斯坦福 NLP NER、Sentiment、SUTime 性能问题

Question

main 方法中的文本似乎需要超过 2 秒才能返回 NER。我不是 NLP 方面的专家，而且这段代码根本无法扩展。我在已确定瓶颈的 2 个地方添加了 cmets。您能否提出改进建议以提高程序的性能。

谢谢。

 public class NERSentimentUtil
{
private static final Logger logger = Logger.getLogger(NERSentimentUtil.class);

private static final String serializedClassifier7 = "edu/stanford/nlp/models/ner/english.muc.7class.distsim.crf.ser.gz";
private static final String serializedClassifier4 = "edu/stanford/nlp/models/ner/english.conll.4class.distsim.crf.ser.gz";
private static final String serializedClassifier3 = "edu/stanford/nlp/models/ner/english.all.3class.distsim.crf.ser.gz";

private static NERClassifierCombiner ncc;
private static StanfordCoreNLP pipeline;

static
{       
    try
    {
        ncc = new NERClassifierCombiner(serializedClassifier3,serializedClassifier4,serializedClassifier7);
    } catch (IOException e) {
        e.printStackTrace();
        logger.error(e);
    }
}

static
{               
    Properties props = new Properties();
    props.setProperty("annotators", "tokenize, ssplit, parse, sentiment, sutime");
    /*props.setProperty("ner.useSUTime", "0");*/

    String defs_sutime = "/edu/stanford/nlp/models/sutime/defs.sutime.txt";
    String holiday_sutime = "/edu/stanford/nlp/models/sutime/english.holidays.sutime.txt";
    String _sutime = "/edu/stanford/nlp/models/sutime/english.sutime.txt";

    String sutimeRules = defs_sutime + "," + holiday_sutime + "," + _sutime;
    props.setProperty("ner.useSUTime", "true");
    props.setProperty("-sutime.rules", sutimeRules);
    props.setProperty("sutime.binders", "0");
    props.setProperty("sutime.markTimeRanges", "false");
    props.setProperty("sutime.includeRange", "false");
    props.setProperty("customAnnotatorClass.sutime", "edu.stanford.nlp.time.TimeAnnotator");
    props.setProperty("parse.maxlen", "20");
    //props.setProperty("ner.applyNumericClassifiers", "false");
    //props.setProperty("nthreads", "16");
    //props.setProperty("threads", "16");
    //props.setProperty("parse.nthreads","16");
    //props.setProperty("ssplit.eolonly","true");

    props.setProperty("-parse.model", "edu/stanford/nlp/models/srparser/englishSR.ser.gz");
    RedwoodConfiguration.current().clear().apply();
    pipeline = new StanfordCoreNLP(props);
    //RedwoodConfiguration.empty().capture(System.err).apply();
}

//A sentiment score of 0 or 1 is negative, 2 neutral and 3 or 4 positive.
private static int getScore(int score)
{
    if(score<2)
        return -1;
    else if(score==2)
        return 0;
    else
        return 1;       
}

public static HashMap<String,Object> getStanford(String s, long dateString)//"2013-07-14"
{   
    int finalScore =0;

    HashMap<String,Object> map = new HashMap<String,Object>();

    HashMap<String, Integer> dateMap = new HashMap<String, Integer>();
    HashMap<String, Integer> dateCountMap = new HashMap<String, Integer>();
    HashMap<String, String> dateSentenceMap = new HashMap<String, String>();

    HashMap<String, Integer> personMap = new HashMap<String, Integer>();
    HashMap<String, Integer> personCountMap = new HashMap<String, Integer>();

    HashMap<String, Integer> orgMap = new HashMap<String, Integer>();
    HashMap<String, Integer> orgCountMap = new HashMap<String, Integer>();

    HashMap<String, Integer> locationMap = new HashMap<String, Integer>();
    HashMap<String, Integer> locationCountMap = new HashMap<String, Integer>();

    HashMap<String, Article_Location> locationArticleMap = new HashMap<String, Article_Location>();

    ArrayList<Articel_Ner> organisationlist = new ArrayList<Articel_Ner>();
    ArrayList<Articel_Ner> personlist = new ArrayList<Articel_Ner>();
    ArrayList<Artilcle_Ner_Date> datelist = new ArrayList<Artilcle_Ner_Date>();
    ArrayList<Article_NerLocation> locationList = new ArrayList<Article_NerLocation>();     

    try
    {
        Annotation annotation = pipeline.process(s);//1/3 rd time is taken up by this line

        List<CoreMap> sentences = annotation.get(CoreAnnotations.SentencesAnnotation.class);

        for (CoreMap sentence : sentences)
        {
             String str = sentence.toString();

             int score = getSentiment(sentence);

             finalScore+=score;
             boolean dFlag = true;

             List<Triple<String,Integer,Integer>> triples = ncc.classifyToCharacterOffsets(str);

             for (Triple<String,Integer,Integer> trip : triples)
             {
                 String ne = trip.first();
                 String word = str.substring(trip.second(), trip.third).toLowerCase();

                 switch(ne)
                 {
                    case "LOCATION":                        
                        extractLocation(locationMap, locationCountMap, locationArticleMap, score, word);
                        break;

                    case "ORGANIZATION":                        
                        extractOrg(orgMap, orgCountMap, score, word);                       
                        break;

                    case "PERSON":                      
                        extractPerson(personMap, personCountMap, score, word);
                        break;

                    case "DATE":
                        if(dFlag)
                        {
                         extractSUDate(dateString, dateMap, dateCountMap, dateSentenceMap, str, score);
                         dFlag = false;
                        }
                        break;

                    default:
                        break;
                 }
             }
        }
            //2/3rd of the time taken by these 4 methods:: can be obtimized
        mapDate(dateMap, dateCountMap, dateSentenceMap, datelist);
        mapLocation(locationMap, locationCountMap, locationArticleMap, locationList);   
        mapOrg(orgMap, orgCountMap, organisationlist);  
        mapPerson(personMap, personCountMap, personlist);
        //
    }
    catch(Exception e)
    {
        logger.error(e);
        logger.error(s);
        e.printStackTrace();
    }

    if(finalScore>0)
        finalScore = 1;
    else if(finalScore<0)
        finalScore = -1;
    else
        finalScore = 0;

    map.put("ORGANISATION", organisationlist);
    map.put("PERSON", personlist);
    map.put("DATE", datelist);
    map.put("LOCATION", locationList);
    map.put("SENTIMENT", finalScore);

    return map;
}

private static void extractPerson(HashMap<String, Integer> personMap, HashMap<String, Integer> personCountMap,
        int score, String word)
{       
    if(personMap.get(word)!=null)
    {
        personMap.put(word, personMap.get(word)+score);
        personCountMap.put(word, personCountMap.get(word)+1);
    }
    else
    {
        personMap.put(word, score);
        personCountMap.put(word, 1);
        //personSentenceMap.put(pname, str);
    }   
}

private static void extractOrg(HashMap<String, Integer> orgMap, HashMap<String, Integer> orgCountMap,
        int score, String word)
{
    if(orgMap.get(word)!=null)
    {
        orgMap.put(word, orgMap.get(word)+score);
        orgCountMap.put(word, orgCountMap.get(word)+1);                             
    }
    else
    {
        orgMap.put(word, score);
        orgCountMap.put(word, 1);
        //orgSentenceMap.put(oname, str);
    }
}

private static void extractLocation(HashMap<String, Integer> locationMap,
        HashMap<String, Integer> locationCountMap,
        HashMap<String, Article_Location> locationArticleMap,
        int score,
        String word)
{
    if(locationMap.get(word)!=null)
    {
        locationMap.put(word, locationMap.get(word)+score);
        locationCountMap.put(word, locationCountMap.get(word)+1);                               
    }
    else
    {
        Article_Location articleLocation = LocationUtil.getLocation(word);

        locationMap.put(word, score);
        locationCountMap.put(word, 1);
        locationArticleMap.put(word, articleLocation);
    }   
}

private static void extractSUDate(long dateString,
        HashMap<String, Integer> dateMap,
        HashMap<String, Integer> dateCountMap,
        HashMap<String, String> dateSentenceMap, 
        String str,
        int score) {

    Annotation dateAnnotation = new Annotation(str);
    dateAnnotation.set(CoreAnnotations.DocDateAnnotation.class, FormatUtil.getDate(dateString));
    pipeline.annotate(dateAnnotation);

    for(CoreMap timex:dateAnnotation.get(TimeAnnotations.TimexAnnotations.class))
    {
        TimeExpression timeExpression = timex.get(TimeExpression.Annotation.class);

         if(timeExpression!=null && timeExpression.getTemporal()!=null &&
            timeExpression.getTemporal().getTimexValue()!=null)
         {           
             String word = checkDate(timeExpression.getTemporal().getTimexValue());

             if(word!=null)
             {
                 if(dateMap.get(word)!=null)
                 {
                     dateMap.put(word, dateMap.get(word)+score);
                     dateCountMap.put(word, dateCountMap.get(word)+1);
                     dateSentenceMap.put(word, dateSentenceMap.get(word)+" "+str);
                 }
                 else
                 {
                     dateMap.put(word, score);
                     dateCountMap.put(word, 1);
                     dateSentenceMap.put(word, str);
                 }                       
             }
         }
    }
}

private static int getSentiment(CoreMap sentence) {
    Tree annotatedTree = sentence.get(SentimentAnnotatedTree.class);
     int localScore = RNNCoreAnnotations.getPredictedClass(annotatedTree);
     int score = getScore(localScore);       
    return score;
}   


private static void mapLocation(HashMap<String, Integer> locationMap,
        HashMap<String, Integer> locationCountMap,
        HashMap<String, Article_Location> locationArticleMap,
        ArrayList<Article_NerLocation> locationList)
{       
    for(Map.Entry<String, Integer> entry : locationMap.entrySet())
    {
        String key = entry.getKey();
        Integer value = entry.getValue();

        Article_Location articleLocation = locationArticleMap.get(key);

        Article_NerLocation l1 = new Article_NerLocation();
        if(value>=1)
            l1.setNerSentiment(1);
        else if(value<=-1)
            l1.setNerSentiment(-1);
        else
            l1.setNerSentiment(0);            

        l1.setKeyword(key);
        l1.setCount(locationCountMap.get(key));

        if(articleLocation!=null)
        {                   
            l1.setNerCountry(articleLocation.getCountryCode());
            l1.setNerLatLong(articleLocation.getLatitude()+","+articleLocation.getLongitude());
            l1.setTimeZone(articleLocation.getTimeZone());
            l1.setCountryName(articleLocation.getCountryName());
        }

        locationList.add(l1);
    }
}

private static void mapDate(HashMap<String, Integer> dateMap,
        HashMap<String, Integer> dateCountMap,
        HashMap<String, String> dateSentenceMap,
        ArrayList<Artilcle_Ner_Date> datelist)
{               
    for(Map.Entry<String, Integer> entry : dateMap.entrySet())
    {
        String key = entry.getKey();
        Integer value = entry.getValue();

        Artilcle_Ner_Date d1 = new Artilcle_Ner_Date();

        if(value>=1)
            d1.setNerSentiment(1);
        else if(value<=-1)
            d1.setNerSentiment(-1);
        else
            d1.setNerSentiment(0);

        d1.setKeyword(key);
        d1.setCount(dateCountMap.get(key));
        d1.setSentence(dateSentenceMap.get(key));
        d1.setNerDateTheme1(SummaryThemeUtil.getSTByDate(dateSentenceMap.get(key)));
        datelist.add(d1);
    }   
}   

private static void mapOrg(HashMap<String, Integer> orgMap,
        HashMap<String, Integer> orgCountMap,
        ArrayList<Articel_Ner> organisationlist) 
{
    for(Map.Entry<String, Integer> entry : orgMap.entrySet())
    {
        String key = entry.getKey();
        Integer value = entry.getValue();

        Articel_Ner o1 = new Articel_Ner();
        if(value>=1)
            o1.setNerSentiment(1);
        else if(value<=-1)
            o1.setNerSentiment(-1); 
        else
            o1.setNerSentiment(0);            


        o1.setKeyword(key);
        o1.setCount(orgCountMap.get(key));
        organisationlist.add(o1);            
    }       
}

private static void mapPerson(HashMap<String, Integer> personMap,
        HashMap<String, Integer> personCountMap,
        ArrayList<Articel_Ner> personlist)
{
    for(Map.Entry<String, Integer> entry : personMap.entrySet())
    {
        String key = entry.getKey();
        Integer value = entry.getValue();

        Articel_Ner p1 = new Articel_Ner();
        if(value>=1)
            p1.setNerSentiment(1);
        else if(value<=-1)
            p1.setNerSentiment(-1);
        else
            p1.setNerSentiment(0);            

        p1.setKeyword(key);
        p1.setCount(personCountMap.get(key));
        personlist.add(p1);      
    }
}   

private static String checkDate(String date)
{               
    if(date.length()<10)
        return null;
    else if(date.length()>10)
        date = date.substring(0,10);

    if (date.matches("\\d{4}-\\d{2}-\\d{2}"))
        return date; 
    else
        return null;
}

public static void main(String args[])
{
    String text = "Lets meet on every 2nd week. Night is young. Happy new Year. The festival will be held on the following dates are 18 Feb 1997, the 20th of july and 4 days from today.";
    long pre = System.currentTimeMillis();
    HashMap<String, Object> map = getStanford(text, 1508745558);
    long post = System.currentTimeMillis();
    long diff = post-pre;

    System.out.println(diff);
    System.out.println(map);
}
}

Answer 1

我在您的示例文本上运行了这个命令：

java -Xmx8g edu.stanford.nlp.pipeline.StanfordCoreNLP -annotators tokenize,ssplit,pos,lemma,ner,parse,sentiment -parse.model edu/stanford/nlp/models/srparser/englishSR.ser.gz  -file example-1.txt -outputFormat text -ner.nthreads 4 -parse.nthreads 4

我得到了这个时间信息：

Annotation pipeline timing information:
TokenizerAnnotator: 0.0 sec.
WordsToSentencesAnnotator: 0.0 sec.
POSTaggerAnnotator: 0.0 sec.
MorphaAnnotator: 0.0 sec.
NERCombinerAnnotator: 0.0 sec.
ParserAnnotator: 0.2 sec.
SentimentAnnotator: 0.0 sec.
TOTAL: 0.4 sec. for 39 tokens at 105.1 tokens/sec.
Pipeline setup: 14.8 sec.
Total time for StanfordCoreNLP pipeline: 15.2 sec.

我看到0.4 sec 处理时间。

1.) 确保不要每次都重建管道。看来您的代码没有在 main() 方法中重建管道

2.) 我的命令对ner 和parse 使用多线程。请注意，我还使用了 shift-reduce 解析器，它比默认解析器快得多。

3.) 所有管道设置都可以在 Java API 代码中设置，方法是将它们分配给用于构建管道的 Properties 对象。以下是使用 Java API 的详尽文档（您必须将其转换为 Scala）：

Java API：https://stanfordnlp.github.io/CoreNLP/api.html

命令行：https://stanfordnlp.github.io/CoreNLP/cmdline.html

4.) 您不需要构建单独的 NERClassifierCombiner，您可以使用 ner 注释器，它也将运行 SUTime。

5.) 我应该注意时间将由解析器控制。您可以选择不使用 -parse.maxlen N 解析非常长的句子，并将 N 设置为您想要的任何标记长度。

6.) 如果您想获得完整实体提及的字符偏移量，请确保将 entitymentions 注释器添加到注释器列表的末尾。每个句子都有一个实体提及列表。每个实体提及都是一个 CoreMap。您可以使用以下代码访问实体提及的开始和结束字符偏移：

    List<CoreMap> mentions = sentence.get(CoreAnnotations.MentionsAnnotation.class);
    for (CoreMap mention : mentions) {
        int docCharBegin = mention.get(CoreAnnotations.CharacterOffsetBeginAnnotation.class);
        int docCharEnd = mention.get(CoreAnnotations.CharacterOffsetEndAnnotation.class);
    }

如果您对将其转换为 Scala 代码有任何疑问，请告诉我。

Answer 2

经过几天和几天的黑眼睛疼痛。这就是问题所在：

• Stanford "parse" 模型无论 PCFG 还是 SRparser 都是 CPU 杀手。您将永远无法扩展。充其量我每秒做 70 个文档。这是我能够在 tomcat 上管理的 15 个线程。从 RabbitMQ 使用的文档。使用 15 GB RAM 的机器 Intel Xeon 8Core VM。 CPU 一直是 90%。

• 所以如果你想做NER,sentiment,sutime。最好使用单独的库，而不是全部使用 stanford 3。对于 NER，您可以使用来自 stanford 的 NERClassifierCombiner。对于情绪，您可以使用 weka。要提取日期，您可以使用 natty。

• 现在我们能够做到 2,000 文档/秒。