使用Standford coreNLP进行中文命名实体识别

因为工作需要，调研了一下Stanford coreNLP的命名实体识别功能。

Stanford CoreNLP是一个比较厉害的自然语言处理工具，很多模型都是基于深度学习方法训练得到的。

先附上其官网链接：

• https://stanfordnlp.github.io/CoreNLP/index.html
• https://nlp.stanford.edu/nlp/javadoc/javanlp/
• https://github.com/stanfordnlp/CoreNLP

本文主要讲解如何在java工程中使用Stanford CoreNLP；

1.环境准备

3.5之后的版本都需要java8以上的环境才能运行。需要进行中文处理的话，比较占用内存，3G左右的内存消耗。

笔者使用的maven进行依赖的引入，使用的是3.9.1版本。

直接在pom文件中加入下面的依赖：

        <dependency>
            <groupId>edu.stanford.nlp</groupId>
            <artifactId>stanford-corenlp</artifactId>
            <version>3.9.1</version>
        </dependency>
        <dependency>
            <groupId>edu.stanford.nlp</groupId>
            <artifactId>stanford-corenlp</artifactId>
            <version>3.9.1</version>
            <classifier>models</classifier>
        </dependency>
        <dependency>
            <groupId>edu.stanford.nlp</groupId>
            <artifactId>stanford-corenlp</artifactId>
            <version>3.9.1</version>
            <classifier>models-chinese</classifier>
        </dependency>

3个包分别是CoreNLP的算法包、英文语料包、中文预料包。这3个包的总大小为1.43G。maven默认镜像在国外，而这几个依赖包特别大，可以找有着三个依赖的国内镜像试一下。笔者用的是自己公司的maven仓库。

2.代码调用

需要注意的是，因为我是需要进行中文的命名实体识别，因此需要使用中文分词和中文的词典。我们可以先打开引入的jar包的结构：

其中有个StanfordCoreNLP-chinese.properties文件，这里面设定了进行中文自然语言处理的一些参数。主要指定相应的pipeline的操作步骤以及对应的预料文件的位置。实际上我们可能用不到所有的步骤，或者要使用不同的语料库，因此可以自定义配置文件，然后再引入。那在我的项目中，我就直接读取了该properties文件。

attention：此处笔者要使用的是ner功能，但可能不想使用其他的一些annotation，想去掉。然而，Stanford CoreNLP有一些局限，就是在ner执行之前，一定需要

tokenize, ssplit, pos, lemma

的引入，当然这增加了很大的时间耗时。

其实我们可以先来分析一下这个properties文件：

# Pipeline options - lemma is no-op for Chinese but currently needed because coref demands it (bad old requirements system)
annotators = tokenize, ssplit, pos, lemma, ner, parse, coref

# segment
tokenize.language = zh
segment.model = edu/stanford/nlp/models/segmenter/chinese/ctb.gz
segment.sighanCorporaDict = edu/stanford/nlp/models/segmenter/chinese
segment.serDictionary = edu/stanford/nlp/models/segmenter/chinese/dict-chris6.ser.gz
segment.sighanPostProcessing = true

# sentence split
ssplit.boundaryTokenRegex = [.。]|[!?！？]+

# pos
pos.model = edu/stanford/nlp/models/pos-tagger/chinese-distsim/chinese-distsim.tagger

# ner 此处设定了ner使用的语言、模型（crf），目前SUTime只支持英文，不支持中文，所以设置为false。
ner.language = chinese
ner.model = edu/stanford/nlp/models/ner/chinese.misc.distsim.crf.ser.gz
ner.applyNumericClassifiers = true
ner.useSUTime = false

# regexner
ner.fine.regexner.mapping = edu/stanford/nlp/models/kbp/chinese/cn_regexner_mapping.tab
ner.fine.regexner.noDefaultOverwriteLabels = CITY,COUNTRY,STATE_OR_PROVINCE

# parse
parse.model = edu/stanford/nlp/models/srparser/chineseSR.ser.gz

# depparse
depparse.model    = edu/stanford/nlp/models/parser/nndep/UD_Chinese.gz
depparse.language = chinese

# coref
coref.sieves = ChineseHeadMatch, ExactStringMatch, PreciseConstructs, StrictHeadMatch1, StrictHeadMatch2, StrictHeadMatch3, StrictHeadMatch4, PronounMatch
coref.input.type = raw
coref.postprocessing = true
coref.calculateFeatureImportance = false
coref.useConstituencyTree = true
coref.useSemantics = false
coref.algorithm = hybrid
coref.path.word2vec =
coref.language = zh
coref.defaultPronounAgreement = true
coref.zh.dict = edu/stanford/nlp/models/dcoref/zh-attributes.txt.gz
coref.print.md.log = false
coref.md.type = RULE
coref.md.liberalChineseMD = false

# kbp
kbp.semgrex = edu/stanford/nlp/models/kbp/chinese/semgrex
kbp.tokensregex = edu/stanford/nlp/models/kbp/chinese/tokensregex
kbp.language = zh
kbp.model = none

# entitylink
entitylink.wikidict = edu/stanford/nlp/models/kbp/chinese/wikidict_chinese.tsv.gz

那我们就直接在代码中引入这个properties文件，参考代码如下：

package com.baidu.corenlp;

import java.util.List;
import java.util.Map;
import java.util.Properties;

import edu.stanford.nlp.coref.CorefCoreAnnotations;
import edu.stanford.nlp.coref.data.CorefChain;
import edu.stanford.nlp.ling.CoreAnnotations;
import edu.stanford.nlp.ling.CoreLabel;
import edu.stanford.nlp.pipeline.Annotation;
import edu.stanford.nlp.pipeline.StanfordCoreNLP;
import edu.stanford.nlp.semgraph.SemanticGraph;
import edu.stanford.nlp.semgraph.SemanticGraphCoreAnnotations;
import edu.stanford.nlp.trees.Tree;
import edu.stanford.nlp.trees.TreeCoreAnnotations;
import edu.stanford.nlp.util.CoreMap;

/**
 * Created by sonofelice on 2018/3/27.
 */
public class TestNLP {
    public void test() throws Exception {
        //构造一个StanfordCoreNLP对象，配置NLP的功能，如lemma是词干化，ner是命名实体识别等
        Properties props = new Properties();
        props.load(this.getClass().getResourceAsStream("/StanfordCoreNLP-chinese.properties"));
        StanfordCoreNLP pipeline = new StanfordCoreNLP(props);
        String text = "袁隆平是中国科学院的院士,他于2009年10月到中国山东省东营市东营区永乐机场附近承包了一千亩盐碱地,"
                + "开始种植棉花, 年产量达到一万吨, 哈哈, 反正棣琦说的是假的,逗你玩儿,明天下午2点来我家吃饭吧。"
                + "棣琦是山东大学毕业的,目前在百度做java开发,位置是东北旺东路102号院,手机号14366778890";

        long startTime = System.currentTimeMillis();
        // 创造一个空的Annotation对象
        Annotation document = new Annotation(text);

        // 对文本进行分析
        pipeline.annotate(document);

        //获取文本处理结果
        List<CoreMap> sentences = document.get(CoreAnnotations.SentencesAnnotation.class);
        for (CoreMap sentence : sentences) {
            // traversing the words in the current sentence
            // a CoreLabel is a CoreMap with additional token-specific methods
            for (CoreLabel token : sentence.get(CoreAnnotations.TokensAnnotation.class)) {
                //                // 获取句子的token（可以是作为分词后的词语）
                String word = token.get(CoreAnnotations.TextAnnotation.class);
                System.out.println(word);
                //词性标注
                String pos = token.get(CoreAnnotations.PartOfSpeechAnnotation.class);
                System.out.println(pos);
                // 命名实体识别
                String ne = token.get(CoreAnnotations.NormalizedNamedEntityTagAnnotation.class);
                String ner = token.get(CoreAnnotations.NamedEntityTagAnnotation.class);
                System.out.println(word + " | analysis : {  original : " + ner + "," + " normalized : "
                        + ne + "}");
                //词干化处理
                String lema = token.get(CoreAnnotations.LemmaAnnotation.class);
                System.out.println(lema);
            }

            // 句子的解析树
            Tree tree = sentence.get(TreeCoreAnnotations.TreeAnnotation.class);
            System.out.println("句子的解析树:");
            tree.pennPrint();

            // 句子的依赖图
            SemanticGraph graph =
                    sentence.get(SemanticGraphCoreAnnotations.CollapsedCCProcessedDependenciesAnnotation.class);
            System.out.println("句子的依赖图");
            System.out.println(graph.toString(SemanticGraph.OutputFormat.LIST));

        }

        long endTime = System.currentTimeMillis();
        long time = endTime - startTime;
        System.out.println("The analysis lasts " + time + " seconds * 1000");

        // 指代词链
        //每条链保存指代的集合
        // 句子和偏移量都从1开始
        Map<Integer, CorefChain> corefChains = document.get(CorefCoreAnnotations.CorefChainAnnotation.class);
        if (corefChains == null) {
            return;
        }
        for (Map.Entry<Integer, CorefChain> entry : corefChains.entrySet()) {
            System.out.println("Chain " + entry.getKey() + " ");
            for (CorefChain.CorefMention m : entry.getValue().getMentionsInTextualOrder()) {
                // We need to subtract one since the indices count from 1 but the Lists start from 0
                List<CoreLabel> tokens = sentences.get(m.sentNum - 1).get(CoreAnnotations.TokensAnnotation.class);
                // We subtract two for end: one for 0-based indexing, and one because we want last token of mention 
                // not one following.
                System.out.println(
                        "  " + m + ", i.e., 0-based character offsets [" + tokens.get(m.startIndex - 1).beginPosition()
                                +
                                ", " + tokens.get(m.endIndex - 2).endPosition() + ")");
            }
        }
    }
}

public static void main(String[] args) throws  Exception {
    TestNLP nlp=new TestNLP();
    nlp.test();
}

 当然，我在运行过程中，只保留了ner相关的分析，别的功能注释掉了。输出结果如下：

19:46:16.000 [main] INFO  e.s.nlp.pipeline.StanfordCoreNLP - Adding annotator pos
19:46:19.387 [main] INFO  e.s.nlp.tagger.maxent.MaxentTagger - Loading POS tagger from edu/stanford/nlp/models/pos-tagger/chinese-distsim/chinese-distsim.tagger ... done [3.4 sec].
19:46:19.388 [main] INFO  e.s.nlp.pipeline.StanfordCoreNLP - Adding annotator lemma
19:46:19.389 [main] INFO  e.s.nlp.pipeline.StanfordCoreNLP - Adding annotator ner
19:46:21.938 [main] INFO  e.s.n.ie.AbstractSequenceClassifier - Loading classifier from edu/stanford/nlp/models/ner/chinese.misc.distsim.crf.ser.gz ... done [2.5 sec].
19:46:22.099 [main] WARN  e.s.n.p.TokensRegexNERAnnotator - TokensRegexNERAnnotator ner.fine.regexner: Entry has multiple types for ner: 巴伐利亚 STATE_OR_PROVINCE    MISC,GPE,LOCATION    1.  Taking type to be MISC
19:46:22.100 [main] WARN  e.s.n.p.TokensRegexNERAnnotator - TokensRegexNERAnnotator ner.fine.regexner: Entry has multiple types for ner: 巴伐利亚 州 STATE_OR_PROVINCE    MISC,GPE,LOCATION    1.  Taking type to be MISC
19:46:22.100 [main] INFO  e.s.n.p.TokensRegexNERAnnotator - TokensRegexNERAnnotator ner.fine.regexner: Read 21238 unique entries out of 21249 from edu/stanford/nlp/models/kbp/chinese/cn_regexner_mapping.tab, 0 TokensRegex patterns.
19:46:22.532 [main] INFO  e.s.nlp.pipeline.StanfordCoreNLP - Adding annotator parse
19:46:35.855 [main] INFO  e.s.nlp.parser.common.ParserGrammar - Loading parser from serialized file edu/stanford/nlp/models/srparser/chineseSR.ser.gz ... done [13.3 sec].
19:46:35.859 [main] INFO  e.s.nlp.pipeline.StanfordCoreNLP - Adding annotator coref
19:46:43.139 [main] INFO  e.s.n.pipeline.CorefMentionAnnotator - Using mention detector type: rule
19:46:43.148 [main] INFO  e.s.nlp.wordseg.ChineseDictionary - Loading Chinese dictionaries from 1 file:
19:46:43.148 [main] INFO  e.s.nlp.wordseg.ChineseDictionary -   edu/stanford/nlp/models/segmenter/chinese/dict-chris6.ser.gz
19:46:43.329 [main] INFO  e.s.nlp.wordseg.ChineseDictionary - Done. Unique words in ChineseDictionary is: 423200.
19:46:43.379 [main] INFO  edu.stanford.nlp.wordseg.CorpusChar - Loading character dictionary file from edu/stanford/nlp/models/segmenter/chinese/dict/character_list [done].
19:46:43.380 [main] INFO  e.s.nlp.wordseg.AffixDictionary - Loading affix dictionary from edu/stanford/nlp/models/segmenter/chinese/dict/in.ctb [done].
袁隆平 | analysis : {  original : PERSON, normalized : null}
是 | analysis : {  original : O, normalized : null}
中国 | analysis : {  original : ORGANIZATION, normalized : null}
科学院 | analysis : {  original : ORGANIZATION, normalized : null}
的 | analysis : {  original : O, normalized : null}
院士 | analysis : {  original : TITLE, normalized : null}
, | analysis : {  original : O, normalized : null}
他 | analysis : {  original : O, normalized : null}
于 | analysis : {  original : O, normalized : null}
2009年 | analysis : {  original : DATE, normalized : 2009-10-XX}
10月 | analysis : {  original : DATE, normalized : 2009-10-XX}
到 | analysis : {  original : O, normalized : null}
中国 | analysis : {  original : COUNTRY, normalized : null}
山东省 | analysis : {  original : STATE_OR_PROVINCE, normalized : null}
东营市 | analysis : {  original : CITY, normalized : null}
东营区 | analysis : {  original : FACILITY, normalized : null}
永乐 | analysis : {  original : FACILITY, normalized : null}
机场 | analysis : {  original : FACILITY, normalized : null}
附近 | analysis : {  original : O, normalized : null}
承包 | analysis : {  original : O, normalized : null}
了 | analysis : {  original : O, normalized : null}
一千 | analysis : {  original : NUMBER, normalized : 1000}
亩 | analysis : {  original : O, normalized : null}
盐 | analysis : {  original : O, normalized : null}
碱地 | analysis : {  original : O, normalized : null}
, | analysis : {  original : O, normalized : null}
开始 | analysis : {  original : O, normalized : null}
种植 | analysis : {  original : O, normalized : null}
棉花 | analysis : {  original : O, normalized : null}
, | analysis : {  original : O, normalized : null}
年产量 | analysis : {  original : O, normalized : null}
达到 | analysis : {  original : O, normalized : null}
一万 | analysis : {  original : NUMBER, normalized : 10000}
吨 | analysis : {  original : O, normalized : null}
, | analysis : {  original : O, normalized : null}
哈哈 | analysis : {  original : O, normalized : null}
, | analysis : {  original : O, normalized : null}
反正 | analysis : {  original : O, normalized : null}
棣琦 | analysis : {  original : PERSON, normalized : null}
说 | analysis : {  original : O, normalized : null}
的 | analysis : {  original : O, normalized : null}
是 | analysis : {  original : O, normalized : null}
假 | analysis : {  original : O, normalized : null}
的 | analysis : {  original : O, normalized : null}
, | analysis : {  original : O, normalized : null}
逗 | analysis : {  original : O, normalized : null}
你 | analysis : {  original : O, normalized : null}
玩儿 | analysis : {  original : O, normalized : null}
, | analysis : {  original : O, normalized : null}
明天 | analysis : {  original : DATE, normalized : XXXX-XX-XX}
下午 | analysis : {  original : TIME, normalized : null}
2点 | analysis : {  original : TIME, normalized : null}
来 | analysis : {  original : O, normalized : null}
我 | analysis : {  original : O, normalized : null}
家 | analysis : {  original : O, normalized : null}
吃饭 | analysis : {  original : O, normalized : null}
吧 | analysis : {  original : O, normalized : null}
。 | analysis : {  original : O, normalized : null}
棣琦 | analysis : {  original : PERSON, normalized : null}
是 | analysis : {  original : O, normalized : null}
山东 | analysis : {  original : ORGANIZATION, normalized : null}
大学 | analysis : {  original : ORGANIZATION, normalized : null}
毕业 | analysis : {  original : O, normalized : null}
的 | analysis : {  original : O, normalized : null}
, | analysis : {  original : O, normalized : null}
目前 | analysis : {  original : DATE, normalized : null}
在 | analysis : {  original : O, normalized : null}
百度 | analysis : {  original : ORGANIZATION, normalized : null}
做 | analysis : {  original : O, normalized : null}
java | analysis : {  original : O, normalized : null}
开发 | analysis : {  original : O, normalized : null}
, | analysis : {  original : O, normalized : null}
位置 | analysis : {  original : O, normalized : null}
是 | analysis : {  original : O, normalized : null}
东北 | analysis : {  original : LOCATION, normalized : null}
旺 | analysis : {  original : O, normalized : null}
东路 | analysis : {  original : O, normalized : null}
102 | analysis : {  original : NUMBER, normalized : 102}
号院 | analysis : {  original : O, normalized : null}
, | analysis : {  original : O, normalized : null}
手机号 | analysis : {  original : O, normalized : null}
143667788 | analysis : {  original : NUMBER, normalized : 14366778890}
90 | analysis : {  original : NUMBER, normalized : 14366778890}
The analysis lasts 819 seconds * 1000

Process finished with exit code 0

我们可以看到，整个工程的启动耗时还是挺久的。分析过程也比较耗时，819毫秒。

并且结果也不够准确，跟我在其官网在线demo得到的结果还是有些差异的：