6 CTX_CLS Package

TRAIN

Use this procedure to generate query rules that select document categories. You must supply a training set consisting of categorized documents. Documents can be in any format supported by Oracle Text and must belong to one or more categories. This procedure generates the queries that define the categories and then writes the results to a table.

You must also have a document table and a category table. The category table must contain at least two categories.

For example, your document and category tables can be defined as:

create table trainingdoc(

docid number primary key,
text varchar2(4000));

create table category (

docid trainingdoc(docid),
categoryid number);

You can use one of two syntaxes depending on the classification algorithm you need. The query compatible syntax uses the RULE_CLASSIFIER preference and generates rules as query strings. The support vector machine syntax uses the SVM_CLASSIFER preference and generates rules in binary format. The SVM_CLASSIFIER is good for high classification accuracy, but because its rules are generated in binary format, they cannot be examined like the query strings generated with the RULE_CLASSIFIER. Note that only those document ids that appear in both the document table and the category table will impact RULE_CLASSIFIER and SVM_CLASSIFIER learning.

The CTX_CLS.TRAIN procedure requires that your document table have an associated context index. For best results, the index should be synchronized before running this procedure. SVM_CLASSIFIER syntax enables the use of an unpopulated context index, while query-compatible syntax requires that the context index be populated.

Query Compatible Syntax

The following syntax generates query-compatible rules and is used with the RULE_CLASSIFIER preference. Use this syntax and preference when different categories are separated from others by several key words. An advantage of generating your rules as query strings is that you can easily examine the generated rules. This is different from generating SVM rules, which are in binary format.

CTX_CLS.TRAIN(

index_name in varchar2,
doc_id in varchar2,
cattab in varchar2,
catdocid in varchar2,
catid in varchar2,
restab in varchar2,
rescatid in varchar2,
resquery in varchar2,
resconfid in varchar2,
preference_name in varchar2 DEFAULT NULL

);

index_name

Specify the name of the context index associated with your document training set.

doc_id

Specify the name of the document id column in the document table. This column must contain unique document ids. This column must a NUMBER.

cattab

Specify the name of the category table. You must have SELECT privilege on this table.

catdocid

Specify the name of the document id column in the category table. The document ids in this table must also exist in the document table. This column must a NUMBER.

catid

Specify the name of the category ID column in the category table. This column must a NUMBER.

restab

Specify the name of the result table. You must have INSERT privilege on this table.

rescatid

Specify the name of the category ID column in the result table. This column must a NUMBER.

resquery

Specify the name of the query column in the result table. This column must be VARACHAR2, CHAR CLOB, NVARCHAR2, or NCHAR.

The queries generated in this column connects terms with AND or NOT operators, such as:

'T1 & T2 ~ T3'

Terms can also be theme tokens and be connected with the ABOUT operator, such as:

'about(T1) & about(T2) ~ about(T3)'

Generated rules also support WITHIN queries on field sections.

resconfid

Specify the name of the confidence column in result table. This column contains the estimated probability from training data that a document is relevant if that document satisfies the query.

preference_name

Specify the name of the preference. For classifier types and attributes, see "Classifier Types" in Chapter 2, " Oracle Text Indexing Elements".

Syntax for Support Vector Machine Rules

The following syntax generates support vector machine (SVM) rules with the SVM_CLASSIFIER preference. This preference generates rules in binary format. Use this syntax when your application requires high classification accuracy.

CTX_CLS.TRAIN(
    index_name in varchar2,    
    docid      in varchar2,         
    cattab     in varchar2, 
    catdocid   in varchar2, 
    catid      in varchar2,  
    restab      in varchar2,  
    preference_name in varchar2 );

index_name

Specify the name of the text index.

docid

Specify the name of docid column in document table.

cattab

Specify the name of category table.

catdocid

Specify the name of docid column in category table.

catid

Specify the name of category ID column in category table.

restab

Specify the name of result table.

The result table has the following format:

Column Name	Datatype	Description
CAT_ID	NUMBER	The ID of the category.
TYPE	NUMBER(3) NOT NULL	0 for the actual rule or catid; 1 for other.
RULE	BLOB	The returned rule.

preference_name

Specify the name of user preference. For classifier types and attributes, see "Classifier Types" in Chapter 2, " Oracle Text Indexing Elements".

Example

The CTX_CLS.TRAIN procedure is used in supervised classification. For an extended example, see the Oracle Text Application Developer's Guide.

CLUSTERING

Use this procedure to cluster a collection of documents. A cluster is a group of documents similar to each other in content. Clustering is also known as unsupervised classification.

Given a set of documents, this procedure assigns each document into a cluster according to the similarity with documents already in the cluster. The result is that documents in a cluster are more similar to one another than documents across different clusters. The more clusters produced, the greater the accuracy and quality of each cluster; however, producing more clusters requires more computing time.

Cluster output may be flat or hierarchical. Hierarchical clustering affords greater specificity of each cluster; however, it may require more computing power. In the case where you want to produce only a few clusters, non-hierarchical clustering may suffice.

A clustering result set is composed of document assignments and cluster descriptions. The document assignment result set contains information about the cluster to which the procedure assigned a document, and how similar the document is to the assigned cluster. This result set contains document identification, cluster identification, and similarity score between the cluster and assigned document.

The cluster description result set contains information about what topic a generated cluster is about. This result set contains cluster identification, cluster description text, suggested cluster label, number of documents assigned, and a quality score of the cluster.

There are two versions of this procedure: one with a table result set, and one with an in-memory result set.

Syntax: Table Result Set

ctx_cls.clustering (
 index_name  IN VARCHAR2, 
 docid       IN VARCHAR2, 
 doctab_name IN VARCHAR2, 
 clstab_name IN VARCHAR2, 
 pref_name   IN VARCHAR2  DEFAULT NULL
);

index_name

Specify the name of the context index on collection table.

docid

Specify the name of document ID column of the collection table.

doctab_name

Specify the name of document assignment table. This procedure creates the table with the following structure:

doc_assign(
   docid number,
   clusterid number,
   score number
);

Column	Description
DOCID	Document ID to identify document.
CLUSTERID	ID of the cluster the document is assigned to. If CLUSTERID is -1, then the cluster contains "miscellaneous" documents; for example, documents that cannot be assigned to any other cluster category.
SCORE	The associated score between the document and cluster.

If you require more columns, you can create the table before you call this procedure.

clstab_name

Specify the name of the cluster description table. This procedure creates the table with the following structure:

cluster_desc(
  clusterid NUMBER,
  descript VARCHAR2(4000),
  label VARCHAR2(200),
  sze NUMBER,
  quality_score NUMBER,
  parent NUMBER
);

Column	Description
CLUSTERID	Cluster ID to identify cluster. If CLUSTERID is -1, then the cluster contains "miscellaneous" documents; for example, documents that cannot be assigned to any other cluster category.
DESCRIPT	String to describe the cluster.
LABEL	A suggested label for the cluster.
SZE	Number of documents assigned to this cluster.
QUALITY_SCORE	The quality score of the cluster, higher is better.
PARENT	The parent cluster id. A negative number means no parent cluster.

If you require more columns, you can create the table before you call this procedure.

pref_name

Specify the name of the preference.

Syntax: In-Memory Result Set

You can put the result set into in-memory structures for better performance. Two in-memory tables are defined in CTX_CLS package for document assignment and cluster description respectively.

CTX_CLS.CLUSTERING(
  index_name     IN VARCHAR2, 
  docid          IN VARCHAR2,
  dids           IN DOCID_TAB,
  doctab_name    IN OUT NOCOPY DOC_TAB,
  clstab_name    IN OUT NOCOPY CLUSTER_TAB,
  pref_name      IN VARCHAR2  DEFAULT NULL 
          );

index_name

Specify the name of context index on the collection table.

docid

Specify the document id column of the collection table.

dids

Specify the name of the in-memory docid_tab.

TYPE docid_tab IS TABLE OF number INDEX BY BINARY_INTEGER;

doctab_name

Specify name of the document assignment in-memory table. This table is defined as follows:

TYPE doc_rec IS RECORD (
   docid NUMBER,
   clusterid NUMBER,
   score NUMBER
)
TYPE doc_tab IS TABLE OF doc_rec INDEX BY BINARY_INTEGER;

Column	Description
DOCID	Document ID to identify document.
CLUSTERID	ID of the cluster the document is assigned to If CLUSTERID is -1, then the cluster contains "miscellaneous" documents; for example, documents that cannot be assigned to any other cluster category.
SCORE	The associated score between the document and cluster.

cls_tab

Specify the name of cluster description in-memory table

TYPE cluster_rec IS RECORD(
     clusterid NUMBER,
     descript VARCHAR2(4000),
     label VARCHAR2(200),
     sze NUMBER,
     quality_score NUMBER,
     parent NUMBER
);
TYPE cluster_tab IS TABLE OF cluster_rec INDEX BY BINARY_INTEGER;

Column	Description
CLUSTERID	Cluster ID to identify cluster. If CLUSTERID is -1, then the cluster contains "miscellaneous" documents; for example, documents that cannot be assigned to any other cluster category.
DESCRIPT	String to describe the cluster.
LABEL	A suggested label for the cluster.
SZE	Number of documents assigned to this cluster.
QUALITY_SCORE	The quality score of the cluster, higher is better.
PARENT	The parent cluster id. A negative number means no parent cluster.

pref_name

Specify the name of the preference. For cluster types and attributes, see "Cluster Types" in Chapter 2, " Oracle Text Indexing Elements".

Example