Java 余弦相似性始终为1.0
我试图用Java中的DeepLearning4J框架构建段落向量并对其进行一些推断。当我将段落向量构建到ZIP文件夹中时,我可以通过使用行号获得相似性,如下所示:Java 余弦相似性始终为1.0,java,nlp,deeplearning4j,Java,Nlp,Deeplearning4j,我试图用Java中的DeepLearning4J框架构建段落向量并对其进行一些推断。当我将段落向量构建到ZIP文件夹中时,我可以通过使用行号获得相似性,如下所示: SentenceIterator sentenceIterator = new BasicLineIterator(new File(inputFilePath)); AbstractCache<VocabWord> abstractCache = new AbstractCache<VocabWord>();
SentenceIterator sentenceIterator = new BasicLineIterator(new File(inputFilePath));
AbstractCache<VocabWord> abstractCache = new AbstractCache<VocabWord>();
TokenizerFactory tokenizerFactory = new DefaultTokenizerFactory();
tokenizerFactory.setTokenPreProcessor(new CommonPreprocessor());
LabelsSource labelsSource = new LabelsSource("LINE_");
ParagraphVectors paragraphVectors = new ParagraphVectors.Builder()
.minWordFrequency(1)
.iterations(5)
.epochs(1)
.layerSize(100)
.learningRate(0.025)
.labelsSource(labelsSource)
.windowSize(5)
.iterate(sentenceIterator)
.trainWordVectors(false)
.vocabCache(abstractCache)
.tokenizerFactory(tokenizerFactory)
.sampling(0)
.build();
paragraphVectors.fit();
double similarity1 = paragraphVectors.similarity("LINE_9835", "LINE_100");
System.out.println("Similarity: " + similarity1);
WordVectorSerializer.writeParagraphVectors(paragraphVectors, outputParagraphVectorsFilePath);
inputFilePath
变量指磁盘上包含向量的ZIP文件夹所在的位置。当我运行此函数时,我得到以下信息:
余弦相似性A/B:1.0
余弦相似性A/B2:1.0
即使我改变周围的向量并将它们与其他向量进行比较,我也会得到相同的1.0。我做错什么了吗?任何帮助都将不胜感激。根据GitHub上发布的问题,余弦相似性不准确的主要原因是dl4j
和nd4j
使用的版本不正确。我的项目中使用的版本是0.7.1
。更新到0.9.1
后,我能够得到准确的答案。以下是一些重要的指导原则:
TokenizerFactory tokenizerFactory = new DefaultTokenizerFactory();
tokenizerFactory.setTokenPreProcessor(new CommonPreprocessor());
ParagraphVectors paragraphVectors = WordVectorSerializer.readParagraphVectors(new File(inputFilePath));
paragraphVectors.setTokenizerFactory(tokenizerFactory);
paragraphVectors.getConfiguration().setIterations(1);
INDArray inferredVectorA = paragraphVectors.inferVector("This is my world .");
INDArray inferredVectorA2 = paragraphVectors.inferVector("This is my world .");
INDArray inferredVectorB = paragraphVectors.inferVector("This is my way .");
System.out.println("Cosine similarity A/B:" + Transforms.cosineSim(inferredVectorA, inferredVectorB));
System.out.println("Cosine similarity A/B2:" + Transforms.cosineSim(inferredVectorA, inferredVectorA2));