之所以相关性比较困难,是因为搜索是一个严重的信息不对等的用户交互过程,所有的交互基本就限定在一个搜索框中,用户提供的搜索词也就寥寥几个,而搜索的数据往往是海量的,包括各种各样的类型和质量,用户的预期却是返回相关性非递增的搜索结果排序展示。

其中为了增加搜索的准确性,可以使用一些上下文信息来帮助搜索引擎。 例如用户搜索时候所在的页面,用户的偏好设置(语言,地理位置),以及累计的用户历史记录等等。

还有一些特定领域的方法,例如具有时效性的新闻,媒体类。

搜索结果的相关性不仅为用户提供了方便,还在潜移默化的影响着用户,甚至可以起到引导用户的作用,比如推荐内容。

Table of Contents

Effectiveness

how well does the system satisfy the user’s information need?

  • algorithm
  • interaction
  • evaluation

Algorithms

  • term importance (wordrank)
    which words are important when ranking a document (e.g. frequent vs. discriminative words)
  • stemming
    how to collapse words which are morphologically equivalent (e.g. bicycles -> bicycle)
  • query expansion (synonym)
    how to collapse words which are semantically equivalent (e.g. bicycles -> bicycle)
  • document structure
    do matches in different parts of the document matter (e.g. title vs. body match)?
  • personalization
    can we exploit user information to improve ranking?
  • relevance feedback
    ask a user which documents are relevant
  • disambiguation
    how to ask a user which words are important

Evaluation

  • relevance
    how to define a good document
  • metrics
    how to measure if the ranking is good
  • comparison
    how to compare two systems

Efficiency

how efficiently does the system satisfy the user’s information need?

  • indexing architectures
  • fast score computation
  • evaluation

Indexing architectures

  • parsing
    how should a document be split into a set of terms?
  • indexing
    which words should be kept?
  • weighting
    what information needs to be stored with terms?
  • compression
    how to compress the index size

Fast score computation

  • inverted indices
    fast retrieval and scoring of short queries
  • tiering
    can we tier retrieval and ranking to improve performance?
  • branch and bound
    how to efficiently prevent scoring unnecessary documents?

Modeling

information retrieval often involes formally modeling the retrieval process in order to optimize performance.

Modeling tasks

  • abstractly represent the documents
  • abstractly represent the queries
  • model the relationship between query and document representations

Boolean Retrieval Model

  • represent each document as an unweighted bag of words
  • represent the query as an unweighted bag of words
  • retrieva an unordered set of documents containing the query words

Simple Ranked Retrieval Model

  • represent each document as an weighted bag of words (based on document frequency)
  • represent the query as an unweighted bag of words
  • retrieve a ranking documents containing the query words

Modeling

  • much of the history of information retrieval effectiveness research involes developing new models or extending existing models
  • as modeling becomes more complicated, mathematics and statistics become necessary
  • new models still being developed

Vector Space Model

  • represent each document as a vector in a very high dimensional space
  • represent each query as a vector in the same high dimensional space
  • rank documents according to some vector similarity measure

Fundamentals

  • vector componenets
    what does each dimension represent?
  • text abstraction
    how to embed a query, document into this space?
  • vector similarity
    how to compare documents in this space?

Vector Componenets

  • Documents and queries should be represented as lineary independent basis vectors
    • Orthogonal ‘concepts’ such as topics or genres: ideal but difficult to define
    • Controlled keyword vocabulary: flexible and compact but difficult to maintain, may not be linearly independent
    • Free text vocabulary: easy to generate but grows quickly, definitely not linearly independent
  • In mose cases, when someone refers to the vector space model, each component of a vector represents the presence of a unique term in the entire corpus

Text Abstraction - Documents

  • Parse document in a vector of normalized strings
    • words are split on white space, removing punctuation (e.g. ‘Cants are lazy.’ -> [‘Cats’, ‘are’, ‘lazy’])
    • words are down-cased (e.g. -> [‘cats’, ‘are’, ‘lazy’])
    • stop words are removed (e.g. -> [‘cats’, ‘lazy’])
    • words are stemmed (e.g. -> [‘cat’, ‘lazy’])
  • Weight each word present in the document
    • Count the frequency of each word in the document (e.g. -> [<’cat’, 25>, <’lazy’, 25>, <’kitten’, 10>, …])
    • Re-weight according to each word’s discriminatory power; Very dependent on corpus
      • corpus about lazy animals:
        • [<’cat’, 189.70>, <’lazy’, 4.50>, <’kitten’, 120.31>, …]
      • corpus about cats:
        • [<’cat’, 0.12>, <’lazy’, 54.44>, <’kitten’, 5.43>, …]
    • Alternatively, can use binary weights (e.g. -> [<’cat’, 1>, <’lazy’, 1>, <’kitten’, 1>, <’dog’, 0>, …])

Text Discrimination Weights

  • Terms which appear in all or very many documents in the corpus may not be useful for indexing or retrieval
  • Stop words or ‘linguistic glue’ can be safely detected using corpus-independent lists
  • Other frequent terms may be domain-specific and are detected using corpus statistics
  • Inverse document frequency (IDF) summarizes this. For a corpus of n documents, the IDF of term t is defined as
\[IDF_t = 1 + log_2\frac{n}{df_t}\]

where df_t is the document frequency of t

Text Abstraction - Queries

  • Queries are processed using the exact same process
  • Because queries are treated as ‘short documents’, we can support keyword or document-length query-by-example style queries
  • Can be easily generalized to non-english languages

Vector Similarity

similarity binary weighted
inner product \(X \cap Y\) \(\sum_ix_iy_i\)
Dice coefficient \(2\frac{\vert X \cap Y \vert}{\vert X \vert+\vert Y \vert}\) \(2\frac{\sum_i{x_iy_i}}{\sum_i{x_i}+\sum_i{y_i}}\)
Cosine coefficient \(\frac{\vert X \cap Y \vert}{\sqrt{\vert X \vert}\sqrt{\vert Y \vert}}\) \(\frac{\sum_i{x_iy_i}}{\sqrt{\sum_ix_i^2}\sqrt{\sum_iy_i^2}}\)
Jaccard coefficient \(\frac{\vert X \cap Y \vert}{\vert X \vert + \vert Y \vert - \vert X \cap Y \vert}\) \(\frac{\sum_ix_iy_i}{\sum_ix_i^2+\sum_iy_i^2-\sum_ix_iy_i}\)

Vector Space Model - Efficiency

  • precompute length-normalized document vectors \(\frac{\sum_ix_iy_i}{\sqrt{\sum_ix_i^2}\sqrt{\sum_iy_i^2}} = \sum_i\frac{x_i}{\sqrt{\sum_ix_i^2}}\frac{y_i}{\sqrt{\sum_iy_i^2}} = \sum_i{x_i^~y_i^\~}\)
  • branch and bound to avoid scoring unnecessary documents
  • locality sensitive hashing can be used to do very fast approximate search (Indyk an Motwani 1998)

Latent Semantic Indexing

(Deerwester et al. 1990)

  • Vector space model suffers when there is a query term mismatch (e.g. ‘bike’ instead of ‘bicycle’)
  • By-product of not having independent basis vectors
  • Latent semantic indexing (LSI) attempts to resolve issues with correlated basis vectors
    • Use singular value decomposition (SVD) to find orthogonal basis vectors in the corpus
    • Project documents (and queries) into the lower dimensional ‘concept space’
    • Index and retrieve lower dimensional documents as with the classic vector space model
  • Similar to clustering documents in the corpus (i.e. k-means, PLSI, LDA)
  • In practice, very difficult to determine the appropriate number of lower dimensions (i.e. concepts)
  • Information retrieval needs to support retrieval at all granularities; clustering commits the system to a single granularity

Summary

  • Vector space model is a straightforward, easy to implement retrieval model
  • Principles underlie many modern commercial retrieval systems
  • However, there is more ranking than just term matches…

PageRank

Beyond Bag of Words

  • Document content, especially when using unordered bags of words has limited expressiveness
    • does not capture phrases
    • does not capture metadata
    • does not capture quality
  • Oftentimes we are intrested in how a document is consumed
    • do people like me think this is relevant?
    • does this document receive a lot of buzz?
    • is this document authoritative?

The Value of Credible Information

  • There is a lot of junk on the web (e.g. spam, irrelevant forums)
  • Knowing what users are reading is a valuable source for knowing what is not junk
  • Ideally, we would be able to monitor everything the user is reading and use that information for ranking; this is achieved through toolbars, browsers, operating systems, DNS.
  • In 1998, no search companies had browsing data. How did they address this lack of data?

Random Surfer Model

(Brin and Page 1998)

  • Simulate a very large number of users browsing the entire web
  • Let users browse randomly. This ia a naive assumption but works okay in practice
  • Observe how often pages get visited
  • The authoritativeness of a page is a function of its popularity in the simulation

Transition Matrix

  • The matrix, G defines a transition matrix over the web graph
  • In order to run the simulation, we take the matrix-vector product G x [1,1,1]^T
  • The result is a distribution over graph nodes representing where users would have gone have a single simulation step
  • We can run the simulation for an arbitrary number of steps, t, by taking powers of the matrix G^T x [1,1,1]^T
  • The result of this simulation is the PageRank score for every document in the graph

PageRank - Extensions

  • Build a non-random surfer (Meiss et al. 2010)
  • Personalized PageRank (Haveliwala 2003)
  • PageRank-directed crawling (Cho and Schonfeld 2007)
  • PageRank without links (Kurland and Lee 2005)

PageRank - Issues

  • Need a web graph stored in an efficient data structure
  • PageRank requires taking powers of an extemely large matrix
  • The PageRank is an approximation of visitation

Summary

  • At the time, PageRank provided a nice surrogate for real user data
  • Nowadays, search engines have access to toolbar data, click logs, GPS data, IM, email, social networks, …
  • Nonetheless, the random surfer model is important since the size of the web is much larger than even these data

Ranking in Practice

  • The vector space model measures text-based similarity
  • PageRank is claimed to be an important ranking feature, how does it compare?

Retrieval Metrics

  • Retrieval metrics attempt to quantify the quality of a ranked list
  • Usually assumed the top of the ranked list is more important that the bottom
  • Metrics used in this study,
    • Normalized Discounted Cumulative Gain (NDGG):
      • measures the amount of relevant information in the top of the ranking
      • importance of rank position drops quickly as you scroll down the list
    • Reciprocal Rank (RR)
      • measures the rank of the top relevant result
      • important for navigational queries
    • Mean Average Precision (MAP)
      • measures the amount of relevant information in the top of the ranking
      • decays slightly more slowly than NDGG

Ranking Signals

  • PageRank: static page ranking algorithm
  • HITS: static page ranking algorithm based on Kleinberg model (Kleinberg 1998)
  • FBM25F: a classic text-based retrieval function similar to the vector space model (Robertson et al. 2004)

Experiment

  • Corpus: 463 billion web pages
  • Relevance: 28k queries, with 17 relevance judgements each
  • Experiment: for each query, for each signal, rank documents, compute metric; compare average performance

Summary

  • Text-based ranking measures are necessary but not sufficient for high quality retrieval
  • Link-baesd ranking measures are important but subtle
  • Extemely import to confirm intuition with experiments

Precision & Recall

Precision

可以对不分词的查询给较高的boost

1
2
3
4
5
6
7
8
9
10
11
12
{
  "query": {
    "match": {
      "title": "",
      "boost": 5
    },
    "multi_match": {
      "query": "",
      "fields": ["title", "body"]
    }
  }
}

Recall

NLP

wordseg

{
  "text": "黑帆第二季第二集",
  "tokens": [
   {
      "token": "黑帆",
      "start_offset": 0,
      "end_offset": 2,
      "type": "<noun>",
      "position": 0
   },
   {
     "token": "第二季",
     "start_offset": 2,
     "end_offset": 5,
     "type": "<keyword>",
     "position": 1
   },
   {
     "token": "第二集",
     "start_offset": 5,
     "end_offset": 8,
     "type": "<keyword>",
     "position": 2
   }
  ]
}

synonym

{
  "text": "黑帆第二季第二集",
  "tokens": [
  {
    "token": "黑帆",
    "synonyms": ["black sails"]
  },
  {
    "token": "第二季",
    "synonyms": ["第2季", "S02", "S2"]
  },
  {
    "token": "第二集",
    "synonyms": ["第2集", "E02", "E2"]
  }
  ]
}

wordrank

{
  "text": "黑帆第二季第二集",
  "tokens": [
  {
    "token": "黑帆",
    "weight": 0.70
  },
  {
    "token": "第二季",
    "weight": 0.20
  },
  {
    "token": "第二集",
    "weight": 0.10
  }
  ]
}

ES

Control relevance

Similarity

Elasticsearch 5.0 has changed the default similarity from TF/IDF to Okapi BM25.

IDF: how popular is the term in the corpus

language is anbivalent, verbose and many topics in one document.

no clear way to formulate your query, no much information available but a few keywords.

normalization:

  • normalize
  • synonym

idf: common is less important, but will cause some less common word to be non-related.

bool query and coord-factor

TF/IDF is a heuristic that makes sense intuitively but it is somewhat a guess (Adhoc)

TF/IDF

BM25

termNorm

The root of BM25:

If retrieved documents are ordered by decreasing probability of relevance on the data available, then the system’s effectiveness is the best that can be obtained for the data.

Estimate relevancy:

  • simplification: relevance is binary (relevant or irrelevant)
  • get a dataset queries (relevant/irrelevant documents)
  • use that to estimate

the binary independence model

query term occurs in a document or doesn’t, don’t care how often. a dramatic but useful simplification

TF saturation curve

  • limits influence of TF
  • allows to tune influence by tweaking k1
  • less influence of common words
  • no more coord factor
  • disable corrd for bool queries
    • index.similarity.default.type = BM25
  • lower automatic boost for short fields
    • with TF/IDF, short fields are automatically scored higher
    • BM25: scales field length with average
  • field length treatment does not automatically boost short fields (you have to explicitly boost)
  • might need to adjust boost

https://github.com/apache/lucene-solr/blob/master/lucene/core/src/java/org/apache/lucene/search/similarities/BM25Similarity.java

/**
 * BM25 Similarity. Introduced in Stephen E. Robertson, Steve Walker,
 * Susan Jones, Micheline Hancock-Beaulieu, and Mike Gatford. Okapi at TREC-3.
 * In Proceedings of the Third <b>T</b>ext <b>RE</b>trieval <b>C</b>onference (TREC 1994).
 * Gaithersburg, USA, November 1994.
 */
public class BM25Similarity extends Similarity {
  private final float k1;
  private final float b;

  /**
   * BM25 with the supplied parameter values.
   * @param k1 Controls non-linear term frequency normalization (saturation).
   * @param b Controls to what degree document length normalizes tf values.
   * @throws IllegalArgumentException if {@code k1} is infinite or negative, or if {@code b} is 
   *         not within the range {@code [0..1]}
   */
  public BM25Similarity(float k1, float b) {
    if (Float.isFinite(k1) == false || k1 < 0) {
      throw new IllegalArgumentException("illegal k1 value: " + k1 + ", must be a non-negative finite value");
    }
    if (Float.isNaN(b) || b < 0 || b > 1) {
      throw new IllegalArgumentException("illegal b value: " + b + ", must be between 0 and 1");
    }
    this.k1 = k1;
    this.b  = b;
  }

  /** BM25 with these default values:
   * <ul>
   *   <li>{@code k1 = 1.2}</li>
   *   <li>{@code b = 0.75}</li>
   * </ul>
   */
  public BM25Similarity() {
    this(1.2f, 0.75f);
  }
 
  /** Implemented as <code>log(1 + (docCount - docFreq + 0.5)/(docFreq + 0.5))</code>. */
  protected float idf(long docFreq, long docCount) {
    return (float) Math.log(1 + (docCount - docFreq + 0.5D)/(docFreq + 0.5D));
  }
  
  /** Implemented as <code>1 / (distance + 1)</code>. */
  protected float sloppyFreq(int distance) {
    return 1.0f / (distance + 1);
  }
  
  /** The default implementation returns <code>1</code> */
  protected float scorePayload(int doc, int start, int end, BytesRef payload) {
    return 1;
  }
  
  /** The default implementation computes the average as <code>sumTotalTermFreq / docCount</code>,
   * or returns <code>1</code> if the index does not store sumTotalTermFreq:
   * any field that omits frequency information). */
  protected float avgFieldLength(CollectionStatistics collectionStats) {
    final long sumTotalTermFreq = collectionStats.sumTotalTermFreq();
    if (sumTotalTermFreq <= 0) {
      return 1f;       // field does not exist, or stat is unsupported
    } else {
      final long docCount = collectionStats.docCount() == -1 ? collectionStats.maxDoc() : collectionStats.docCount();
      return (float) (sumTotalTermFreq / (double) docCount);
    }
  }
  
  /** The default implementation encodes <code>boost / sqrt(length)</code>
   * with {@link SmallFloat#floatToByte315(float)}.  This is compatible with 
   * Lucene's default implementation.  If you change this, then you should 
   * change {@link #decodeNormValue(byte)} to match. */
  protected byte encodeNormValue(float boost, int fieldLength) {
    return SmallFloat.floatToByte315(boost / (float) Math.sqrt(fieldLength));
  }

  /** The default implementation returns <code>1 / f<sup>2</sup></code>
   * where <code>f</code> is {@link SmallFloat#byte315ToFloat(byte)}. */
  protected float decodeNormValue(byte b) {
    return NORM_TABLE[b & 0xFF];
  }
  
  /** 
   * True if overlap tokens (tokens with a position of increment of zero) are
   * discounted from the document's length.
   */
  protected boolean discountOverlaps = true;

  /** Sets whether overlap tokens (Tokens with 0 position increment) are 
   *  ignored when computing norm.  By default this is true, meaning overlap
   *  tokens do not count when computing norms. */
  public void setDiscountOverlaps(boolean v) {
    discountOverlaps = v;
  }

  /**
   * Returns true if overlap tokens are discounted from the document's length. 
   * @see #setDiscountOverlaps 
   */
  public boolean getDiscountOverlaps() {
    return discountOverlaps;
  }
  
  /** Cache of decoded bytes. */
  private static final float[] NORM_TABLE = new float[256];

  static {
    for (int i = 1; i < 256; i++) {
      float f = SmallFloat.byte315ToFloat((byte)i);
      NORM_TABLE[i] = 1.0f / (f*f);
    }
    NORM_TABLE[0] = 1.0f / NORM_TABLE[255]; // otherwise inf
  }


  @Override
  public final long computeNorm(FieldInvertState state) {
    final int numTerms = discountOverlaps ? state.getLength() - state.getNumOverlap() : state.getLength();
    return encodeNormValue(state.getBoost(), numTerms);
  }

  /**
   * Computes a score factor for a simple term and returns an explanation
   * for that score factor.
   * 
   * <p>
   * The default implementation uses:
   * 
   * <pre class="prettyprint">
   * idf(docFreq, docCount);
   * </pre>
   * 
   * Note that {@link CollectionStatistics#docCount()} is used instead of
   * {@link org.apache.lucene.index.IndexReader#numDocs() IndexReader#numDocs()} because also 
   * {@link TermStatistics#docFreq()} is used, and when the latter 
   * is inaccurate, so is {@link CollectionStatistics#docCount()}, and in the same direction.
   * In addition, {@link CollectionStatistics#docCount()} does not skew when fields are sparse.
   *   
   * @param collectionStats collection-level statistics
   * @param termStats term-level statistics for the term
   * @return an Explain object that includes both an idf score factor 
             and an explanation for the term.
   */
  public Explanation idfExplain(CollectionStatistics collectionStats, TermStatistics termStats) {
    final long df = termStats.docFreq();
    final long docCount = collectionStats.docCount() == -1 ? collectionStats.maxDoc() : collectionStats.docCount();
    final float idf = idf(df, docCount);
    return Explanation.match(idf, "idf, computed as log(1 + (docCount - docFreq + 0.5) / (docFreq + 0.5)) from:",
        Explanation.match(df, "docFreq"),
        Explanation.match(docCount, "docCount"));
  }

  /**
   * Computes a score factor for a phrase.
   * 
   * <p>
   * The default implementation sums the idf factor for
   * each term in the phrase.
   * 
   * @param collectionStats collection-level statistics
   * @param termStats term-level statistics for the terms in the phrase
   * @return an Explain object that includes both an idf 
   *         score factor for the phrase and an explanation 
   *         for each term.
   */
  public Explanation idfExplain(CollectionStatistics collectionStats, TermStatistics termStats[]) {
    double idf = 0d; // sum into a double before casting into a float
    List<Explanation> details = new ArrayList<>();
    for (final TermStatistics stat : termStats ) {
      Explanation idfExplain = idfExplain(collectionStats, stat);
      details.add(idfExplain);
      idf += idfExplain.getValue();
    }
    return Explanation.match((float) idf, "idf(), sum of:", details);
  }

  @Override
  public final SimWeight computeWeight(float boost, CollectionStatistics collectionStats, TermStatistics... termStats) {
    Explanation idf = termStats.length == 1 ? idfExplain(collectionStats, termStats[0]) : idfExplain(collectionStats, termStats);

    float avgdl = avgFieldLength(collectionStats);

    // compute freq-independent part of bm25 equation across all norm values
    float cache[] = new float[256];
    for (int i = 0; i < cache.length; i++) {
      cache[i] = k1 * ((1 - b) + b * decodeNormValue((byte)i) / avgdl);
    }
    return new BM25Stats(collectionStats.field(), boost, idf, avgdl, cache);
  }

  @Override
  public final SimScorer simScorer(SimWeight stats, LeafReaderContext context) throws IOException {
    BM25Stats bm25stats = (BM25Stats) stats;
    return new BM25DocScorer(bm25stats, context.reader().getNormValues(bm25stats.field));
  }
  
  private class BM25DocScorer extends SimScorer {
    private final BM25Stats stats;
    private final float weightValue; // boost * idf * (k1 + 1)
    private final NumericDocValues norms;
    private final float[] cache;
    
    BM25DocScorer(BM25Stats stats, NumericDocValues norms) throws IOException {
      this.stats = stats;
      this.weightValue = stats.weight * (k1 + 1);
      this.cache = stats.cache;
      this.norms = norms;
    }
    
    @Override
    public float score(int doc, float freq) throws IOException {
      // if there are no norms, we act as if b=0
      float norm;
      if (norms == null) {
        norm = k1;
      } else {
        if (norms.advanceExact(doc)) {
          norm = cache[(byte)norms.longValue() & 0xFF];
        } else {
          norm = cache[0];
        }
      }
      return weightValue * freq / (freq + norm);
    }
    
    @Override
    public Explanation explain(int doc, Explanation freq) throws IOException {
      return explainScore(doc, freq, stats, norms);
    }

    @Override
    public float computeSlopFactor(int distance) {
      return sloppyFreq(distance);
    }

    @Override
    public float computePayloadFactor(int doc, int start, int end, BytesRef payload) {
      return scorePayload(doc, start, end, payload);
    }
  }
  
  /** Collection statistics for the BM25 model. */
  private static class BM25Stats extends SimWeight {
    /** BM25's idf */
    private final Explanation idf;
    /** The average document length. */
    private final float avgdl;
    /** query boost */
    private final float boost;
    /** weight (idf * boost) */
    private final float weight;
    /** field name, for pulling norms */
    private final String field;
    /** precomputed norm[256] with k1 * ((1 - b) + b * dl / avgdl) */
    private final float cache[];

    BM25Stats(String field, float boost, Explanation idf, float avgdl, float cache[]) {
      this.field = field;
      this.boost = boost;
      this.idf = idf;
      this.avgdl = avgdl;
      this.cache = cache;
      this.weight = idf.getValue() * boost;
    }

  }

  private Explanation explainTFNorm(int doc, Explanation freq, BM25Stats stats, NumericDocValues norms) throws IOException {
    List<Explanation> subs = new ArrayList<>();
    subs.add(freq);
    subs.add(Explanation.match(k1, "parameter k1"));
    if (norms == null) {
      subs.add(Explanation.match(0, "parameter b (norms omitted for field)"));
      return Explanation.match(
          (freq.getValue() * (k1 + 1)) / (freq.getValue() + k1),
          "tfNorm, computed as (freq * (k1 + 1)) / (freq + k1) from:", subs);
    } else {
      byte norm;
      if (norms.advanceExact(doc)) {
        norm = (byte) norms.longValue();
      } else {
        norm = 0;
      }
      float doclen = decodeNormValue(norm);
      subs.add(Explanation.match(b, "parameter b"));
      subs.add(Explanation.match(stats.avgdl, "avgFieldLength"));
      subs.add(Explanation.match(doclen, "fieldLength"));
      return Explanation.match(
          (freq.getValue() * (k1 + 1)) / (freq.getValue() + k1 * (1 - b + b * doclen/stats.avgdl)),
          "tfNorm, computed as (freq * (k1 + 1)) / (freq + k1 * (1 - b + b * fieldLength / avgFieldLength)) from:", subs);
    }
  }

  private Explanation explainScore(int doc, Explanation freq, BM25Stats stats, NumericDocValues norms) throws IOException {
    Explanation boostExpl = Explanation.match(stats.boost, "boost");
    List<Explanation> subs = new ArrayList<>();
    if (boostExpl.getValue() != 1.0f)
      subs.add(boostExpl);
    subs.add(stats.idf);
    Explanation tfNormExpl = explainTFNorm(doc, freq, stats, norms);
    subs.add(tfNormExpl);
    return Explanation.match(
        boostExpl.getValue() * stats.idf.getValue() * tfNormExpl.getValue(),
        "score(doc="+doc+",freq="+freq+"), product of:", subs);
  }

  @Override
  public String toString() {
    return "BM25(k1=" + k1 + ",b=" + b + ")";
  }
  
  /** 
   * Returns the <code>k1</code> parameter
   * @see #BM25Similarity(float, float) 
   */
  public final float getK1() {
    return k1;
  }
  
  /**
   * Returns the <code>b</code> parameter 
   * @see #BM25Similarity(float, float) 
   */
  public final float getB() {
    return b;
  }
}

actor 字段:

  • analyzed
  • not_analyzed

actor, director 来构造词典

cutoff_frequency

  • 提升性能
  • 降低高频词干扰

dis_max

取最大得分,可以使用tie_breaker来综合其它匹配的影响。

"minimum_should_match": "1<50%"

可以用来过滤

title:

  • exact match
    • stopword 丢失
    • 低频词降权
    • lengthNorm
  • analyzed
  • shingles

文本相关性评分:

  • bool
    • actor
    • dis_max
      • title
      • alias

Boosting

  • Lucene 是通用搜索引擎,需要使用boost来完成specific-domain适配
  • Adding domain-specific logic
  • Incorporating additional signals
    • recency/freshness boost
    • boosting by popularity (more comments or views)
  • Filter boost
    • boost articles which has comments or reviews or audited or from official, eliminate function score overhead
    • filter by user preferences
      • content type
      • categories of interest
      • source of interest

Boost should be a multiplier on the relevancy score.

索引文档类型不同可能导致termFreq差异较大,此时可以给因文本较长导致较高的文档类型一个boost(需要注意的是Lucene可以通过lengthNorm自适应,所以boost不宜太大)

Boosting with Recency/Freshness

Time Decay

  • freshness boost
  • staleness penalty
if updatedTime == null:
    updatedTime = createdTime

Bottom out the old age penalty using min(recip(...), 0.20)

  • Solr recip
    • y = a/(m*x + b)
    • a = 0.08, b = 0.05, m = 3.16E-11
  • Newtow’s Law of Cooling

Boosting with Popularity

view count should be broken into time slots:

  • Past
  • Last Month
  • Last Week
  • Recent

view count can be derived from log analysis:

log analysis -> Map/Reduce -> view count

minimum popularity is 1 (not zero), up to 2 or more

1 + (0.4*Recent + 0.3*LastWeek + 0.2*LastMonth ...)

Testing

Focus Group

In-house Testing

A/B Test