4
我在同時註釋多個語句時遇到併發問題。我不清楚自己是否做錯了什麼，或者CoreNLP中有bug。使用斯坦福CoreNLP併發處理（3.5.2）
我的目標是使用多個並行運行的線程，通過管道「標記化，ssplit，pos，引理，ner，parse，dcoref」對句子進行註釋。每個線程分配自己的StanfordCoreNLP實例，然後將其用於註釋。
的問題是，在某些時候拋出一個異常：
java.util.ConcurrentModificationException
\t at java.util.ArrayList$Itr.checkForComodification(ArrayList.java:901)
\t at java.util.ArrayList$Itr.next(ArrayList.java:851)
\t at java.util.Collections$UnmodifiableCollection$1.next(Collections.java:1042)
\t at edu.stanford.nlp.trees.GrammaticalStructure.analyzeNode(GrammaticalStructure.java:463)
\t at edu.stanford.nlp.trees.GrammaticalStructure.analyzeNode(GrammaticalStructure.java:488)
\t at edu.stanford.nlp.trees.GrammaticalStructure.analyzeNode(GrammaticalStructure.java:488)
\t at edu.stanford.nlp.trees.GrammaticalStructure.analyzeNode(GrammaticalStructure.java:488)
\t at edu.stanford.nlp.trees.GrammaticalStructure.analyzeNode(GrammaticalStructure.java:488)
\t at edu.stanford.nlp.trees.GrammaticalStructure.analyzeNode(GrammaticalStructure.java:488)
\t at edu.stanford.nlp.trees.GrammaticalStructure.analyzeNode(GrammaticalStructure.java:488)
\t at edu.stanford.nlp.trees.GrammaticalStructure.analyzeNode(GrammaticalStructure.java:488)
\t at edu.stanford.nlp.trees.GrammaticalStructure.analyzeNode(GrammaticalStructure.java:488)
\t at edu.stanford.nlp.trees.GrammaticalStructure.analyzeNode(GrammaticalStructure.java:488)
\t at edu.stanford.nlp.trees.GrammaticalStructure.analyzeNode(GrammaticalStructure.java:488)
\t at edu.stanford.nlp.trees.GrammaticalStructure.<init>(GrammaticalStructure.java:201)
\t at edu.stanford.nlp.trees.EnglishGrammaticalStructure.<init>(EnglishGrammaticalStructure.java:89)
\t at edu.stanford.nlp.semgraph.SemanticGraphFactory.makeFromTree(SemanticGraphFactory.java:139)
\t at edu.stanford.nlp.pipeline.DeterministicCorefAnnotator.annotate(DeterministicCorefAnnotator.java:89)
\t at edu.stanford.nlp.pipeline.AnnotationPipeline.annotate(AnnotationPipeline.java:68)
\t at edu.stanford.nlp.pipeline.StanfordCoreNLP.annotate(StanfordCoreNLP.java:412)我附加能重現問題在大約20秒對我的核心應用程序的示例代碼i3 370M筆記本電腦（Win 7 64bit，Java 1.8.0.45 64bit）。此應用程序讀取識別文本語言（RTE）語料庫的XML文件，然後使用標準Java併發類同時解析所有語句。需要將本地RTE XML文件的路徑作爲命令行參數提供。在我的測試中，我使用這裏的公開可用的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 {
\t @XmlRootElement(name = "entailment-corpus")
\t @XmlAccessorType (XmlAccessType.FIELD)
\t public static class Corpus {
\t \t @XmlElement(name = "pair")
\t \t private List<Pair> pairList = new ArrayList<Pair>();
\t \t public void addPair(Pair p) {pairList.add(p);}
\t \t public List<Pair> getPairList() {return pairList;}
\t }
\t @XmlRootElement(name="pair")
\t public static class Pair {
\t \t @XmlAttribute(name = "id")
\t \t String id;
\t \t @XmlAttribute(name = "entailment")
\t \t String entailment;
\t \t @XmlElement(name = "t")
\t \t String t;
\t \t @XmlElement(name = "h")
\t \t String h;
\t \t private Pair() {}
\t \t public Pair(int id, boolean entailment, String t, String h) {
\t \t \t this();
\t \t \t this.id = Integer.toString(id);
\t \t \t this.entailment = entailment ? "YES" : "NO";
\t \t \t this.t = t;
\t \t \t this.h = h;
\t \t }
\t \t public String getId() {return id;}
\t \t public String getEntailment() {return entailment;}
\t \t public String getT() {return t;}
\t \t public String getH() {return h;}
\t }
\t
\t class NullStream extends OutputStream {
\t \t @Override
\t \t public void write(int b) {}
\t };
\t private Corpus corpus;
\t private Unmarshaller unmarshaller;
\t private ExecutorService executor;
\t public StanfordMultiThreadingTest() throws Exception {
\t \t javax.xml.bind.JAXBContext jaxbCtx = JAXBContext.newInstance(Pair.class,Corpus.class);
\t \t unmarshaller = jaxbCtx.createUnmarshaller();
\t \t executor = Executors.newFixedThreadPool(Runtime.getRuntime().availableProcessors());
\t }
\t public void readXML(String fileName) throws Exception {
\t \t System.out.println("Reading XML - Started");
\t \t corpus = (Corpus) unmarshaller.unmarshal(new InputStreamReader(new FileInputStream(fileName), StandardCharsets.UTF_8));
\t \t System.out.println("Reading XML - Ended");
\t }
\t public void parseSentences() throws Exception {
\t \t System.out.println("Parsing - Started");
\t \t // turn pairs into a list of sentences
\t \t List<String> sentences = new ArrayList<String>();
\t \t for (Pair pair : corpus.getPairList()) {
\t \t \t sentences.add(pair.getT());
\t \t \t sentences.add(pair.getH());
\t \t }
\t \t // prepare the properties
\t \t final Properties props = new Properties();
\t \t props.put("annotators", "tokenize, ssplit, pos, lemma, ner, parse, dcoref");
\t \t // first run is long since models are loaded
\t \t new StanfordCoreNLP(props);
\t \t // to avoid the CoreNLP initialization prints (e.g. "Adding annotation pos")
\t \t final PrintStream nullPrintStream = new PrintStream(new NullStream());
\t \t PrintStream err = System.err;
\t \t System.setErr(nullPrintStream);
\t \t int totalCount = sentences.size();
\t \t AtomicInteger counter = new AtomicInteger(0);
\t \t // use java concurrency to parallelize the parsing
\t \t for (String sentence : sentences) {
\t \t \t executor.execute(new Runnable() {
\t \t \t \t @Override
\t \t \t \t public void run() {
\t \t \t \t \t try {
\t \t \t \t \t \t StanfordCoreNLP pipeline = new StanfordCoreNLP(props);
\t \t \t \t \t \t Annotation annotation = new Annotation(sentence);
\t \t \t \t \t \t pipeline.annotate(annotation);
\t \t \t \t \t \t if (counter.incrementAndGet() % 20 == 0) {
\t \t \t \t \t \t \t System.out.println("Done: " + String.format("%.2f", counter.get()*100/(double)totalCount));
\t \t \t \t \t \t };
\t \t \t \t \t } catch (Exception e) {
\t \t \t \t \t \t System.setErr(err);
\t \t \t \t \t \t e.printStackTrace();
\t \t \t \t \t \t System.setErr(nullPrintStream);
\t \t \t \t \t \t executor.shutdownNow();
\t \t \t \t \t }
\t \t \t \t }
\t \t \t });
\t \t }
\t \t executor.shutdown();
\t \t
\t \t System.out.println("Waiting for parsing to end."); \t \t
\t \t executor.awaitTermination(10, TimeUnit.MINUTES);
\t \t System.out.println("Parsing - Ended");
\t }
\t public static void main(String[] args) throws Exception {
\t \t StanfordMultiThreadingTest smtt = new StanfordMultiThreadingTest();
\t \t smtt.readXML(args[0]);
\t \t smtt.parseSentences();
\t }
}在我試圖找到一些背景資料，我遇到了來自斯坦福大學的Christopher Manning和Gabor Angeli給出答案，這表明當代版本斯坦福CoreNLP應該是線程安全的。然而，CoreNLP 3.4.1版最近的bug report描述了一個併發問題。正如標題中提到的，我使用的是3.5.2版本。
目前還不清楚我面臨的問題是由於錯誤還是由於我使用軟件包的方式有問題。如果有人對此有所瞭解，我會很感激。我希望示例代碼對於重現問題很有用。謝謝！
[1]：
感謝您的建議。我想嘗試一下，但我不確定如何使用界面。假設我設置了「線索」屬性，我應該如何通過並行註釋句子？使用多個使用StanfordCoreNLP實例的線程？或者通過不同於'annotate（）'的方法一次傳遞幾個句子？謝謝！ 調用 – Assaf
「Annotation」構造函數的參數實際上不是一個句子，而是整個文檔。在「句子」變量中存儲幾個（或甚至所有）句子，並用「\ n」分隔它們。還要將選項「ssplit.eolonly」設置爲「true」，以防止句子拆分器錯誤地分割實際句子。解析後，註釋對象包含一個句子列表，其中每個句子具有解析，pos，引理等註釋。 –
謝謝，我試過了。然而，要麼註釋由'\ n'分隔的多個句子的模式存在問題，要麼我做錯了什麼。我能夠解析100個句子，但不是1000或2000.當用1000或2000個句子餵食時，對annotate（）的調用會無休止地運行。另外，當我用100個句子測試時，1,2或4個線程（我的硬件有4個）在性能上幾乎沒有差別。它比使用單個線程稍微慢一點，並且每次只用一個語句調用annotate（）。 – Assaf