`

生成文本聚类java实现 (3)

 
阅读更多

    很多网友看到我的聚类的研究,到后来基本上都是到carrot2的研究上去了。但由于carrot2对中文的理解很不靠谱,所以参考了网络上的一些资料,现在贡献出来所有代码。

 

 代码的思路就是找字或者词出现的频度,并进行打分,最后按照出现次数和重要性,找出重要的语汇。现在贴出来一些可用的代码。

 ClusterBuilder.java

 /**

 * 
* @author  
* @version 创建时间:2011-3-8 下午02:02:36
* 聚类生成器
 */
public class ClusterBuilder {
	private static final Log LOG;
	private List<DocCluster> clusters;
	private ICTHit[] docs;
	private int maxLevels;
	private ClusteringOptions[] options;
	private boolean useTagsAsTitle;
	private String wordsExcluded;
	private static short[] bit1Table;

	static {
		LOG = LogFactory.getLog(ClusterBuilder.class.getName());

		bit1Table = new short[65536];

		for (int n = 0; n < bit1Table.length; n++) {
			String s = Integer.toBinaryString(n);
			short m = 0;
			for (int k = 0; k < s.length(); k++) {
				if (s.charAt(k) == '1') {
					m = (short) (m + 1);
				}
			}
			bit1Table[n] = m;
		}
	}

	private static int getValidBitCount(long n) {
		int i3 = (int) (n % 65536L);
		n /= 65536L;
		int i2 = (int) (n % 65536L);
		n /= 65536L;
		int i1 = (int) (n % 65536L);
		n /= 65536L;
		int i0 = (int) (n % 65536L);
		return bit1Table[i0] + bit1Table[i1] + bit1Table[i2] + bit1Table[i3];
	}

	private static int getDocHitCount(long[] hits) {
		assert (hits != null);
		if (hits == null)
			return 0;
		int n0 = 0;
		for (int i = 0; i < hits.length; i++) {
			n0 += getValidBitCount(hits[i]);
		}
		return n0;
	}

	public ClusterBuilder() {
		for (int n = 0; n < bit1Table.length; n++)
		{
			String s = Integer.toBinaryString(n);
			short m = 0;
			for (int k = 0; k < s.length(); k++)
			{
				if (s.getBytes()[k] == '1')
				{
					m = (short)(m + 1);
				}
			}
			bit1Table[n] = m;
		}
	}
	/**
	 * 
	 * @param docsToCluster 要聚类的记录列表
	 * @param exWords 不使用的主题词列表,多个词用西文逗号分隔。这些词将不会作为主题词。
	 * @param maxLevels 最大聚类级数
	 * @param useTagsAsTitle 是否使用主题词作为类别主题词。如果不使用,则根据文档标题自动生成类别主题词。
	 */
	public ClusterBuilder(ICTHit[] docsToCluster, String exWords, int maxLevels, boolean useTagsAsTitle) {
		this.useTagsAsTitle = useTagsAsTitle;
		this.wordsExcluded = exWords;
		this.maxLevels = maxLevels;
		this.docs = docsToCluster;
		this.options = new ClusteringOptions[3];
		this.options[0] = new ClusteringOptions();
		this.options[0].setDocMaxTagCount(10);
		this.options[0].setMinTagRelevance(60);
		this.options[0].setMinSameDocPercent(80);

		this.options[1] = new ClusteringOptions();
		this.options[1].setDocMaxTagCount(8);
		this.options[1].setMinTagRelevance(85);
		this.options[1].setMinSameDocPercent(70);
		this.options[1].setTagMinDocCount(2);
		this.options[1].setMinSameDocs(2);

		this.options[2] = new ClusteringOptions();
		this.options[2].setDocMaxTagCount(8);
		this.options[2].setMinTagRelevance(50);
		this.options[2].setMinSameDocPercent(70);
		this.options[2].setTagMinDocCount(2);
		this.options[2].setMinSameDocs(2);
	}
	/**
	 * 对Docs记录列表执行聚类,结果存放于Clusters中
	 */
	public void cluster() {
		this.clusters = createLevelClusters(docs, 0, options[0]);
		List subs = null;
		if (this.maxLevels <= 1) {
			return;
		}
		for (DocCluster dc : this.clusters) {
			if ((dc.getDocList().length < options[0].getMinDocsToCluster()) || (dc.getTags() == "其他"))
				continue;
			subs = createLevelClusters(dc.getDocList(), 1, options[1]);
			if (subs.size() > 1)
				dc.setSubclusters(subs);
		}
	}
	/**
	 * 创建一个层级的聚类
	 * @param docs 文档列表
	 * @param level 层级号
	 * @param levelOpt 该层级的聚类选项
	 * @return
	 */
	private List<DocCluster> createLevelClusters(ICTHit[] docs, int level, ClusteringOptions levelOpt) {
		TagHitMatrix matrix = new TagHitMatrix(docs.length, levelOpt.getDocMaxTagCount());
		List clusters = new ArrayList();
		int i, ValidTagCount;
		int DocCount = 0;
		// 扫描文档列表,根据每个文档的主题词列表,初始化主题词文档对照表。
		for (i = 0; i < docs.length; i++) {
			ICTHit d = docs[i];
			int validTagCount = 0;
			if (d.getTagList() != null) {
				String[] tagList = d.getTagList();
				for (int tagIdx = 0; (tagIdx < tagList.length) && (validTagCount < levelOpt.getDocMaxTagCount()); tagIdx++) {
					String tag = tagList[tagIdx].trim();
					 // 主题词长度大于6个字的丢弃
					if ((tag.length() <= 0)
							|| (tag.length() > 20)
							|| ((this.wordsExcluded.length() != 0) && ((tag.contains(this.wordsExcluded)) || (this.wordsExcluded
									.contains(tag)))))
						continue;
					matrix.AddDocHit(tag, i);
					validTagCount++;
				}
			}

		}

		int maxKwDocCount = 0;
		List entryListToRemove = new ArrayList();
		String kwWithMaxDocCount = "";
		LOG.debug("有效关键词:");
		for (Map.Entry entry : matrix.entrySet()) {
			// 统计当前主题词的命中文档数,文档数小于预设值,则该主题词将被删除
			int n = getDocHitCount((long[]) entry.getValue());
			if (n < levelOpt.getTagMinDocCount()) {
				entryListToRemove.add((String) entry.getKey());
			} else {
				LOG.debug((String) entry.getKey() + "(" + n + "), ");

				DocCount += n;
			}
			if (n > maxKwDocCount) {
				maxKwDocCount = n;
				kwWithMaxDocCount = (String) entry.getKey();
			}
		}
		LOG.debug("");

		LOG.debug("被忽略的关键词:");

		for (i = 0; i < entryListToRemove.size(); i++) {
			LOG.debug((String) entryListToRemove.get(i) + ", ");
			matrix.remove(entryListToRemove.get(i));
		}

		LOG.debug("");

		LOG.debug(entryListToRemove.size() + "个关键词被忽略。剩余" + matrix.size() + "个关键词。");

		LOG.debug("最大文档数的关键词:" + kwWithMaxDocCount + ",文档数:" + maxKwDocCount + "。");

		double docCountPerTag = matrix.size() > 0 ? DocCount / matrix.size() : 0.0D;
		LOG.debug("关键词平均文档数:" + docCountPerTag);

		levelOpt.setMinSameDocs((int) (docCountPerTag / (2.0D + level)));
		if (levelOpt.getMinSameDocs() < 1) {
			levelOpt.setMinSameDocs(1);
		}

		while (mergeClusters(matrix, levelOpt) > 0) {
		}
		return createResult(matrix, docs, level, levelOpt);
	}

	private int mergeClusters(TagHitMatrix matrix, ClusteringOptions opt) {
		if (matrix.size() == 0)
			return 0;
		long[] docHitsMerged = (long[]) null;
		long[] maxDocHitsMerged = (long[]) null;
		String word1 = "";
		String word2 = "";
		String word1ToMerge = "";
		String word2ToMerge = "";
		int i,j;
		int sameDocs = 0;
		// 初始化一个相关度数组,0到100分,共101项
		List rankMatrix = new ArrayList();
		for (i = 0; i < 101; i++) {
			rankMatrix.add(new ArrayList());
		}
		List matrix2List = new ArrayList();
		matrix2List.addAll(matrix.entrySet());
		// 将主题词文档映射表中的主题词两两比对
		for (int i1 = 0; i1 < matrix2List.size() - 1; i1++) {
			Map.Entry hits1 = (Map.Entry) matrix2List.get(i1);
			word1 = (String) hits1.getKey();
			for (int i2 = i1 + 1; i2 < matrix2List.size(); i2++) {
				Map.Entry hits2 = (Map.Entry) matrix2List.get(i2);
				word2 = (String) hits2.getKey();
				Object[] re = getWordsRelevance(mapEntry2TagHitEntry(hits1), mapEntry2TagHitEntry(hits2),
						docHitsMerged, sameDocs, opt, matrix.hitsItemCount);
				// 计算两个词的相关性,获取两词的文档汇总表,以及相同文档数
				int nRank = ((Integer) re[0]).intValue();
				docHitsMerged = (long[]) re[1];
				sameDocs = ((Integer) re[2]).intValue();
				// 相关度小于预设阈值的忽略
				if (nRank >= opt.getMinTagRelevance()) {
					((List) rankMatrix.get(nRank)).add(new IdPair(i1, i2));
				}

			}

		}

		List tagListToRemove = new ArrayList();
		List entryListMerged = new ArrayList();
		entryListMerged.add(new TagHitEntry("", null));
		HashSet idPairTable = new HashSet();
		TagHitEntry entryToMerge1;
		while (true) {
			// 找到最大相关性的两个主题词
			for (i = 100; (i >= opt.getMinTagRelevance()) && (((List) rankMatrix.get(i)).size() == 0); i--){};
			if (i < opt.getMinTagRelevance()) {
				break;
			}
			IdPair ip = (IdPair) ((List) rankMatrix.get(i)).get(0);
			// 合并两个类别
			((List) rankMatrix.get(i)).remove(0);
			
			entryToMerge1 = ip.Id1 >= 0 ? mapEntry2TagHitEntry((Map.Entry) matrix2List.get(ip.Id1))
					: (TagHitEntry) entryListMerged.get(-ip.Id1);
			TagHitEntry entryToMerge2 = ip.Id2 >= 0 ? mapEntry2TagHitEntry((Map.Entry) matrix2List.get(ip.Id2))
					: (TagHitEntry) entryListMerged.get(-ip.Id2);
			word1ToMerge = entryToMerge1.key;
			word2ToMerge = entryToMerge2.key;
			assert ((word1ToMerge.length() > 0) && (word2ToMerge.length() > 0));

			String wordsMerged = word1ToMerge + "," + word2ToMerge;
			long[] lDocs0 = entryToMerge1.value;
			long[] lDocs1 = entryToMerge2.value;
			maxDocHitsMerged = new long[matrix.hitsItemCount];
			for (i = 0; i < lDocs0.length; i++) {
				lDocs0[i] |= lDocs1[i];// 获取合并的文档集
			}
			if (ip.Id1 >= 0)
				tagListToRemove.add(word1ToMerge);
			else
				entryListMerged.set(-ip.Id1, new TagHitEntry("", null));
			if (ip.Id2 >= 0)
				tagListToRemove.add(word2ToMerge);
			else {
				entryListMerged.set(-ip.Id2, new TagHitEntry("", null));
			}
			entryListMerged.add(new TagHitEntry(wordsMerged, maxDocHitsMerged));
			// 替换与合并主题词有关联的其他相关主题词对的评分
			int idMerged = -(entryListMerged.size() - 1);
			int id2 = 0;

			boolean CanDelete = false;

			for (i = 0; i <= 100; i++) {
				int ListCount = ((List) rankMatrix.get(i)).size();
				if (ListCount == 0) {
					continue;
				}

				for (j = 0; j < ListCount; j++) {
					IdPair p = (IdPair) ((List) rankMatrix.get(i)).get(j);
					CanDelete = false;
					if ((ip.Id1 == p.Id1) || (ip.Id2 == p.Id1)) {
						id2 = p.Id2;
						CanDelete = true;
					} else if ((ip.Id1 == p.Id2) || (ip.Id2 == p.Id2)) {
						id2 = p.Id1;
						CanDelete = true;
					}
					if (!CanDelete)
						continue;
					if (idMerged == id2) {
						continue;
					}

					((List) rankMatrix.get(i)).remove(j);
					j--;
					ListCount--;

					IdPair pairMerged = new IdPair(idMerged, id2);
					if (idPairTable.contains(pairMerged)) {
						continue;
					}

					TagHitEntry e2 = id2 >= 0 ? mapEntry2TagHitEntry((Map.Entry) matrix2List.get(id2))
							: (TagHitEntry) entryListMerged.get(-id2);

					assert ((e2.key.length() != 0) && (e2.key != wordsMerged));

					Object[] re = getWordsRelevance(new TagHitEntry(wordsMerged, maxDocHitsMerged), e2, docHitsMerged,
							sameDocs, opt, matrix.hitsItemCount);
					int rank = ((Integer) re[0]).intValue();
					docHitsMerged = (long[]) re[1];
					sameDocs = ((Integer) re[2]).intValue();

					if (rank <= opt.getMinTagRelevance())
						continue;
					((List) rankMatrix.get(rank)).add(pairMerged);
					idPairTable.add(pairMerged);
				}

			}

		}
		// 删除被合并的主题词
		for (int m =0;m<tagListToRemove.size();m++){
			matrix.remove(tagListToRemove.get(m));
		}
		/**
		for (String w : tagListToRemove)
			matrix.remove(w);
		**/	
		// 添加合并而成的新主题词
		for (int n=0;n<entryListMerged.size();n++){
			TagHitEntry e = (TagHitEntry) entryListMerged.get(n);
			matrix.put(e.getKey(), e.getValue());
		}
		/**
		for (TagHitEntry e : entryListMerged) {
			if (e.getKey().length() > 0)
				matrix.put(e.getKey(), e.getValue());
		}
		**/
		return 0;
	}

	private int mergeClusters1(TagHitMatrix matrix, ClusteringOptions opt) {
		if (matrix.size() == 0)
			return 0;
		long[] docHitsMerged = (long[]) null;
		long[] maxDocHitsMerged = (long[]) null;
		int nMaxRank = 0;
		String word1 = "";
		String word2 = "";
		String word1ToMerge = "";
		String word2ToMerge = "";
		int sameDocs = 0;

		List matrix2List = new ArrayList();
		matrix2List.addAll(matrix.entrySet());

		for (int i1 = 0; i1 < matrix2List.size() - 1; i1++) {
			TagHitEntry hits1 = mapEntry2TagHitEntry((Map.Entry) matrix2List.get(i1));
			word1 = hits1.getKey();
			for (int i2 = i1 + 1; i2 < matrix2List.size(); i2++) {
				TagHitEntry hits2 = mapEntry2TagHitEntry((Map.Entry) matrix2List.get(i2));
				word2 = hits2.getKey();
				Object[] re = getWordsRelevance(hits1, hits2, docHitsMerged, sameDocs, opt, matrix.hitsItemCount);
				int nRank = ((Integer) re[0]).intValue();
				docHitsMerged = (long[]) re[1];
				sameDocs = ((Integer) re[2]).intValue();

				if ((nRank <= nMaxRank) || (nRank <= opt.getMinTagRelevance()))
					continue;
				nMaxRank = nRank;
				maxDocHitsMerged = docHitsMerged;
				word1ToMerge = word1;
				word2ToMerge = word2;
			}

		}

		if ((word1ToMerge.length() == 0) || (word2ToMerge.length() == 0)) {
			return 0;
		}

		String wordsMerged = word1ToMerge + "," + word2ToMerge;
		if ((nMaxRank > opt.getMinTagRelevance()) && (wordsMerged != "")) {
			matrix.remove(word1ToMerge);
			matrix.remove(word2ToMerge);
			matrix.put(wordsMerged, maxDocHitsMerged);
			LOG.debug("(" + word1ToMerge + ") - (" + word2ToMerge + ")");

			return 1;
		}

		return 0;
	}

	private Object[] getWordsRelevance(TagHitEntry entry1, TagHitEntry entry2, long[] docHitsMerged, int sameDocCount,
			ClusteringOptions opt, int hitsItemCount) {
		Object[] re = new Object[3];
		docHitsMerged = new long[hitsItemCount];
		sameDocCount = 0;

		String tag1 = entry1.getKey();
		String tag2 = entry2.getKey();
		assert (tag2 != tag1);

		long[] lDocs0 = entry1.getValue();
		long[] lDocs1 = entry2.getValue();
		int n0 = 0;
		int n1 = 0;
		n0 = getDocHitCount(lDocs0);
		n1 = getDocHitCount(lDocs1);
		int docCountMin = Math.min(n0, n1);
		int docCountMax = Math.max(n0, n1);
		int docCountMerged = 0;

		long sameDocBits = 0L;
		long diffDocBits = 0L;
		int diffDocCount = 0;
		for (int i = 0; i < lDocs0.length; i++) {
			docHitsMerged[i] = lDocs0[i] | lDocs1[i];// 获取合并的文档集
			docCountMerged += getValidBitCount(docHitsMerged[i]);
			diffDocBits = lDocs0[i] ^ lDocs1[i];// 获取不同的文档集
			diffDocCount += getValidBitCount(diffDocBits);
			sameDocBits = lDocs0[i] & lDocs1[i];// 获取相同的文档集
			sameDocCount += getValidBitCount(sameDocBits);
		}

		boolean IsSubstring = false;
		// 一个主题词是另一个的子串,则得分较高
		if ((tag2.contains(tag1)) || (tag1.contains(tag2))) {
			IsSubstring = true;
			docCountMin += opt.getTagMinDocCount();
		}

		if ((sameDocCount == 0) && (!IsSubstring)) {
			re[0] = Integer.valueOf(0);
			re[1] = docHitsMerged;
			re[2] = Integer.valueOf(sameDocCount);
			return re;
		}

		if (docCountMin < opt.getTagMinDocCount()) {
			re[0] = Integer.valueOf(0);
			re[1] = docHitsMerged;
			re[2] = Integer.valueOf(sameDocCount);
			return re;
		}

		int samePercent = (int) Math.round(sameDocCount * 100.0D / docCountMerged);
		int samePercentMin = (int) Math.round(sameDocCount * 100.0D / docCountMin);
		int diffPercent = (int) Math.round(diffDocCount * 100.0D / docCountMerged);
		LOG.debug("相关性:" + tag1 + "(" + n0 + ")-(" + n1 + ")" + tag2);
		LOG.debug(", SamePercent=" + samePercent);
		LOG.debug(", SamePercentMin=" + samePercentMin);
		LOG.debug(", DiffPercent=" + diffPercent);
		int nRank;
		if ((sameDocCount >= opt.getMinSameDocs())
				&& ((docCountMin < 10) || (samePercentMin >= opt.getMinSameDocPercent()))) {
			nRank = (int) Math.round((samePercentMin + samePercent) * 0.85D - diffPercent * 0.2D);
		} else {
			nRank = 0;
		}
		if (IsSubstring)
			nRank += 80;
		LOG.debug(", Rank=" + nRank);

		re[0] = Integer.valueOf(Math.min(nRank, 100));
		re[1] = docHitsMerged;
		re[2] = Integer.valueOf(sameDocCount);
		return re;
	}

	private TagHitEntry mapEntry2TagHitEntry(Map.Entry<String, long[]> e) {
		return new TagHitEntry((String) e.getKey(), (long[]) e.getValue());
	}

	@SuppressWarnings("unchecked")
	private List<DocCluster> createResult(TagHitMatrix matrix, ICTHit[] docs, int level, ClusteringOptions opt) {
		int i,j;
		Map<String,DocValue> clsIdList = new HashMap();
		List ClassTitleList = new ArrayList();
		for (Map.Entry de : matrix.entrySet()) {
			DocValue dv = new DocValue();
			clsIdList.put((String) de.getKey(), dv);
		}

		List<Integer> otherIdList = new ArrayList();
		TagHitEntry maxTagHitEntry = new TagHitEntry();
		int clsCount;
		String tag;
		// 确定每个文档所属的类别
		for (i = 0; i < docs.length; i++) {
			ICTHit d = docs[i];
			TagHitMatrix.ClusterDocInfo di = matrix.docs[i];
			assert (docs[i] != null);
			int maxTagHit = 0;
			clsCount = 0;

			for (Map.Entry hits : matrix.entrySet()) {
				int tagHitCount = 0;
				int score = 0;
				String clsWordListStr = "," + (String) hits.getKey() + ",";
				// 那个类别包含当前文档的主题词最多,该文档就属于哪个类别
				for (j = 0; j < di.TagCount; j++) {
					tag = di.TagList[j];
					score = j < 3 ? 2 : 1;
					assert (tag.length() > 0);
					if (!clsWordListStr.contains("," + tag + ","))
						continue;
					tagHitCount += score;
					clsCount++;
				}

				if (maxTagHit >= tagHitCount)
					continue;
				maxTagHit = tagHitCount;
				maxTagHitEntry = mapEntry2TagHitEntry(hits);
			}

			if (maxTagHit > 0) {
				DocValue dv = (DocValue) clsIdList.get(maxTagHitEntry.getKey());
				dv.idList.add(Integer.valueOf(i));
			} else {
				otherIdList.add(Integer.valueOf(i));
			}

		}
		// 生成类别列表
		List<DocCluster> clusterList = new ArrayList();
		String[] TagList;
		Object dc;
		for (Map.Entry<String,DocValue> kv : clsIdList.entrySet()) {
			DocValue dv = (DocValue) kv.getValue();
			if (dv.idList.size() <= 0)
				continue;
			if (dv.idList.size() == 1) {
				otherIdList.add((Integer) dv.idList.get(0));
			} else {
				dc = new DocCluster();
				((DocCluster) dc).setDocIdList(new String[dv.idList.size()]);
				((DocCluster) dc).setDocList(new ICTHit[dv.idList.size()]);
				for (i = 0; i < dv.idList.size(); i++) {
					((DocCluster) dc).getDocIdList()[i] = docs[((Integer) dv.idList.get(i)).intValue()].getDocId();
					((DocCluster) dc).getDocList()[i] = docs[((Integer) dv.idList.get(i)).intValue()];
				}
				((DocCluster) dc).setLevel(level);
				((DocCluster) dc).setTags((String) kv.getKey());

				for (i = 0; (i < clusterList.size())
						&& (((DocCluster) dc).getDocIdList().length <= ((DocCluster) clusterList.get(i)).getDocIdList().length);) {
					i++;
				}
				clusterList.add(i, (DocCluster) dc);
			}
		}
		for (i = opt.getMaxClusterCount(); i < clusterList.size();) {
			DocCluster c = (DocCluster) clusterList.get(i);
			List idList = ((DocValue) clsIdList.get(c.getTags())).idList;
			for (dc = idList.iterator(); ((Iterator) dc).hasNext();) {
				int idx = ((Integer) ((Iterator) dc).next()).intValue();
				otherIdList.add(Integer.valueOf(idx));
			}
			clusterList.remove(i);
		}
		int i1;
		for (i = 0; i < clusterList.size(); i++) {
			DocCluster dc1 = (DocCluster) clusterList.get(i);
			String[] tagList = dc1.getTags().split(",");
			String newTags = "";

			for (j = 0; j < tagList.length; j++) {
				i1 = dc1.getTags().indexOf(tagList[j]);
				int i2 = dc1.getTags().lastIndexOf(tagList[j]);
				if (i1 == i2)
					newTags = newTags + tagList[j] + ",";
			}
			if ((newTags.trim().length() > 0) && (newTags.endsWith(","))) {
				newTags = newTags.substring(0, newTags.length() - 1);
			}
			dc1.setTags(newTags);

			dc1.setTitle("");

			if (this.useTagsAsTitle) {
				tagList = dc1.getTags().split(",");
				for (j = 0; (tagList != null) && (j < tagList.length); j++) {
					if ((dc1.getTitle() + tagList[j]).length() > 16)
						break;
					boolean isSubstr = false;
					for (DocCluster c : clusterList) {
						if ((c.getTitle().length() <= 0)
								|| ((!c.getTitle().contains(tagList[j])) && (!tagList[j].contains(c.getTitle()))))
							continue;
						isSubstr = true;
						break;
					}
					if (!isSubstr)
						dc1.setTitle(dc1.getTitle() + tagList[j] + ",");
				}
				if ((dc1.getTitle().trim().length() > 0) && (dc1.getTitle().endsWith(","))) {
					dc1.setTitle(dc1.getTitle().substring(0, dc1.getTitle().length() - 1));
				}

			}

			if (dc1.getTitle() != "")
				continue;
			dc1.setTitle(dc1.getTags());
			if (dc1.getTitle().length() <= 16)
				continue;
			String s = dc1.getTitle().substring(0, 16);
			int li = s.lastIndexOf(',');
			if (li > 0) {
				dc1.setTitle(s.substring(0, li));
			}

		}

		if (otherIdList.size() > 0) {
			DocCluster clusterOther = new DocCluster();
			clusterOther.setDocIdList(new String[otherIdList.size()]);
			clusterOther.setDocList(new ICTHit[otherIdList.size()]);
			clusterOther.setLevel(level);
			clusterOther.setTitle("其他");
			clusterOther.setTags("其他");
			i = 0;
			for (int k=0;k<otherIdList.size();k++) {
				int idx = otherIdList.get(k);

				clusterOther.getDocIdList()[i] = docs[idx].getDocId();
				clusterOther.getDocList()[i] = docs[idx];
				i++;
			}
			clusterList.add(clusterOther);
		}

		return (List<DocCluster>) clusterList;
	}

	public List<DocCluster> getClusters() {
		return this.clusters;
	}

	public void setClusters(List<DocCluster> clusters) {
		this.clusters = clusters;
	}

	public ICTHit[] getDocs() {
		return this.docs;
	}

	public void setDocs(ICTHit[] docs) {
		this.docs = docs;
	}

	public int getMaxLevels() {
		return this.maxLevels;
	}

	public void setMaxLevels(int maxLevels) {
		this.maxLevels = maxLevels;
	}

	public ClusteringOptions[] getOptions() {
		return this.options;
	}

	public void setOptions(ClusteringOptions[] options) {
		this.options = options;
	}

	public boolean isUseTagsAsTitle() {
		return this.useTagsAsTitle;
	}

	public void setUseTagsAsTitle(boolean useTagsAsTitle) {
		this.useTagsAsTitle = useTagsAsTitle;
	}

	public String getWordsExcluded() {
		return this.wordsExcluded;
	}

	public void setWordsExcluded(String wordsExcluded) {
		this.wordsExcluded = wordsExcluded;
	}

	private class DocValue {
		public List<Integer> idList = new ArrayList();
		public String titleListStr = "";

		private DocValue() {
		}
	}
	/**
	 * 主题词ID对,主题词ID为该主题词在主题词文档映射表中的主键位置。
	* @author 
	* @version 创建时间:2011-3-9 下午02:52:44
	 */
	private class IdPair {
		public int Id1;
		public int Id2;

		public IdPair(int id1, int id2) {
			assert (id1 != id2);
			if (id1 < id2) {
				this.Id1 = id1;
				this.Id2 = id2;
			} else {
				this.Id1 = id2;
				this.Id2 = id1;
			}
		}

		public int hashCode() {
			return -1;
		}

		public boolean equals(Object o) {
			return (((IdPair) o).Id1 == this.Id1) && (((IdPair) o).Id2 == this.Id2);
		}
	}

	public static class TagHitEntry {
		public String key;
		public long[] value;

		public TagHitEntry() {
		}

		public TagHitEntry(String k, long[] v) {
			this.key = k;
			this.value = v;
		}

		public String getKey() {
			return this.key;
		}

		public long[] getValue() {
			return this.value;
		}
	}
}

 

   ClusteringOptions.java

 

 

/**
 * 
* @author 
* @version 创建时间:2011-3-8 上午10:23:27
 */
public class ClusteringOptions {
	public static int DefMaxClusterCount = 20;
	public static int DefMaxKeywordCount = 6;
	public static int DefMinWordsRelevance = 10;
	public static int DefTagMinDocCount = 3;
	public static int DefIgnoreSameDocs = 2;
	public static int DefSameDocPercent = 50;
	public static int DefMinDocsToCluster = 8;
	private int docMaxTagCount;
	private int maxClusterCount;
	private int minDocsToCluster;
	private int minSameDocPercent;
	private int minSameDocs;
	private int minTagRelevance;
	private int tagMinDocCount;

	public ClusteringOptions() {
		this.maxClusterCount = DefMaxClusterCount;
		this.minTagRelevance = DefMinWordsRelevance;
		this.tagMinDocCount = DefTagMinDocCount;
		this.minSameDocs = DefIgnoreSameDocs;
		this.minSameDocPercent = DefSameDocPercent;
		this.docMaxTagCount = DefMaxKeywordCount;
		this.minDocsToCluster = DefMinDocsToCluster;
	}

	public int getDocMaxTagCount() {
		return this.docMaxTagCount;
	}

	public void setDocMaxTagCount(int docMaxTagCount) {
		this.docMaxTagCount = docMaxTagCount;
	}

	public int getMaxClusterCount() {
		return this.maxClusterCount;
	}

	public void setMaxClusterCount(int maxClusterCount) {
		this.maxClusterCount = maxClusterCount;
	}

	public int getMinDocsToCluster() {
		return this.minDocsToCluster;
	}

	public void setMinDocsToCluster(int minDocsToCluster) {
		this.minDocsToCluster = minDocsToCluster;
	}

	public int getMinSameDocPercent() {
		return this.minSameDocPercent;
	}

	public void setMinSameDocPercent(int minSameDocPercent) {
		this.minSameDocPercent = minSameDocPercent;
	}

	public int getMinSameDocs() {
		return this.minSameDocs;
	}

	public void setMinSameDocs(int minSameDocs) {
		this.minSameDocs = minSameDocs;
	}

	public int getMinTagRelevance() {
		return this.minTagRelevance;
	}

	public void setMinTagRelevance(int minTagRelevance) {
		this.minTagRelevance = minTagRelevance;
	}

	public int getTagMinDocCount() {
		return this.tagMinDocCount;
	}

	public void setTagMinDocCount(int tagMinDocCount) {
		this.tagMinDocCount = tagMinDocCount;
	}
}

 

 

 

   DocCluster.java

 

/**
 * 
* @author
* @version 创建时间:2011-3-8 上午10:23:35
 */
public class DocCluster {
	private String[] docIdList;
	private ICTHit[] docList;
	private int level;
	private List<DocCluster> subclusters;
	private String tags;
	private String title;

	public String[] getDocIdList() {
		return this.docIdList;
	}

	public void setDocIdList(String[] docIdList) {
		this.docIdList = docIdList;
	}

	public ICTHit[] getDocList() {
		return this.docList;
	}

	public void setDocList(ICTHit[] docList) {
		this.docList = docList;
	}

	public int getLevel() {
		return level;
	}

	public void setLevel(int level) {
		this.level = level;
	}

	public List<DocCluster> getSubclusters() {
		return this.subclusters;
	}

	public void setSubclusters(List<DocCluster> subclusters) {
		this.subclusters = subclusters;
	}

	public String getTags() {
		return this.tags;
	}

	public void setTags(String tags) {
		this.tags = tags;
	}

	public String getTitle() {
		if (title == null)
			title = "";
		return this.title;
	}

	public void setTitle(String title) {
		this.title = title;
	}
}

 ICTHit.java

 

 

public class ICTHit implements Serializable {
	/*
	 * 关键词数组
	 */
	private String[] TagList;
	private String docId;
	private String title;

	public String[] getTagList() {
		return TagList;
	}

	public void setTagList(String[] tagList) {
		TagList = tagList;
	}

	public String getDocId() {
		return docId;
	}

	public void setDocId(String docId) {
		this.docId = docId;
	}

	public String getTitle() {
		return title;
	}

	public void setTitle(String title) {
		this.title = title;
	}	
	
}
 

 

TagHitMatrix.java

 

 

 

 

public class TagHitMatrix extends LinkedHashMap<String, long[]> {
	/**
	 * 
	 */
	private static final long serialVersionUID = -7511464445378974433L;
	public static int ii = 0;
	public ClusterDocInfo[] docs;
	public int hitsItemCount;

	public TagHitMatrix(int DocCount, int MaxTagCount) {
		this.hitsItemCount = (int) (DocCount / 62.0D + 0.984375D);
		this.docs = new ClusterDocInfo[DocCount];

		for (int i = 0; i < this.docs.length; i++)
			this.docs[i] = new ClusterDocInfo(MaxTagCount);
	}

	public void AddDocHit(String TagStr, int Position) {
		TagStr = TagStr.trim();

		int n = Position / 62;
		int m = Position % 62;
		long[] DocHits = (long[]) get(TagStr);
		if (DocHits == null) {
			DocHits = new long[this.hitsItemCount];
			put(TagStr, DocHits);
		}
		DocHits[n] |= Math.round(Math.pow(2.0D, m));
		ClusterDocInfo di = this.docs[Position];
		di.TagList[(di.TagCount++)] = TagStr;
	}

	class ClusterDocInfo {
		public String[] TagList;
		public int TagCount;

		public ClusterDocInfo(int MaxTagCount) {
			this.TagList = new String[MaxTagCount];
			this.TagCount = 0;
		}
	}
}

 

 

   测试方法:

 

 

public void test(ICTHit[] icthits) throws IOException {
		ClusterBuilder clusterBuilder = new ClusterBuilder();
		// 设置需要聚类的数据集合,测试中用的null。
		clusterBuilder.setDocs(icthits);
		// 设置聚类级别,只使用1级
		clusterBuilder.setMaxLevels(10);
		clusterBuilder.setUseTagsAsTitle(true);
		// 一般将检索词设置为wordsExcluded
		clusterBuilder.setWordsExcluded("万美元,日本,公司,视频,北京时间,图文,新华网,新浪,消息,通讯,互联网,美国,中国");
		clusterBuilder
				.setOptions(new ClusteringOptions[] { new ClusteringOptions(),new ClusteringOptions() });

		// 开始聚类
		clusterBuilder.cluster();
		FileWriter fw1 = new FileWriter("c:/today-20110509-cluster.txt ", true);
		BufferedWriter bw1 = new BufferedWriter(fw1);

		// 打印结果
		if (clusterBuilder.getClusters() != null) {
			int i = 0;
			for (DocCluster docCluster : clusterBuilder.getClusters()) {
				i++;
				System.out.println("tag:" + docCluster.getTags() + "("
						+ docCluster.getDocIdList().length + ")");
				bw1.write(docCluster.getTags() + "("+ docCluster.getDocIdList().length + ")"+"\r\n ");				
				
				if (docCluster.getDocList() != null
						&& docCluster.getDocList().length > 0) {
					for (ICTHit co : docCluster.getDocList()) {
						System.out.println("     DocID: " + co.getDocId());
						bw1.write("标题为: "	+ co.getTitle()+",ID为"+co.getDocId()+"\r\n ");	
						for (int m = 0; m < co.getTagList().length; m++) {							
							bw1.write("标题为: "	+ co.getTitle()+",ID为"+co.getDocId()+"\r\n ");	
							System.out.println("     Key Word: "
									+ co.getTagList()[m]);
						}
						System.out.println("");
					}
					System.out.println("");
				} else {
					bw1.write("      该分类下无数据!"+"\r\n ");	
				}
				bw1.write("-------------------------------------------------------------------------------\r\n");
			}
		}
		bw1.close();
		fw1.close();
	}
 

 如上方法可以,是一个示例性的,没有用在生产当中。核心方法有了。大家可以引用到项目当中。效果比carrot2标准的方法要好很多。

分享到:
评论
2 楼 hui94781674 2013-04-12  
您好,可以把代码细路说详细点么? 看了代码不是很明白。。
1 楼 hui94781674 2013-04-12  
感谢分享,正做这方面的研究。

相关推荐

    Kmeans文本聚类java实现

    在Java环境中实现KMeans算法进行文本聚类,可以为大数据分析、信息检索和推荐系统等应用场景提供有力支持。 KMeans算法的基本思想是通过迭代过程,不断调整样本的归属,使得同一簇内的样本尽可能接近,不同簇间的...

    计算机研究 -后缀树及其在中文文本聚类中的应用探索.pdf

    在研究的过程中,本人首先实现了将中文文本生成后缀树结构的演示程序,然后用 java 实现了针对单篇文档的高频短语发现器和针对多篇文档的公共短语发现器,并以主题相似的文档集、不同主题的文档集和主题一致的网页...

    凝聚层次聚类算法JAVA实现.docx

    凝聚层次聚类算法的实现相对复杂,但下面是一个简化的JAVA代码示例,可以帮助理解其实现的基本思路: ```java import java.util.ArrayList; import java.util.List; public class AgglomerativeClustering { ...

    dbscan.rar_DBSCAN_DBScan java_clustering in jsp_dbscan.zip_聚类 ja

    3. 执行DBSCAN:遍历数据集中的每个点,对每个点执行密度可达性判断和聚类过程。 4. 输出结果:将聚类结果以适当的形式返回,如颜色编码的点图、表格或其他可视化方式。 在实际应用中,DBSCAN有以下优势: - 自适应...

    doc2vec的java实现

    虽然原生的Gensim是用Python编写的,但通过Jython(Python的Java实现)或其他Java调用Python库的方法,可以在Java项目中使用Gensim的DOC2Vec功能。 **步骤详解** 1. **数据预处理**:首先,你需要对输入的文本进行...

    hadoop kmeans中文网页聚类 实验文档

    根据提供的测试数据,可以通过提取网页正文得到文档(document),再通过分词获取文档中的词汇,从而将聚类问题转化为文本聚类问题。考虑到中文的特殊性,实验采用了Unicode编码范围(\u4e00-\u9fa5)来提取所有的汉字...

    人工智能-LDA-主题分析-LDA(潜在狄利克雷分配)的 Java 实现, 用几行 Java 代码从一组文档中推断主题

    总之,LDA作为一种强大的文本挖掘工具,通过Java实现能够高效处理大量文本数据,揭示文档中的潜在主题。LDA4j等Java库的出现,使得开发人员能够轻松地将主题分析技术融入到各种应用场景中,如新闻分类、用户兴趣分析...

    java利用matlib导出的jar包计算特征值

    2. **编译Matlab代码**:使用Matlab的` mcc -m`命令将Matlab代码编译为Java类,生成相应的jar包。 3. **在Eclipse中导入jar包**:将生成的jar包添加到Eclipse项目的类路径中。 4. **编写Java代码**:在Java代码中...

    java版本的EM

    3. **Kmeans.java**:这是K-means算法的实现,用于生成EM算法的初始均值。K-means通过迭代找到数据的最佳聚类中心,这些中心被用作EM算法的起始点。 4. **TextDataIO.java**:这个文件可能是用于读取和处理文本数据...

    62种常见算法(JAVA,C实现都有)

    "62种常见算法(JAVA,C实现都有)"这个资源集合提供了一套丰富的算法实现,涵盖了多种基础到进阶的算法,对于学习者和开发者来说,这是一个宝贵的资料库。下面将详细讨论这些算法以及它们在Java和C语言中的实现。 ...

    GetFileTimes.rar_IF-IDF_TF_java TF-IDF_tf idf_tf idf java

    在压缩包内的文件中,"GetFileTimes.java"很可能是实现TF-IDF算法的主要源代码文件,可能包括读取文本、计算词频、计算IDF值以及生成输出等功能。而"www.pudn.com.txt"则可能是一个示例文本文件,用于测试代码,这个...

    优质课件 北京大学研究生课程文本挖掘 文本数据挖掘全套PPT教程(共67页) TextMining14-文本挖掘工具与应用.ra

    10. **文本生成**:近年来,随着深度学习的发展,自动生成文本的技术也取得了显著进步,如使用seq2seq模型、transformer等。 11. **文本挖掘工具**:课程可能会介绍一些常用工具,如NLTK、Spacy(Python)、Gensim...

    基于关键词提取的矩阵生成程序

    标题“基于关键词提取的矩阵生成程序”涉及到的是一种在文本分析领域常见的技术,它通过自动提取文本中的关键词并构造对称矩阵,以便于进行文本相似性分析或主题挖掘。这样的程序通常用于信息检索、自然语言处理...

    机器学习/深度学习、数据分析/风险控制/数据挖掘和算法 文本分类、信息提取和文本生成

    在文本处理中,这些算法可以用于特征选择、模型训练和参数调优,以实现更高效和准确的模型。 在文件"charent-main"中,可能包含了关于这些主题的代码、模型、数据集或其他相关资源,供研究者和开发者使用,以进一步...

    LDA.rar_LDA 文档主题_java LDA_lda_lda java_lda模型

    Gensim虽然主要以Python实现,但也有Java接口,允许Java开发者利用其强大的文本处理能力。 **LDA模型的步骤** 1. **数据预处理**:首先,需要对输入的文本数据进行清洗,包括去除标点符号、数字、特殊字符,转换为...

    Lingpipe(自然语言处理的Java开源工具包)入门

    例如,在文本挖掘和信息检索方面,Lingpipe能够帮助用户从海量文本数据中提取有用信息,实现自动化的关键词提取、主题分类、聚类等任务;在情感分析方面,Lingpipe可以分析网络评论、社交媒体帖子等文本的情绪倾向,...

    Latent Dirichlet Allocation 的java代码

    3. **Java实现LDA**: - **数据结构**:在Java中,可能需要使用ArrayLists、Maps等数据结构存储词语、文档、主题分配等信息。 - **Gibbs采样**:编写Gibbs采样算法的Java代码,处理每个文档中的词语,更新主题分配...

    K-shingle算法实现

    4. **文本分类与聚类**:在机器学习中,K-shingle可以作为特征,帮助进行文本的分类和聚类。 五、K-shingle算法的实现 在提供的"K-shingle"压缩包文件中,可能包含了实现K-shingle算法的相关代码。通常,这些代码会...

    通用数据挖掘系统 V3.0DMS JAVA代码.zip

    3. 数据挖掘算法:V3.0DMS可能集成了多种数据挖掘算法,如K-means聚类、Apriori关联规则、决策树(如C4.5、ID3)、随机森林、支持向量机(SVM)等。 4. 可视化界面:为了便于用户操作和理解,系统可能有一个直观的...

    源码地java spark淘宝大数据分析可视化系统(源码+数据+报告)址.zip

    报告可能结合Markdown或LaTeX等工具,实现格式化的文本输出。 8. 源码分享:本项目提供源码,有利于学习者深入理解系统架构和实现细节,也便于其他开发者在此基础上进行二次开发或定制化需求。 总之,这个基于Java...

Global site tag (gtag.js) - Google Analytics