聊聊Spring AI Alibaba的SentenceSplitter

本文主要研究一下Spring AI Alibaba的SentenceSplitter

SentenceSplitter

spring-ai-alibaba-core/src/main/java/com/alibaba/cloud/ai/transformer/splitter/SentenceSplitter.java

复制代码
public class SentenceSplitter extends TextSplitter {

	private final EncodingRegistry registry = Encodings.newLazyEncodingRegistry();

	private final Encoding encoding = registry.getEncoding(EncodingType.CL100K_BASE);

	private static final int DEFAULT_CHUNK_SIZE = 1024;

	private final SentenceModel sentenceModel;

	private final int chunkSize;

	public SentenceSplitter() {
		this(DEFAULT_CHUNK_SIZE);
	}

	public SentenceSplitter(int chunkSize) {
		this.chunkSize = chunkSize;
		this.sentenceModel = getSentenceModel();
	}

	@Override
	protected List<String> splitText(String text) {
		SentenceDetectorME sentenceDetector = new SentenceDetectorME(sentenceModel);
		String[] texts = sentenceDetector.sentDetect(text);
		if (texts == null || texts.length == 0) {
			return Collections.emptyList();
		}

		List<String> chunks = new ArrayList<>();
		StringBuilder chunk = new StringBuilder();
		for (int i = 0; i < texts.length; i++) {
			int currentChunkSize = getEncodedTokens(chunk.toString()).size();
			int textTokenSize = getEncodedTokens(texts[i]).size();
			if (currentChunkSize + textTokenSize > chunkSize) {
				chunks.add(chunk.toString());
				chunk = new StringBuilder(texts[i]);
			}
			else {
				chunk.append(texts[i]);
			}

			if (i == texts.length - 1) {
				chunks.add(chunk.toString());
			}
		}

		return chunks;
	}

	private SentenceModel getSentenceModel() {
		try (InputStream is = getClass().getResourceAsStream("/opennlp/opennlp-en-ud-ewt-sentence-1.2-2.5.0.bin")) {
			if (is == null) {
				throw new RuntimeException("sentence model is invalid");
			}

			return new SentenceModel(is);
		}
		catch (IOException e) {
			throw new RuntimeException(e);
		}
	}

	private List<Integer> getEncodedTokens(String text) {
		Assert.notNull(text, "Text must not be null");
		return this.encoding.encode(text).boxed();
	}

}

SentenceSplitter继承了TextSplitter,其构造器会通过getSentenceModel()来加载/opennlp/opennlp-en-ud-ewt-sentence-1.2-2.5.0.bin这个SentenceModel;splitText方法创建SentenceDetectorME,使用其sentDetect来拆分句子,再根据chunkSize进一步合并或拆分

示例

spring-ai-alibaba-core/src/test/java/com/alibaba/cloud/ai/transformer/splitter/SentenceSplitterTests.java

复制代码
class SentenceSplitterTests {

	private SentenceSplitter splitter;

	private static final int CUSTOM_CHUNK_SIZE = 100;

	@BeforeEach
	void setUp() {
		// Initialize with default chunk size
		splitter = new SentenceSplitter();
	}

	/**
	 * Test default constructor. Verifies that splitter can be created with default chunk
	 * size.
	 */
	@Test
	void testDefaultConstructor() {
		SentenceSplitter defaultSplitter = new SentenceSplitter();
		assertThat(defaultSplitter).isNotNull();
	}

	/**
	 * Test constructor with custom chunk size. Verifies that splitter can be created with
	 * specified chunk size.
	 */
	@Test
	void testCustomChunkSizeConstructor() {
		SentenceSplitter customSplitter = new SentenceSplitter(CUSTOM_CHUNK_SIZE);
		assertThat(customSplitter).isNotNull();
	}

	/**
	 * Test splitting simple sentences. Verifies basic sentence splitting functionality.
	 */
	@Test
	void testSplitSimpleSentences() {
		String text = "This is a test. This is another test. And this is a third test.";
		Document doc = new Document(text);
		List<Document> documents = splitter.apply(Collections.singletonList(doc));

		assertThat(documents).isNotNull();
		assertThat(documents).hasSize(1);
		assertThat(documents.get(0).getText()).contains("This is a test", "This is another test",
				"And this is a third test");
	}

	/**
	 * Test splitting empty text. Verifies handling of empty input.
	 */
	@Test
	void testSplitEmptyText() {
		Document doc = new Document("");
		List<Document> documents = splitter.apply(Collections.singletonList(doc));
		assertThat(documents).isEmpty();
	}

	/**
	 * Test splitting text with special characters. Verifies handling of text with various
	 * punctuation and special characters.
	 */
	@Test
	void testSplitTextWithSpecialCharacters() {
		String text = "Hello, world! How are you? I'm doing great... This is a test; with various punctuation.";
		Document doc = new Document(text);
		List<Document> documents = splitter.apply(Collections.singletonList(doc));

		assertThat(documents).isNotNull();
		assertThat(documents).hasSize(1);
		assertThat(documents.get(0).getText()).contains("Hello, world", "How are you", "I'm doing great",
				"This is a test");
	}

	/**
	 * Test splitting long text. Verifies handling of text that exceeds default chunk
	 * size.
	 */
	@Test
	void testSplitLongText() {
		// Generate a very long text that will exceed the default chunk size (1024
		// tokens)
		StringBuilder longText = new StringBuilder();
		String longSentence = "This is a very long sentence with many words that will contribute to the total token count and eventually force the text to be split into multiple chunks because it exceeds the default chunk size limit of 1024 tokens. ";
		// Repeat the sentence enough times to ensure we exceed the chunk size
		for (int i = 0; i < 50; i++) {
			longText.append(longSentence);
		}
		Document doc = new Document(longText.toString());

		List<Document> documents = splitter.apply(Collections.singletonList(doc));

		// Verify that the text was split into multiple documents
		assertThat(documents).isNotNull();
		assertThat(documents).hasSizeGreaterThan(1);
		// Verify that each document contains part of the original text
		documents.forEach(document -> assertThat(document.getText()).contains("This is a very long sentence"));
	}

	/**
	 * Test splitting text with multiple line breaks. Verifies handling of text with
	 * various types of line breaks.
	 */
	@Test
	void testSplitTextWithLineBreaks() {
		String text = "First sentence.\nSecond sentence.\r\nThird sentence.\rFourth sentence.";
		Document doc = new Document(text);
		List<Document> documents = splitter.apply(Collections.singletonList(doc));

		assertThat(documents).isNotNull();
		assertThat(documents.get(0).getText()).contains("First sentence", "Second sentence", "Third sentence",
				"Fourth sentence");
	}

	/**
	 * Test splitting text with single character sentences. Verifies handling of very
	 * short sentences.
	 */
	@Test
	void testSplitSingleCharacterSentences() {
		String text = "A. B. C. D.";
		Document doc = new Document(text);
		List<Document> documents = splitter.apply(Collections.singletonList(doc));

		assertThat(documents).isNotNull();
		assertThat(documents).hasSize(1);
		assertThat(documents.get(0).getText()).contains("A", "B", "C", "D");
	}

	/**
	 * Test splitting multiple documents. Verifies handling of multiple input documents.
	 */
	@Test
	void testSplitMultipleDocuments() {
		List<Document> inputDocs = new ArrayList<>();
		inputDocs.add(new Document("First document. With multiple sentences."));
		inputDocs.add(new Document("Second document. Also with multiple sentences."));

		List<Document> documents = splitter.apply(inputDocs);
		assertThat(documents).isNotNull();
		assertThat(documents).hasSizeGreaterThan(1);
	}

}

小结

Spring AI Alibaba提供了SentenceSplitter,它使用了opennlp的SentenceDetectorME进行拆分,其构造器会加载/opennlp/opennlp-en-ud-ewt-sentence-1.2-2.5.0.bin这个SentenceModel。

doc

相关推荐
jiuri_121534 分钟前
Docker使用详解:在ARM64嵌入式环境部署Python应用
python·docker·容器
chenchihwen37 分钟前
AI代码开发宝库系列:Function Call
人工智能·python·1024程序员节·dashscope
FreeBuf_38 分钟前
微软Copilot被用于窃取OAuth令牌,AI Agent成为攻击者帮凶
人工智能·microsoft·copilot
学slam的小范39 分钟前
ROS跑ORB-SLAM3遇见的问题总结
人工智能·机器人·自动驾驶
coding消烦员1 小时前
新版 vscode 去除快捷键 Ctrl+I 显示 Copilot 的 AI 对话框
人工智能·vscode·copilot
周杰伦_Jay1 小时前
【自动驾驶开源仿真平台】Carla、AirSim、Udacity self-driving-car-sim、Apollo、Autoware。
人工智能·机器学习·自动驾驶
牛奶还是纯的好2 小时前
双目测距实战5-立体矫正
人工智能·3d
汤姆yu2 小时前
基于python的化妆品销售分析系统
开发语言·python·化妆品销售分析
无风听海2 小时前
神经网络之窗口大小对词语义向量的影响
人工智能·深度学习·神经网络