default ranker you can either use SetRankingMode() with SphinxAPI, or OPTION ranker clause in SELECT statement when using SphinxQL.

As a sidenote, legacy matching modes are internally implemented via the unified syntax anyway. When you use one of those modes, Manticore just internally adjusts the query and sets the associated ranker, then executes the query using the very same unified code path.

Available built-in rankers

Manticore ships with a number of built-in rankers suited for different purposes. A number of them uses two factors, phrase proximity (aka LCS) and BM25. Phrase proximity works on the keyword positions, while BM25 works on the keyword frequencies. Basically, the better the degree of the phrase match between the document body and the query, the higher is the phrase proximity (it maxes out when the document contains the entire query as a verbatim quote). And BM25 is higher when the document contains more rare words. We’ll save the detailed discussion for later.

Currently implemented rankers are:

  • SPH_RANK_PROXIMITY_BM25, the default ranking mode that uses and combines both phrase proximity and BM25 ranking.
  • SPH_RANK_BM25, statistical ranking mode which uses BM25 ranking only (similar to most other full-text engines). This mode is faster but may result in worse quality on queries which contain more than 1 keyword.
  • SPH_RANK_NONE, no ranking mode. This mode is obviously the fastest. A weight of 1 is assigned to all matches. This is sometimes called boolean searching that just matches the documents but does not rank them.
  • SPH_RANK_WORDCOUNT, ranking by the keyword occurrences count. This ranker computes the per-field keyword occurrence counts, then multiplies them by field weights, and sums the resulting values.
  • SPH_RANK_PROXIMITY returns raw phrase proximity value as a result. This mode is internally used to emulate SPH_MATCH_ALL queries.
  • SPH_RANK_MATCHANY returns rank as it was computed in SPH_MATCH_ANY mode earlier, and is internally used to emulate SPH_MATCH_ANY queries.
  • SPH_RANK_FIELDMASK returns a 32-bit mask with N-th bit corresponding to N-th fulltext field, numbering from 0. The bit will only be set when the respective field has any keyword occurrences satisfying the query.
  • SPH_RANK_SPH04 is generally based on the default SPH_RANK_PROXIMITY_BM25 ranker, but additionally boosts the matches when they occur in the very beginning or the very end of a text field. Thus, if a field equals the exact query, SPH04 should rank it higher than a field that contains the exact query but is not equal to it. (For instance, when the query is “Hyde Park”, a document entitled “Hyde Park” should be ranked higher than a one entitled “Hyde Park, London” or “The Hyde Park Cafe”.)
  • SPH_RANK_EXPR lets you specify the ranking formula in run time. It exposes a number of internal text factors and lets you define how the final weight should be computed from those factors. You can find more details about its syntax and a reference available factors in a subsection below.

You should specify the SPH_RANK_ prefix and use capital letters only when using the SetRankingMode() call from the SphinxAPI. The API ports expose these as global constants. Using SphinxQL syntax, the prefix should be omitted and the ranker name is case insensitive. Example:

// SphinxAPI
$client->SetRankingMode ( SPH_RANK_SPH04 );

// SphinxQL
mysql_query ( "SELECT ... OPTION ranker=sph04" );

Legacy matching modes rankers

Legacy matching modes automatically select a ranker as follows:

  • SPH_MATCH_ALL uses SPH_RANK_PROXIMITY ranker;
  • SPH_MATCH_ANY uses SPH_RANK_MATCHANY ranker;
  • SPH_MATCH_PHRASE uses SPH_RANK_PROXIMITY ranker;
  • SPH_MATCH_BOOLEAN uses SPH_RANK_NONE ranker.

Quick summary of the ranking factors

Name Level Type Summary
max_lcs query int maximum possible LCS value for the current query
bm25 document int quick estimate of BM25(1.2, 0) without syntax support
bm25a(k1, b) document int precise BM25() value with configurable K1, B constants and syntax support
bm25f(k1, b, {field=weight, …}) document int precise BM25F() value with extra configurable field weights
field_mask document int bit mask of matched fields
query_word_count document int number of unique inclusive keywords in a query
doc_word_count document int number of unique keywords matched in the document
lcs field int Longest Common Subsequence between query and document, in words
user_weight field int user field weight
hit_count field int total number of keyword occurrences
word_count field int number of unique matched keywords
tf_idf field float sum(tf*idf) over matched keywords == sum(idf) over occurrences
min_hit_pos field int first matched occurrence position, in words, 1-based
min_best_span_pos field int first maximum LCS span position, in words, 1-based
exact_hit field bool whether query == field
min_idf field float min(idf) over matched keywords
max_idf field float max(idf) over matched keywords
sum_idf field float sum(idf) over matched keywords
exact_order field bool whether all query keywords were a) matched and b) in query order
min_gaps field int minimum number of gaps between the matched keywords over the matching spans
lccs field int Longest Common Contiguous Subsequence between query and document, in words
wlccs field float Weighted Longest Common Contiguous Subsequence, sum(idf) over contiguous keyword spans
atc field float Aggregate Term Closeness, log(1+sum(idf1*idf2*pow(distance, -1.75)) over the best pairs of keywords

Document-level ranking factors

A document-level factor is a numeric value computed by the ranking engine for every matched document with regards to the current query. So it differs from a plain document attribute in that the attribute do not depend on the full text query, while factors might. Those factors can be used anywhere in the ranking expression. Currently implemented document-level factors are:

  • bm25 (integer), a document-level BM25 estimate (computed without keyword occurrence filtering).
  • max_lcs (integer), a query-level maximum possible value that the sum(lcs*user_weight) expression can ever take. This can be useful for weight boost scaling. For instance, MATCHANY ranker formula uses this to guarantee that a full phrase match in any field ranks higher than any combination of partial matches in all fields.
  • field_mask (integer), a document-level 32-bit mask of matched fields.
  • query_word_count (integer), the number of unique keywords in a query, adjusted for a number of excluded keywords. For instance, both (one one one one) and (one !two) queries should assign a value of 1 to this factor, because there is just one unique non-excluded keyword.
  • doc_word_count (integer), the number of unique keywords matched in the entire document.

Field-level ranking factors

A field-level factor is a numeric value computed by the ranking engine for every matched in-document text field with regards to the current query. As more than one field can be matched by a query, but the final weight needs to be a single integer value, these values need to be folded into a single one. To achieve that, field-level factors can only be used within a field aggregation function, they can not be used anywhere in the expression. For example, you can not use (lcs+bm25) as your ranking expression, as lcs takes multiple values (one in every matched field). You should use (sum(lcs)+bm25) instead, that expression sums lcs over all matching fields, and then adds bm25 to that per-field sum. Currently implemented field-level factors are:

  • lcs (integer), the length of a maximum verbatim match between the document and the query, counted in words. LCS stands for Longest Common Subsequence (or Subset). Takes a minimum value of 1 when only stray keywords were matched in a field, and a maximum value of query keywords count when the entire query was matched in a field verbatim (in the exact query keywords order). For example, if the query is ‘hello world’ and the field contains these two words quoted from the query (that is, adjacent to each other, and exactly in the query order), lcs will be 2. For example, if the query is ‘hello world program’ and the field contains ‘hello world’, lcs will be 2. Note that any subset of the query keyword works, not just a subset of adjacent keywords. For example, if the query is ‘hello world program’ and the field contains ‘hello (test program)’, lcs will be 2 just as well, because both ‘hello’ and ‘program’ matched in the same respective positions as they were in the query. Finally, if the query is ‘hello world program’ and the field contains ‘hello world program’, lcs will be 3. (Hopefully that is unsurprising at this point.)

  • user_weight (integer), the user specified per-field weight (refer to SetFieldWeights() in SphinxAPI and OPTION field_weights in SphinxQL respectively). The weights default to 1 if not specified explicitly.

  • hit_count (integer), the number of keyword occurrences that matched in the field. Note that a single keyword may occur multiple times. For example, if ‘hello’ occurs 3 times in a field and ‘world’ occurs 5 times, hit_count will be 8.

  • word_count (integer), the number of unique keywords matched in the field. For example, if ‘hello’ and ‘world’ occur anywhere in a field, word_count will be 2, irregardless of how many times do both keywords occur.

  • tf_idf (float), the sum of TF/IDF over all the keywords matched in the field. IDF is the Inverse Document Frequency, a floating point value between 0 and 1 that describes how frequent is the keywords (basically, 0 for a keyword that occurs in every document indexed, and 1 for a unique keyword that occurs in just a single document). TF is the Term Frequency, the number of matched keyword occurrences in the field. As a side note, tf_idf is actually computed by summing IDF over all matched occurrences. That’s by construction equivalent to summing TF*IDF over all matched keywords.

  • min_hit_pos (integer), the position of the first matched keyword occurrence, counted in words. Indexing begins from position 1.

  • min_best_span_pos (integer), the position of the first maximum LCS occurrences span. For example, assume that our query was ‘hello world program’ and ‘hello world’ subphrase was matched twice in the field, in positions 13 and 21. Assume that ‘hello’ and ‘world’ additionally occurred elsewhere in the field, but never next to each other and thus never as a subphrase match. In that case, min_best_span_pos will be 13. Note how for the single keyword queries min_best_span_pos will always equal min_hit_pos.

  • exact_hit (boolean), whether a query was an exact match of the entire current field. Used in the SPH04 ranker.

  • min_idf, max_idf, and sum_idf (float). These factors respectively represent the min(idf), max(idf) and sum(idf) over all keywords that were matched in the field.

  • exact_order (boolean). Whether all of the query keywords were matched in the field in the exact query order. For example, (microsoft office) query would yield exact_order=1 in a field with the following contents: (We use Microsoft software in our office.). However, the very same query in a (Our office is Microsoft free.) field would yield exact_order=0.

  • min_gaps (integer), the minimum number of positional gaps between (just) the keywords matched in field. Always 0 when less than 2 keywords match; always greater or equal than 0 otherwise.

    For example, with a [big wolf] query, [big bad wolf] field would yield min_gaps=1; [big bad hairy wolf] field would yield min_gaps=2; [the wolf was scary and big] field would yield min_gaps=3; etc. However, a field like [i heard a wolf howl] would yield min_gaps=0, because only one keyword would be matching in that field, and, naturally, there would be no gaps between the _matched_keywords.

    Therefore, this is a rather low-level, “raw” factor that you would most likely want to adjust before actually using for ranking. Specific adjustments depend heavily on your data and the resulting formula, but here are a few ideas you can start with: (a) any min_gaps based boosts could be simply ignored when word_count<2; (b) non-trivial min_gaps values (i.e. when word_count>=2) could be clamped with a certain “worst case” constant while trivial values (i.e. when min_gaps=0 and word_count<2) could be replaced by that constant; (c) a transfer function like 1/(1+min_gaps) could be applied (so that better, smaller min_gaps values would maximize it and worse, bigger min_gaps values would fall off slowly); and so on.

  • lccs (integer). Longest Common Contiguous Subsequence. A length of the longest subphrase that is common between the query and the document, computed in keywords.

    LCCS factor is rather similar to LCS but more restrictive, in a sense. While LCS could be greater than 1 though no two query words are matched next to each other, LCCS would only get greater than 1 if there are exact, contiguous query subphrases in the document. For example, (one two three four five) query vs (one hundred three hundred five hundred) document would yield lcs=3, but lccs=1, because even though mutual dispositions of 3 keywords (one, three, five) match between the query and the document, no 2 matching positions are actually next to each other.

    Note that LCCS still does not differentiate between the frequent and rare keywords; for that, see WLCCS.

  • wlccs (float). Weighted Longest Common Contiguous Subsequence. A sum of IDFs of the keywords of the longest subphrase that is common between the query and the document.

    WLCCS is computed very similarly to LCCS, but every “suitable” keyword occurrence increases it by the keyword IDF rather than just by 1 (which is the case with LCS and LCCS). That lets us rank sequences of more rare and important keywords higher than sequences of frequent keywords, even if the latter are longer. For example, a query (Zanzibar bed and breakfast) would yield lccs=1 for a (hotels of Zanzibar) document, but lccs=3 against (London bed and breakfast), even though “Zanzibar” is actually somewhat more rare than the entire “bed and breakfast” phrase. WLCCS factor alleviates that problem by using the keyword frequencies.

  • atc (float). Aggregate Term Closeness. A proximity based measure that grows higher when the document contains more groups of more closely located and more important (rare) query keywords. WARNING: you should use ATC with OPTION idf=‘plain,tfidf_unnormalized’; otherwise you would get unexpected results.

    ATC basically works as follows. For every keyword occurrence in the document, we compute the so called term closeness. For that, we examine all the other closest occurrences of all the query keywords (keyword itself included too) to the left and to the right of the subject occurrence, compute a distance dampening coefficient as k = pow(distance, -1.75) for those occurrences, and sum the dampened IDFs. Thus for every occurrence of every keyword, we get a “closeness” value that describes the “neighbors” of that occurrence. We then multiply those per-occurrence closenesses by their respective subject keyword IDF, sum them all, and finally, compute a logarithm of that sum.

    Or in other words, we process the best (closest) matched keyword pairs in the document, and compute pairwise “closenesses” as the product of their IDFs scaled by the distance coefficient:

    pair_tc = idf(pair_word1) * idf(pair_word2) * pow(pair_distance, -1.75)
    

    We then sum such closenesses, and compute the final, log-dampened ATC value:

    atc = log(1+sum(pair_tc))
    

    Note that this final dampening logarithm is exactly the reason you should use OPTION idf=plain, because without it, the expression inside the log() could be negative.

    Having closer keyword occurrences actually contributes much more to ATC than having more frequent keywords. Indeed, when the keywords are right next to each other, distance=1 and k=1; when there just one word in between them, distance=2 and k=0.297, with two words between, distance=3 and k=0.146, and so on. At the same time IDF attenuates somewhat slower. For example, in a 1 million document collection, the IDF values for keywords that match in 10, 100, and 1000 documents would be respectively 0.833, 0.667, and 0.500. So a keyword pair with two rather rare keywords that occur in just 10 documents each but with 2 other words in between would yield pair_tc = 0.101 and thus just barely outweigh a pair with a 100-doc and a 1000-doc keyword with 1 other word between them and pair_tc = 0.099. Moreover, a pair of two unique, 1-doc keywords with 3 words between them would get a pair_tc = 0.088 and lose to a pair of two 1000-doc keywords located right next to each other and yielding a pair_tc = 0.25. So, basically, while ATC does combine both keyword frequency and proximity, it is still somewhat favoring the proximity.

Ranking factor aggregation functions

A field aggregation function is a single argument function that takes an expression with field-level factors, iterates it over all the matched fields, and computes the final results. Currently implemented field aggregation functions are:

  • sum, sums the argument expression over all matched fields. For instance, sum(1) should return a number of matched fields.
  • top, returns the greatest value of the argument over all matched fields.

Formula expressions for all the built-in rankers

Most of the other rankers can actually be emulated with the expression based ranker. You just need to pass a proper expression. Such emulation is, of course, going to be slower than using the built-in, compiled ranker but still might be of interest if you want to fine-tune your ranking formula starting with one of the existing ones. Also, the formulas define the nitty gritty ranker details in a nicely readable fashion.

  • SPH_RANK_PROXIMITY_BM25 = sum(lcs*user_weight)*1000+bm25
  • SPH_RANK_BM25 = bm25
  • SPH_RANK_NONE = 1
  • SPH_RANK_WORDCOUNT = sum(hit_count*user_weight)
  • SPH_RANK_PROXIMITY = sum(lcs*user_weight)
  • SPH_RANK_MATCHANY = sum((word_count+(lcs-1)*max_lcs)*user_weight)
  • SPH_RANK_FIELDMASK = field_mask
  • SPH_RANK_SPH04 = sum((4*lcs+2*(min_hit_pos==1)+exact_hit)*user_weight)*1000+bm25