This chapter includes the following topics:
A major problem facing businesses and institutions today is that of information overload. Sorting out useful documents from documents that are not of interest challenges the ingenuity and resources of both individuals and organizations.
One way to sift through numerous documents is to use keyword search engines. However, keyword searches have limitations. One major drawback is that keyword searches don't discriminate by context. In many languages, a word or phrase may have multiple meanings, so a search may result in many matches that are not on the desired topic. For example, a query on the phrase river bank might return documents about the Hudson River Bank & Trust Company, because the word bank has two meanings.
An alternative strategy is to have human beings sort through documents and classify them by content, but this is not feasible for very large volumes of documents.
Oracle Text offers various approaches to document classification. Under rule-based classification, you write the classification rules yourself. With supervised classification, Oracle Text creates classification rules based on a set of sample documents that you pre-classify. Finally, with unsupervised classification (also known as clustering), Oracle Text performs all the steps, from writing the classification rules to classifying the documents, for you.
Oracle Text enables you to build document classification applications. A document classification application performs some action based on document content. Actions include assigning category ids to a document for future lookup or sending a document to a user. The result is a set or stream of categorized documents. Figure 6-1 illustrates how the classification process works.
Oracle Text enables you to create document classification applications in different ways. This chapter defines a typical classification scenario and shows how you can use Oracle Text to build a solution.
Figure 6-1 Overview of a Document Classification Application
Oracle Text enables you to classify documents in the following ways:
Rule-Based Classification. In rule-based classification, you group your documents together, decide on categories, and formulate the rules that define those categories; these rules are actually query phrases. You then index the rules and use the MATCHES
operator to classify documents.
Advantage: Rule-based classification is very accurate for small document sets. Results are always based on what you define, since you write the rules.
Disadvantages: Defining rules can be tedious for large document sets with many categories. As your document set grows, you may need to write correspondingly more rules.
Supervised Classification. This method is similar to rule-based classification, but the rule writing step is automated with CTX_CLS.TRAIN
. CTX_CLS.TRAIN
formulates a set of classification rules from a sample set of pre-classified documents that you provide. As with rule-based classification, you use MATCHES
operator to classify documents.
Oracle Text offers two versions of supervised classification, one using the RULE_CLASSIFIER
preference and one using the SVM_CLASSIFIER
preference. These are discussed in "Supervised Classification".
Advantage: Rules are written for you automatically. This is useful for large document sets.
Disadvantages:
You must assign documents to categories before generating the rules.
Rules may not be as specific or accurate as those you write yourself.
Unsupervised Classification (Clustering). All steps from grouping your documents to writing the category rules are automated with CTX_CLS.CLUSTERING
. Oracle Text statistically analyzes your document set and correlates them with clusters according to content.
Advantages:
You don't need to provide either the classification rules or the sample documents as a training set.
Helps to discover patterns and content similarities in your document set that you might overlook.
In fact, you can use unsupervised classification when you do not have a clear idea of rules or classifications. One possible scenario is to use unsupervised classification to provide an initial set of categories, and to subsequently build on these through supervised classification.
Disadvantages:
Clustering might result in unexpected groupings, since the clustering operation is not user-defined, but based on an internal algorithm.
You do not see the rules that create the clusters.
The clustering operation is CPU-intensive and can take at least the same time as indexing.
Rule-based classification (sometimes called "simple classification") is the basic way of creating an Oracle Text classification application.
The basic steps for rule-based classification are as follows. Specific steps are explored in greater detail in the example.
Create a table for the documents to be classified, and populate it.
Create a rule table (also known as a category table). The rule table consists of categories that you name, such as "medicine" or "finance," and the rules that sort documents into those categories.
These rules are actually queries. For example, you might define the "medicine" category as consisting of documents that include the words "hospital," "doctor," or "disease," so you would set up a rule of the form "hospital OR doctor OR disease." See "CTXRULE Parameters and Limitations" for information on which operators are allowed for queries.
Create a CTXRULE
index on the rule table.
Classify the documents.
In this example, we gather news articles on different subjects and then classify them.
Once our rules are created, we can index them and then use the MATCHES
statement to classify documents. The steps are as follows:
Step 1 Create schema
We create the tables to store the data. The news_table
stores the documents to be classified. The news_categories
table stores the categories and rules that define our categories. The news_id_cat
table stores the document ids and their associated categories after classification.
create table news_table ( tk number primary key not null, title varchar2(1000), text clob); create table news_categories ( queryid number primary key not null, category varchar2(100), query varchar2(2000)); create table news_id_cat ( tk number, category_id number);
Step 2 Load Documents with SQLLDR
In this step, we load the HTML news articles into the news_table
using the SQLLDR
program. The filenames and titles are read from loader.dat.
LOAD DATA INFILE 'loader.dat' INTO TABLE news_table REPLACE FIELDS TERMINATED BY ';' (tk INTEGER EXTERNAL, title CHAR, text_file FILLER CHAR, text LOBFILE(text_file) TERMINATED BY EOF)
Step 3 Create Categories
In this step, we define our categories and write the rules for each of our categories.
Defined Categories: | ||
---|---|---|
United States | Europe | Middle East |
Asia | Africa | Conflicts |
Finance | Technology | Consumer Electronics |
Latin America | World Politics | U.S. Politics |
Astronomy | Paleontology | Health |
Natural Disasters | Law | Music News |
A rule is a query that selects documents for the category. For example, the category 'Asia' has a rule of 'China or Pakistan or India or Japan'. We insert our rules in the news_categories
table as follows:
insert into news_categories values (1,'United States','Washington or George Bush or Colin Powell'); insert into news_categories values (2,'Europe','England or Britain or Germany'); insert into news_categories values (3,'Middle East','Israel or Iran or Palestine'); insert into news_categories values(4,'Asia','China or Pakistan or India or Japan'); insert into news_categories values(5,'Africa','Egypt or Kenya or Nigeria'); insert into news_categories values (6,'Conflicts','war or soliders or military or troops'); insert into news_categories values(7,'Finance','profit or loss or wall street'); insert into news_categories values (8,'Technology','software or computer or Oracle or Intel or IBM or Microsoft'); insert into news_categories values (9,'Consumer electronics','HDTV or electronics'); insert into news_categories values (10,'Latin America','Venezuela or Colombia or Argentina or Brazil or Chile'); insert into news_categories values (11,'World Politics','Hugo Chavez or George Bush or Tony Blair or Saddam Hussein or United Nations'); insert into news_categories values (12,'US Politics','George Bush or Democrats or Republicans or civil rights or Senate or White House'); insert into news_categories values (13,'Astronomy','Jupiter or Earth or star or planet or Orion or Venus or Mercury or Mars or Milky Way or Telescope or astronomer or NASA or astronaut'); insert into news_categories values (14,'Paleontology','fossils or scientist or paleontologist or dinosaur or Nature'); insert into news_categories values (15,'Health','stem cells or embryo or health or medical or medicine or World Health Organization or AIDS or HIV or virus or centers for disease control or vaccination'); insert into news_categories values (16,'Natural Disasters','earthquake or hurricane or tornado'); insert into news_categories values (17,'Law','abortion or Supreme Court or illegal or legal or legislation'); insert into news_categories values (18,'Music News','piracy or anti-piracy or Recording Industry Association of America or copyright or copy-protection or CDs or music or artist or song'); commit;
Step 4 Create the CTXRULE index
In this step, we create a CTXRULE
index on our news_categories query column.
create index news_cat_idx on news_categories(query) indextype is ctxsys.ctxrule;
To classify the documents, we use the CLASSIFIER.THIS
PL/SQL procedure (a simple procedure designed for this example), which scrolls through the news_table
, matches each document to a category, and writes the categorized results into the news_id_cat
table.
create or replace package classifier asprocedure this;end;/ show errors create or replace package body classifier as procedure this is v_document clob; v_item number; v_doc number; begin for doc in (select tk, text from news_table) loop v_document := doc.text; v_item := 0; v_doc := doc.tk; for c in (select queryid, category from news_categories where matches(query, v_document) > 0 ) loop v_item := v_item + 1; insert into news_id_cat values (doc.tk,c.queryid); end loop; end loop; end this; end; / show errors exec classifier.this
The following considerations apply to indexing a CTXRULE
index.
If the SVM_CLASSIFIER
classifier is used, then you may use the BASIC_LEXER
, CHINESE_LEXER
, JAPANESE_LEXER
, or KOREAN_MORPH_LEXER
lexers. If SVM_CLASSIFIER
is not used, only the BASIC_LEXER
lexer type may be used for indexing your query set. (See the Oracle Text Reference for more on lexer and classifier preferences.)
Filter, memory, datastore, and [no]populate parameters are not applicable to index type CTXRULE
.
The CREATE INDEX
storage clause is supported for creating the index on the queries.
Wordlists are supported for stemming operations on your query set.
Queries for CTXRULE
are similar to those of CONTAINS
queries. Basic phrasing ("dog house") is supported, as are the following CONTAINS
operators: ABOUT
, AND
, NEAR
, NOT
, OR
, STEM
, WITHIN
, and THESAURUS
. Additionally, wildcards are supported. Section groups are supported for using the MATCHES
operator to classify documents. Field sections are also supported; however, CTXRULE
does not directly support field queries, so you must use a query rewrite on a CONTEXT
query.
With supervised classification, you employ the CTX_CLS.TRAIN
procedure to automate the rule writing step. CTX_CLS.TRAIN
uses a training set of sample documents to deduce classification rules. This is the major advantage over rule-based classification, in which you must write the classification rules.
However, before you can run the CTX_CLS.TRAIN
procedure, you must manually create categories and assign each document in the sample training set to a category. See the Oracle Text Reference for more information on CTX_CLS.TRAIN
.
When the rules are generated, you index them to create a CTXRULE
index. You can then use the MATCHES
operator to classify an incoming stream of new documents.
You may choose between two different classification algorithms for supervised classification:
Decision Tree classification. The advantage of Decision Tree classification is that the generated rules are easily observed (and modified).
SVM-based classification. This method uses the Support Vector Machine (SVM) algorithm for creating rules. The advantage of SVM-based classification is that it is often more accurate than Decision Tree classification. The disadvantage is that it generates binary rules, so the rules themselves are opaque.
To use Decision Tree classification, you set the preference argument to CTX_CLS.TRAIN
to RULE_CLASSIFIER
.
This form of classification uses a decision tree algorithm for creating rules. Generally speaking, a decision tree is a method of deciding between two (or more, but usually two) choices. In document classification, the choices are "the document matches the training set" or "the document does not match the training set."
A decision tree has a set of attributes that can be tested. In this case, these include:
words from the document
stems of words from the document (as an example, the stem of running is run)
themes from the document (if themes are supported for the language in use)
The learning algorithm in Oracle Text builds one or more decision trees for each category provided in the training set. These decision trees are then coded into queries suitable for use by a CTXRULE
index. As a trivial example, if one category is provided with a training document that consists of "Japanese beetle" and another category with a document reading "Japanese currency," the algorithm may create decision trees based on the words "Japanese," "beetle," and "currency," and classify documents accordingly.
The decision trees include the concept of confidence. Each rule that is generated is allocated a percentage value that represents the accuracy of the rule, given the current training set. In trivial examples, this accuracy is almost always 100%, but this merely represents the limitations of the training set. Similarly, the rules generated from a trivial training set may seem to be less than what you might expect, but these are sufficient to distinguish the different categories given the current training set.
The advantage of the Decision Tree method is that it can generate rules that are easily inspected and modified by a human. Using Decision Tree classification makes sense when you want to the computer to generate the bulk of the rules, but you want to fine tune them afterward by editing the rule sets.
The following SQL example steps through creating your document and classification tables, classifying the documents, and generating the rules. It then goes on to generate rules with CTX_CLS.TRAIN
.
Rules are then indexed to create CTXRULE
index and new documents are classified with MATCHES
.
The general steps for supervised classification can be broken down as follows:
The CTX_CLS.TRAIN
procedure requires an input training document set. A training set is a set of documents that have already been assigned a category.
Step 1 Create and populate a training document table
Create and load a table of training documents. This example uses a simple set; three concern fast food and three concern computers.
create table docs ( doc_id number primary key, doc_text clob); insert into docs values (1, 'MacTavishes is a fast-food chain specializing in burgers, fries and - shakes. Burgers are clearly their most important line.'); insert into docs values (2, 'Burger Prince are an up-market chain of burger shops, who sell burgers - and fries in competition with the likes of MacTavishes.'); insert into docs values (3, 'Shakes 2 Go are a new venture in the low-cost restaurant arena, specializing in semi-liquid frozen fruit-flavored vegetable oil products.'); insert into docs values (4, 'TCP/IP network engineers generally need to know about routers, firewalls, hosts, patch cables networking etc'); insert into docs values (5, 'Firewalls are used to protect a network from attack by remote hosts, generally across TCP/IP');
Step 2 Create category tables, category descriptions and ids
----------------------------------------------------------------------------
-- Create category tables -- Note that "category_descriptions" isn't really needed for this demo - -- it just provides a descriptive name for the category numbers in -- doc_categories ---------------------------------------------------------------------------- create table category_descriptions ( cd_category number, cd_description varchar2(80)); create table doc_categories ( dc_category number, dc_doc_id number, primary key (dc_category, dc_doc_id)) organization index; -- descriptons for categories insert into category_descriptions values (1, 'fast food'); insert into category_descriptions values (2, 'computer networking');
Step 3 Assign each document to a category
In this case, the fast food documents all go into category 1, and the computer documents into category 2.
insert into doc_categories values (1, 1); insert into doc_categories values (1, 2); insert into doc_categories values (1, 3); insert into doc_categories values (2, 4); insert into doc_categories values (2, 5);
Step 4 Create a CONTEXT index to be used by CTX_CLS.TRAIN
Create an Oracle Text preference for the index. This enables us to experiment with the effects of turning themes on and off:
exec ctx_ddl.create_preference('my_lex', 'basic_lexer'); exec ctx_ddl.set_attribute ('my_lex', 'index_themes', 'no'); exec ctx_ddl.set_attribute ('my_lex', 'index_text', 'yes'); create index docsindex on docs(doc_text) indextype is ctxsys.context parameters ('lexer my_lex');
Step 5 Create the rules table
Create the table that will be populated by the generated rules.
create table rules( rule_cat_id number, rule_text varchar2(4000), rule_confidence number );
Step 6 Call CTX_CLS.TRAIN procedure to generate category rules
Now call the CTX_CLS.TRAIN
procedure to generate some rules. Note all the arguments are the names of tables, columns or indexes previously created in this example. The rules
table now contains the rules, which you can view.
begin ctx_cls.train( index_name => 'docsindex', docid => 'doc_id', cattab => 'doc_categories', catdocid => 'dc_doc_id', catid => 'dc_category', restab => 'rules', rescatid => 'rule_cat_id', resquery => 'rule_text', resconfid => 'rule_confidence' ); end; /
Step 7 Fetch the generated rules, viewed by category
Fetch the generated rules. For convenience's sake, the rules
table is joined with category_descriptions
so we can see to which category each rule applies:
select cd_description, rule_confidence, rule_text from rules, category_descriptions where cd_category = rule_cat_id;
Once the rules are generated, you can test them by first indexing them and then using MATCHES
to classify new documents. The process is as follows:
Step 1 Index the rules to create the CTXRULE index
Use CREATE INDEX
to create the CTXRULE
index on the previously generated rules:
create index rules_idx on rules (rule_text) indextype is ctxsys.ctxrule;
Step 2 Test an incoming document using MATCHES
set serveroutput on;
declare incoming_doc clob; begin incoming_doc := 'I have spent my entire life managing restaurants selling burgers'; for c in ( select distinct cd_description from rules, category_descriptions where cd_category = rule_cat_id and matches (rule_text, incoming_doc) > 0) loop dbms_output.put_line('CATEGORY: '||c.cd_description); end loop; end; /
The second method we can use for training purposes is known as Support Vector Machine (SVM) classification. SVM is a type of machine learning algorithm derived from statistical learning theory. A property of SVM classification is the ability to learn from a very small sample set.
Using the SVM classifier is much the same as using the Decision Tree classifier, with the following differences.
The preference used in the call to CTX_CLS.TRAIN
should be of type SVM_CLASSIFIER
instead of RULE_CLASSIFIER
. (If you don't want to modify any attributes, you can use the predefined preference CTXSYS.SVM_CLASSIFIER
.)
The CONTEXT
index on the table does not have to be populated; that is, you can use the NOPOPULATE
keyword. The classifier uses it only to find the source of the text, by means of datastore and filter preferences, and to determine how to process the text, through lexer and sectioner preferences.
The table for the generated rules must have (as a minimum) these columns:
cat_id number, type number, rule blob );
As you can see, the generated rule is written into a BLOB
column. It is therefore opaque to the user, and unlike Decision Tree classification rules, it cannot be edited or modified. The trade-off here is that you often get considerably better accuracy with SVM than with Decision Tree classification.
With SVM classification, allocated memory has to be large enough to load the SVM model; otherwise, the application built on SVM will incur an out-of-memory error. Here is how to calculate the memory allocation:
Minimum memory request (in bytes) = number of unique categories x number of features example: (value of MAX_FEATURES attributes) x 8
If necessary to meet the minimum memory requirements, either:
increase SGA memory (if in shared server mode)
increase PGA memory (if in dedicated server mode)
The following example uses SVM-based classification. It uses essentially the same steps as the Decision Tree example. Some differences between the examples:
In this example, we set the SVM_CLASSIFIER
preference with CTX_DDL.CREATE_PREFERENCE
rather than setting it in CTX_CLS.TRAIN
. (You can do it either way.)
In this example, our category table includes category descriptions, unlike the category table in the Decision Tree example. (You can do it either way.)
CTX_CLS.TRAIN
takes fewer arguments than in the Decision Tree example, as rules are opaque to the user.
Step 1 Create and populate the training document table:
create table doc (id number primary key, text varchar2(2000)); insert into doc values(1,'1 2 3 4 5 6'); insert into doc values(2,'3 4 7 8 9 0'); insert into doc values(3,'a b c d e f'); insert into doc values(4,'g h i j k l m n o p q r'); insert into doc values(5,'g h i j k s t u v w x y z');
Step 2 Create and populate the category table:
create table testcategory ( doc_id number, cat_id number, cat_name varchar2(100) ); insert into testcategory values (1,1,'number'); insert into testcategory values (2,1,'number'); insert into testcategory values (3,2,'letter'); insert into testcategory values (4,2,'letter'); insert into testcategory values (5,2,'letter');
Step 3 Create the context index on the document table:
In this case, we create the index without population.
create index docx on doc(text) indextype is ctxsys.context parameters('nopopulate');
Step 4 Set SVM_CLASSIFIER:
This can also be done in CTX.CLS_TRAIN
.
exec ctx_ddl.create_preference('my_classifier','SVM_CLASSIFIER'); exec ctx_ddl.set_attribute('my_classifier','MAX_FEATURES','100');
Step 5 Create the result (rule) table:
create table restab ( cat_id number, type number(3) not null, rule blob );
Step 6 Perform the training:
exec ctx_cls.train('docx', 'id','testcategory','doc_id','cat_id', 'restab','my_classifier');
Step 7 Create a CTXRULE index on the rules table:
exec ctx_ddl.create_preference('my_filter','NULL_FILTER'); create index restabx on restab (rule) indextype is ctxsys.ctxrule parameters ('filter my_filter classifier my_classifier');
Now we can classify two unknown documents:
select cat_id, match_score(1) from restab where matches(rule, '4 5 6',1)>50; select cat_id, match_score(1) from restab where matches(rule, 'f h j',1)>50; drop table doc; drop table testcategory; drop table restab; exec ctx_ddl.drop_preference('my_classifier'); exec ctx_ddl.drop_preference('my_filter');
With Rule-Based Classification, you write the rules for classifying documents yourself. With Supervised Classification, Oracle Text writes the rules for you, but you must provide a set of training documents that you pre-classify. With unsupervised classification (also known as clustering), you don't even have to provide a training set of documents.
Clustering is performed with the CTX_CLS.CLUSTERING
procedure. CTX_CLS.CLUSTERING
creates a hierarchy of document groups, known as clusters, and, for each document, returns relevancy scores for all leaf clusters.
For example, suppose that you have a large collection of documents concerning animals. CTX_CLS.CLUSTERING
might create one leaf cluster about dogs, another about cats, another about fish, and a fourth cluster about bears. (The first three might be grouped under a node cluster concerning pets.) Suppose further that you have a document about one breed of dogs, such as chihuahuas. CTX_CLS.CLUSTERING
would assign the dog cluster to the document with a very high relevancy score, while the cat cluster would be assigned with a lower score and the fish and bear clusters with still lower scores. Once scores for all clusters have been assigned to all documents, an application can then take action based on the scores.
As noted in "Decision Tree Supervised Classification", attributes used for determining clusters may consist of simple words (or tokens), word stems, and themes (where supported).
CTX_CLS.CLUSTERING
assigns output to two tables (which may be in-memory tables):
A document assignment table showing how similar the document is to each leaf cluster. This information takes the form of document identification, cluster identification, and a similarity score between the document and a cluster.
A cluster description table containing information about what a generated cluster is about. This table contains cluster identification, cluster description text, a suggested cluster label, and a quality score for the cluster.
CTX_CLS.CLUSTERING
employs a K-MEAN
algorithm to perform clustering. Use the KMEAN_CLUSTERING
preference to determine how CTX_CLS.CLUSTERING
works.
The following SQL example creates a small collection of documents in the collection table and creates a CONTEXT
index. It then creates a document assignment and cluster description table, which are populated with a call to the CLUSTERING
procedure. The output would then be viewed with a select statement:
set serverout on /* collect document into a table */ create table collection (id number primary key, text varchar2(4000)); insert into collection values (1, 'Oracle Text can index any document or textual content.'); insert into collection values (2, 'Ultra Search uses a crawler to access documents.'); insert into collection values (3, 'XML is a tag-based markup language.'); insert into collection values (4, 'Oracle Database 10g XML DB treats XML as a native datatype in the database.'); insert into collection values (5, 'There are three Text index types to cover all text search needs.'); insert into collection values (6, 'Ultra Search also provides API for content management solutions.'); create index collectionx on collection(text) indextype is ctxsys.context parameters('nopopulate'); /* prepare result tables, if you omit this step, procedure will create table automatically */ create table restab ( docid NUMBER, clusterid NUMBER, score NUMBER); create table clusters ( clusterid NUMBER, descript varchar2(4000), label varchar2(200), sze number, quality_score number, parent number); /* set the preference */ exec ctx_ddl.drop_preference('my_cluster'); exec ctx_ddl.create_preference('my_cluster','KMEAN_CLUSTERING'); exec ctx_ddl.set_attribute('my_cluster','CLUSTER_NUM','3'); /* do the clustering */ exec ctx_output.start_log('my_log'); exec ctx_cls.clustering('collectionx','id','restab','clusters','my_cluster'); exec ctx_output.end_log;