lucene实战--打分算法没有那么难？

2021-03-13 07:32

阅读：557

标签：模型 get from Lucene false val prot hive gic

准备工作

1.1 下载最新源码，https://github.com/apache/lucene-solr

1.2 编译，按照说明，使用ant进行编译(我使用了ant eclipse)

1.3.将编译后的文件导入到eclipse，sts或者idea中

2.新建测试类

public void test() throws IOException, ParseException {
        Analyzer analyzer = new NGramAnalyzer();

        // Store the index in memory:
        Directory directory = new RAMDirectory();
        // To store an index on disk, use this instead:
        //Path path = FileSystems.getDefault().getPath("E:\\demo\\data", "access.data");
        //Directory directory = FSDirectory.open(path);
        IndexWriterConfig config = new IndexWriterConfig(analyzer);
        IndexWriter iwriter = new IndexWriter(directory, config);
        Document doc = new Document();
        String text = "我是中国人.";
        doc.add(new Field("fieldname", text, TextField.TYPE_STORED));
        iwriter.addDocument(doc);
        iwriter.close();

        // Now search the index:
        DirectoryReader ireader = DirectoryReader.open(directory);
        IndexSearcher isearcher = new IndexSearcher(ireader);
        isearcher.setSimilarity(new BM25Similarity());
        // Parse a simple query that searches for "text":
        QueryParser parser = new QueryParser("fieldname", analyzer);
        Query query = parser.parse("中国,人");
        ScoreDoc[] hits = isearcher.search(query, 1000).scoreDocs;
        // Iterate through the results:
        for (int i = 0; i

其中，分词使用自定义的NGramAnalyzer，它继承自Analyzer，Analyzer分析文本，并将文本转换为TokenStream。详细如下：

/**
 * An Analyzer builds TokenStreams, which analyze text.  It thus represents a
 * policy for extracting index terms from text.
 * 
 * In order to define what analysis is done, subclasses must define their
 * {@link TokenStreamComponents TokenStreamComponents} in {@link #createComponents(String)}.
 * The components are then reused in each call to {@link #tokenStream(String, Reader)}.
 * 

 * Simple example:
 * 
 * Analyzer analyzer = new Analyzer() {
 *  {@literal @Override}
 *   protected TokenStreamComponents createComponents(String fieldName) {
 *     Tokenizer source = new FooTokenizer(reader);
 *     TokenStream filter = new FooFilter(source);
 *     filter = new BarFilter(filter);
 *     return new TokenStreamComponents(source, filter);
 *   }
 *   {@literal @Override}
 *   protected TokenStream normalize(TokenStream in) {
 *     // Assuming FooFilter is about normalization and BarFilter is about
 *     // stemming, only FooFilter should be applied
 *     return new FooFilter(in);
 *   }
 * };
 * 
 * For more examples, see the {@link org.apache.lucene.analysis Analysis package documentation}.
 * 
 * For some concrete implementations bundled with Lucene, look in the analysis modules:
 * 

 *   Common:
 *       Analyzers for indexing content in different languages and domains.
 *   
ICU:
 *       Exposes functionality from ICU to Apache Lucene. 
 *   
Kuromoji:
 *       Morphological analyzer for Japanese text.
 *   
Morfologik:
 *       Dictionary-driven lemmatization for the Polish language.
 *   
Phonetic:
 *       Analysis for indexing phonetic signatures (for sounds-alike search).
 *   
Smart Chinese:
 *       Analyzer for Simplified Chinese, which indexes words.
 *   
Stempel:
 *       Algorithmic Stemmer for the Polish Language.
 * 

 *
 * @since 3.1
 */

ClassicSimilarity是TFIDFSimilarity的封装，因TFIDFSimilarity是抽象方法，无法直接new出实例.这个算法是lucene早期的默认打分实现。

将测试类放入solr-lucene源码中，并进行debug，如果想要分析TFIDF算法，可以直接new ClassicSimilarity 然后放入IndexSearch，其它的类似。

3.算法介绍

新版的lucene使用了BM25Similarity作为默认打分实现。这里显式使用了BM25Similarity，算法详细。这里简要介绍一下：

技术图片

其中:

　　 D即文档(Document),Q即查询语句(Query),score(D,Q)指使用Q的查询语句在该文档下的打分函数。

　　IDF即倒排文件频次(Inverse Document Frequency)指在倒排文档中出现的次数，qi是Q分词后term

 ![](https://s4.51cto.com/images/blog/202011/29/c2041d2f7a39e25d7f18abe98f4b48af.png?x-oss-process=image/watermark,size_16,text_QDUxQ1RP5Y2a5a6i,color_FFFFFF,t_100,g_se,x_10,y_10,shadow_90,type_ZmFuZ3poZW5naGVpdGk=)其中，N是总的文档数目，n(qi)是出现分词qi的文档数目。

　　f(qi,D)是qi分词在文档Document出现的频次

　　 k1和b是可调参数，默认值为1.2，0.75

　　|D|是文档的单词的个数，avgdl 指库里的平均文档长度。

4.算法实现

1.IDF实现

　　单个IDF实现

  /** Implemented as log(1 + (docCount - docFreq + 0.5)/(docFreq + 0.5)). */
  protected float idf(long docFreq, long docCount) {
    return (float) Math.log(1 + (docCount - docFreq + 0.5D)/(docFreq + 0.5D));
  }

IDF的集合实现

@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);

    float[] oldCache = new float[256];
    float[] cache = new float[256];
    for (int i = 0; i 
   * 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 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);
  }

2.k1和b参数实现

public BM25Similarity(float k1, float b) {
    if (Float.isFinite(k1) == false || k1  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:
   * 
   *   {@code k1 = 1.2}
   *   {@code b = 0.75}
   * 
   */
  public BM25Similarity() {
    this(1.2f, 0.75f);
  }

3.平均文档长度avgdl 计算

/** The default implementation computes the average as sumTotalTermFreq / docCount */
  protected float avgFieldLength(CollectionStatistics collectionStats) {
    final long sumTotalTermFreq;
    if (collectionStats.sumTotalTermFreq() == -1) {
      // frequencies are omitted (tf=1), its # of postings
      if (collectionStats.sumDocFreq() == -1) {
        // theoretical case only: remove!
        return 1f;
      }
      sumTotalTermFreq = collectionStats.sumDocFreq();
    } else {
      sumTotalTermFreq = collectionStats.sumTotalTermFreq();
    }
    final long docCount = collectionStats.docCount() == -1 ? collectionStats.maxDoc() : collectionStats.docCount();
    return (float) (sumTotalTermFreq / (double) docCount);
  }

4.参数Weigh的计算

/** Cache of decoded bytes. */
  private static final float[] OLD_LENGTH_TABLE = new float[256];
  private static final float[] LENGTH_TABLE = new float[256];

  static {
    for (int i = 1; i

相当于技术图片

5.WeightValue计算


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

    BM25DocScorer(BM25Stats stats, int indexCreatedVersionMajor, NumericDocValues norms) throws IOException {
      this.stats = stats;
      this.weightValue = stats.weight * (k1 + 1);
      this.norms = norms;
      if (indexCreatedVersionMajor >= 7) {
        lengthCache = LENGTH_TABLE;
        cache = stats.cache;
      } else {
        lengthCache = OLD_LENGTH_TABLE;
        cache = stats.oldCache;
      }
    }

相当于
技术图片
红色部分相乘

6.总的得分计算

   @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);
    }

其中norm是从cache里取的，cache是放入了技术图片

那么整个公式就完整的出来了

7.深入

打分的数据来源于CollectionStatistics,TermStatistics及freq，那么它们是哪里得到的？

    SynonymWeight(Query query, IndexSearcher searcher, float boost) throws IOException {
      super(query);
      CollectionStatistics collectionStats = searcher.collectionStatistics(terms[0].field());//1
      long docFreq = 0;
      long totalTermFreq = 0;
      termContexts = new TermContext[terms.length];
      for (int i = 0; i

CollectionStatistics的来源

 /**
   * Returns {@link CollectionStatistics} for a field.
   * 
   * This can be overridden for example, to return a field‘s statistics
   * across a distributed collection.
   * @lucene.experimental
   */
  public CollectionStatistics collectionStatistics(String field) throws IOException {
    final int docCount;
    final long sumTotalTermFreq;
    final long sumDocFreq;

    assert field != null;

    Terms terms = MultiFields.getTerms(reader, field);
    if (terms == null) {
      docCount = 0;
      sumTotalTermFreq = 0;
      sumDocFreq = 0;
    } else {
      docCount = terms.getDocCount();
      sumTotalTermFreq = terms.getSumTotalTermFreq();
      sumDocFreq = terms.getSumDocFreq();
    }

    return new CollectionStatistics(field, reader.maxDoc(), docCount, sumTotalTermFreq, sumDocFreq);
  }

TermStatistics的来源

 /**
   * Returns {@link TermStatistics} for a term.
   * 
   * This can be overridden for example, to return a term‘s statistics
   * across a distributed collection.
   * @lucene.experimental
   */
  public TermStatistics termStatistics(Term term, TermContext context) throws IOException {
    return new TermStatistics(term.bytes(), context.docFreq(), context.totalTermFreq(),term.text());
  }

freq的来源(tf)


     @Override
    protected float score(DisiWrapper topList) throws IOException {
      return similarity.score(topList.doc, tf(topList));
    }

    /** combines TF of all subs. */
    final int tf(DisiWrapper topList) throws IOException {
      int tf = 0;
      for (DisiWrapper w = topList; w != null; w = w.next) {
        tf += ((TermScorer)w.scorer).freq();
      }
      return tf;
    }

底层实现

Lucene50PostingsReader.BlockPostingsEnum

 @Override
    public int nextDoc() throws IOException {
      if (docUpto == docFreq) {
        return doc = NO_MORE_DOCS;
      }
      if (docBufferUpto == BLOCK_SIZE) {
        refillDocs();
      }

      accum += docDeltaBuffer[docBufferUpto];
      freq = freqBuffer[docBufferUpto];
      posPendingCount += freq;
      docBufferUpto++;
      docUpto++;

      doc = accum;
      position = 0;
      return doc;
    }

8.总结

BM25算法的全称是 Okapi BM25，是一种二元独立模型的扩展，也可以用来做搜索的相关度排序。本文通过和lucene的BM25Similarity的实现来深入理解整个打分公式。

在此基础之上，又分析了CollectionStatistics,TermStatistics及freq这些参数是如何计算的。

通过整个分析过程，我们想要定制自己的打分公式，只需要实现Similarity或者SimilarityBase类，然后实现业务上的打分公式即可。