使用 Stanford CoreNLP (3.5.2) 的并发处理

Question

我在同时注释多个句子时遇到了并发问题。我不清楚是我做错了什么还是 CoreNLP 中存在错误。

我的目标是使用多个并行运行的线程使用管道“tokenize、ssplit、pos、lemma、ner、parse、dcoref”来注释句子。每个线程分配自己的 StanfordCoreNLP 实例，然后将其用于注释。

问题是有时会抛出异常：

java.util.ConcurrentModificationException
	at java.util.ArrayList$Itr.checkForComodification(ArrayList.java:901)
	at java.util.ArrayList$Itr.next(ArrayList.java:851)
	at java.util.Collections$UnmodifiableCollection$1.next(Collections.java:1042)
	at edu.stanford.nlp.trees.GrammaticalStructure.analyzeNode(GrammaticalStructure.java:463)
	at edu.stanford.nlp.trees.GrammaticalStructure.analyzeNode(GrammaticalStructure.java:488)
	at edu.stanford.nlp.trees.GrammaticalStructure.analyzeNode(GrammaticalStructure.java:488)
	at edu.stanford.nlp.trees.GrammaticalStructure.analyzeNode(GrammaticalStructure.java:488)
	at edu.stanford.nlp.trees.GrammaticalStructure.analyzeNode(GrammaticalStructure.java:488)
	at edu.stanford.nlp.trees.GrammaticalStructure.analyzeNode(GrammaticalStructure.java:488)
	at edu.stanford.nlp.trees.GrammaticalStructure.analyzeNode(GrammaticalStructure.java:488)
	at edu.stanford.nlp.trees.GrammaticalStructure.analyzeNode(GrammaticalStructure.java:488)
	at edu.stanford.nlp.trees.GrammaticalStructure.analyzeNode(GrammaticalStructure.java:488)
	at edu.stanford.nlp.trees.GrammaticalStructure.analyzeNode(GrammaticalStructure.java:488)
	at edu.stanford.nlp.trees.GrammaticalStructure.analyzeNode(GrammaticalStructure.java:488)
	at edu.stanford.nlp.trees.GrammaticalStructure.<init>(GrammaticalStructure.java:201)
	at edu.stanford.nlp.trees.EnglishGrammaticalStructure.<init>(EnglishGrammaticalStructure.java:89)
	at edu.stanford.nlp.semgraph.SemanticGraphFactory.makeFromTree(SemanticGraphFactory.java:139)
	at edu.stanford.nlp.pipeline.DeterministicCorefAnnotator.annotate(DeterministicCorefAnnotator.java:89)
	at edu.stanford.nlp.pipeline.AnnotationPipeline.annotate(AnnotationPipeline.java:68)
	at edu.stanford.nlp.pipeline.StanfordCoreNLP.annotate(StanfordCoreNLP.java:412)

我附上了一个应用程序的示例代码，它在我的 Core i3 370M 笔记本电脑（Win 7 64bit，Java 1.8.0.45 64bit）上大约 20 秒内重现了该问题。此应用程序读取识别文本内涵 (RTE) 语料库的 XML 文件，然后使用标准 Java 并发类同时解析所有句子。本地 RTE XML 文件的路径需要作为命令行参数给出。在我的测试中，我在这里使用了公开可用的 XML 文件： http://www.nist.gov/tac/data/RTE/RTE3-DEV-FINAL.tar.gz

package semante.parser.stanford.server;

import java.io.FileInputStream;
import java.io.InputStreamReader;
import java.io.OutputStream;
import java.io.PrintStream;
import java.nio.charset.StandardCharsets;
import java.util.ArrayList;
import java.util.List;
import java.util.Properties;
import java.util.concurrent.ExecutorService;
import java.util.concurrent.Executors;
import java.util.concurrent.TimeUnit;
import java.util.concurrent.atomic.AtomicInteger;

import javax.xml.bind.JAXBContext;
import javax.xml.bind.Unmarshaller;
import javax.xml.bind.annotation.XmlAccessType;
import javax.xml.bind.annotation.XmlAccessorType;
import javax.xml.bind.annotation.XmlAttribute;
import javax.xml.bind.annotation.XmlElement;
import javax.xml.bind.annotation.XmlRootElement;

import edu.stanford.nlp.pipeline.Annotation;
import edu.stanford.nlp.pipeline.StanfordCoreNLP;

public class StanfordMultiThreadingTest {

	@XmlRootElement(name = "entailment-corpus")
	@XmlAccessorType (XmlAccessType.FIELD)
	public static class Corpus {
		@XmlElement(name = "pair")
		private List<Pair> pairList = new ArrayList<Pair>();

		public void addPair(Pair p) {pairList.add(p);}
		public List<Pair> getPairList() {return pairList;}
	}

	@XmlRootElement(name="pair")
	public static class Pair {

		@XmlAttribute(name = "id")
		String id;

		@XmlAttribute(name = "entailment")
		String entailment;

		@XmlElement(name = "t")
		String t;

		@XmlElement(name = "h")
		String h;

		private Pair() {}

		public Pair(int id, boolean entailment, String t, String h) {
			this();
			this.id = Integer.toString(id);
			this.entailment = entailment ? "YES" : "NO";
			this.t = t;
			this.h = h;
		}

		public String getId() {return id;}
		public String getEntailment() {return entailment;}
		public String getT() {return t;}
		public String getH() {return h;}
	}
	
	class NullStream extends OutputStream {
		@Override 
		public void write(int b) {}
	};

	private Corpus corpus;
	private Unmarshaller unmarshaller;
	private ExecutorService executor;

	public StanfordMultiThreadingTest() throws Exception {
		javax.xml.bind.JAXBContext jaxbCtx = JAXBContext.newInstance(Pair.class,Corpus.class);
		unmarshaller = jaxbCtx.createUnmarshaller();
		executor = Executors.newFixedThreadPool(Runtime.getRuntime().availableProcessors());
	}

	public void readXML(String fileName) throws Exception {
		System.out.println("Reading XML - Started");
		corpus = (Corpus) unmarshaller.unmarshal(new InputStreamReader(new FileInputStream(fileName), StandardCharsets.UTF_8));
		System.out.println("Reading XML - Ended");
	}

	public void parseSentences() throws Exception {
		System.out.println("Parsing - Started");

		// turn pairs into a list of sentences
		List<String> sentences = new ArrayList<String>();
		for (Pair pair : corpus.getPairList()) {
			sentences.add(pair.getT());
			sentences.add(pair.getH());
		}

		// prepare the properties
		final Properties props = new Properties();
		props.put("annotators", "tokenize, ssplit, pos, lemma, ner, parse, dcoref");

		// first run is long since models are loaded
		new StanfordCoreNLP(props);

		// to avoid the CoreNLP initialization prints (e.g. "Adding annotation pos")
		final PrintStream nullPrintStream = new PrintStream(new NullStream());
		PrintStream err = System.err;
		System.setErr(nullPrintStream);

		int totalCount = sentences.size();
		AtomicInteger counter = new AtomicInteger(0);

		// use java concurrency to parallelize the parsing
		for (String sentence : sentences) {
			executor.execute(new Runnable() {
				@Override
				public void run() {
					try {
						StanfordCoreNLP pipeline = new StanfordCoreNLP(props);
						Annotation annotation = new Annotation(sentence);
						pipeline.annotate(annotation);
						if (counter.incrementAndGet() % 20 == 0) {
							System.out.println("Done: " + String.format("%.2f", counter.get()*100/(double)totalCount));
						};
					} catch (Exception e) {
						System.setErr(err);
						e.printStackTrace();
						System.setErr(nullPrintStream);
						executor.shutdownNow();
					}
				}
			});
		}
		executor.shutdown();
		
		System.out.println("Waiting for parsing to end.");		
		executor.awaitTermination(10, TimeUnit.MINUTES);

		System.out.println("Parsing - Ended");
	}

	public static void main(String[] args) throws Exception {
		StanfordMultiThreadingTest smtt = new StanfordMultiThreadingTest();
		smtt.readXML(args[0]);
		smtt.parseSentences();
	}

}

在我尝试查找一些背景信息时，我遇到了来自斯坦福的 Christopher Manning 和 Gabor Angeli 给出的答案，这表明斯坦福 CoreNLP 的当代版本应该是线程安全的。然而，最近在 CoreNLP 3.4.1 版上的bug report 描述了一个并发问题。如标题所述，我使用的是 3.5.2 版本。

我不清楚我所面临的问题是由于错误还是由于我使用软件包的方式有问题。如果有更多知识的人可以对此有所了解，我将不胜感激。我希望示例代码对重现问题有用。谢谢！

[1]：

Answer 1

您是否尝试过使用threads 选项？您可以为单个StanfordCoreNLP 管道指定多个线程，然后它将并行处理句子。

例如，如果要在 8 个核心上处理句子，请将 threads 选项设置为 8：

Properties props = new Properties();
props.put("annotators", "tokenize, ssplit, pos, lemma, ner, parse, dcoref");
props.put("threads", "8")
StanfordCoreNLP pipeline  = new StanfordCoreNLP(props);

尽管如此，我认为您的解决方案也应该可以工作，我们将检查是否存在一些并发错误，但同时使用此选项可能会解决您的问题。

Answer 2

我遇到了同样的问题，使用最新的 github 修订版（今天）解决了这个问题。所以我认为这是一个CoreNLP问题，从3.5.2开始就解决了。

另见CoreNLP on Apache Spark