最近做一些音乐类、读物类的自然语言理解,就调研使用了下Stanford corenlp,记录下来。
功能
Stanford Corenlp是一套自然语言分析工具集包括:
- POS(part of speech tagger)-标注词性
- NER(named entity recognizer)-实体名识别
- Parser树-分析句子的语法结构,如识别出短语词组、主谓宾等
- Coreference Resolution-指代消解,找出句子中代表同一个实体的词。下文的I/my,Nader/he表示的是同一个人
- Sentiment Analysis-情感分析
- Bootstrapped pattern learning-自展的模式学习(也不知道翻译对不对,大概就是可以无监督的提取一些模式,如提取实体名)
- Open IE(Information Extraction)-从纯文本中提取有结构关系组,如"Barack Obama was born in Hawaii" =》 (Barack Obama; was born in; Hawaii)
需求
语音交互类的应用(如语音助手、智能音箱echo)收到的通常是口语化的自然语言,如:我想听一个段子,给我来个牛郎织女的故事,要想精确的返回结果,就需要提出有用的主题词,段子/牛郎织女/故事。看了一圈就想使用下corenlp的TokensRegex,基于tokens序列的正则表达式。因为它提供的可用的工具有:正则表达式、分词、词性、实体类别,另外还可以自己指定实体类别,如指定牛郎织女是READ类别的实体。
Pattern语法
规则格式
{// ruleType is "text", "tokens", "composite", or "filter"ruleType: "tokens",//tokens是基于切词用于tokens正则,text是文本串用于文本正则,composite/filter还没搞明白// pattern to be matched pattern: ( ( [ { ner:PERSON } ]) /was/ /born/ /on/ ([ { ner:DATE } ]) ),// value associated with the expression for which the pattern was matched// matched expressions are returned with "DATE_OF_BIRTH" as the value// (as part of the MatchedExpression class)result: "DATE_OF_BIRTH"
}
除了上面的字段外还有action/name/stage/active/priority等,可以参考文后的文献。
ruleTypes是tokens,pattern中的基本元素是token,整体用(),1个token用[<expression>],1个expression用{tag:xx;ner:xx}来表述
ruleTypes是text,pattern就是常规的正则表达式,基本元素就是字符了,整体用//包围
实例
corenlp提供了单条/多条正则表达式的提取,本文就介绍从文件中加载规则来拦截我们需要的文本,并从中提取主题词。
依赖包
<dependency><groupId>edu.stanford.nlp</groupId><artifactId>stanford-corenlp</artifactId><version>3.4.1</version>
</dependency>
<dependency><groupId>edu.stanford.nlp</groupId><artifactId>stanford-corenlp</artifactId><version>3.4.1</version><classifier>models</classifier>
</dependency>
<!--中文支持-->
<dependency><groupId>edu.stanford.nlp</groupId><artifactId>stanford-corenlp</artifactId><version>3.6.0</version><classifier>models-chinese</classifier>
</dependency>
属性配置CoreNLP-chinese.properties(可以参考stanford-corenlp-models-chinese中的配置)
annotators = segment, ssplit, pos, ner, regexner, parse regexner.mapping = regexner.txt//自定义的实体正则表达式文件customAnnotatorClass.segment = edu.stanford.nlp.pipeline.ChineseSegmenterAnnotatorsegment.model = edu/stanford/nlp/models/segmenter/chinese/pku.gz segment.sighanCorporaDict = edu/stanford/nlp/models/segmenter/chinese segment.serDictionary = edu/stanford/nlp/models/segmenter/chinese/dict-chris6.ser.gz segment.sighanPostProcessing = truessplit.boundaryTokenRegex = [.]|[!?]+|[。]|[!?]+ //句子切分符pos.model = edu/stanford/nlp/models/pos-tagger/chinese-distsim/chinese-distsim.taggerner.model = edu/stanford/nlp/models/ner/chinese.misc.distsim.crf.ser.gz ner.applyNumericClassifiers = false ner.useSUTime = falseparse.model = edu/stanford/nlp/models/lexparser/chinesePCFG.ser.gz
corenlp中对文本的一次处理称为一个pipeline,annotators代表一个处理节点,如segment切词、ssplit句子切割(将一段话分为多个句子)、pos词性、ner实体命名、regexner是用自定义正则表达式来标注实体类型、parse是句子结构解析。后面就是各annotator的属性。
自定义的规则文件
regexner.txt(将'牛郎织女'的实体类别识别为READ)
牛郎织女 READ
rule.txt(tokensregex规则)
$TYPE="/笑话|故事|段子|口技|谜语|寓言|评书|相声|小品|唐诗|古诗|宋词|绕口令|故事|小说/ | /脑筋/ /急转弯/" //单类型 {ruleType: "tokens",pattern: ((?$type $TYPE)),result: Format("%s;%s;%s", "", $$type.text.replace(" ",""), "") }
(?type xx)代表一个命名group,提取该group将结果组装成xx;xx;xx形式返回
代码
//加载tokens正则表达
CoreMapExpressionExtractor extractor = CoreMapExpressionExtractor.createExtractorFromFile(TokenSequencePattern.getNewEnv(), "rule.txt");
//创建pipeline
StanfordCoreNLP coreNLP = new StanfordCoreNLP("CoreNLP-chinese.properties");
//处理文本
Annotation annotation = coreNLP.process("听个故事");
List<CoreMap> sentences = annotation.get(CoreAnnotations.SentencesAnnotation.class);
CoreMap sentence = sentences.get(0); //获得第一个句子分析结果
//过一遍tokens正则
List<MatchedExpression> matchedExpressions = extractor.extractExpressions(sentence);
for (MatchedExpression match : matchedExpressions) {System.out.println("Matched expression: " + match.getText() + " with value " + match.getValue());
}
想看下分析结果,如切词、词性、实体名,可以使用下面的函数
private void debug(CoreMap sentence) {// 从CoreMap中取出CoreLabel List,逐一打印出来List<CoreLabel> tokens = sentence.get(CoreAnnotations.TokensAnnotation.class);System.out.println("字/词" + "\t " + "词性" + "\t " + "实体标记");System.out.println("-----------------------------");for (CoreLabel token : tokens) {String word = token.getString(CoreAnnotations.TextAnnotation.class);String pos = token.getString(CoreAnnotations.PartOfSpeechAnnotation.class);String ner = token.getString(CoreAnnotations.NamedEntityTagAnnotation.class);System.out.println(word + "\t " + pos + "\t " + ner);}}
功能还是很强大的,毕竟可以用的东西多了,遇到问题时方法就多了。
参考文献
TokensRegex: http://nlp.stanford.edu/software/tokensregex.shtml
SequenceMatchRules: http://nlp.stanford.edu/nlp/javadoc/javanlp-3.5.0/edu/stanford/nlp/ling/tokensregex/SequenceMatchRules.html
Regexner: http://nlp.stanford.edu/software/regexner.html