You are viewing a plain text version of this content. The canonical link for it is here.
Posted to commits@uima.apache.org by pk...@apache.org on 2012/11/06 13:35:49 UTC
svn commit: r1406113 [2/2] - in
/uima/sandbox/trunk/TextMarker/uima-docbook-textmarker/src/docbook:
images/tools/tm/workbench/query/ images/tools/tm/workbench/testing/
workbench/
Modified: uima/sandbox/trunk/TextMarker/uima-docbook-textmarker/src/docbook/workbench/tools.textmarker.workbench.textruler.xml
URL: http://svn.apache.org/viewvc/uima/sandbox/trunk/TextMarker/uima-docbook-textmarker/src/docbook/workbench/tools.textmarker.workbench.textruler.xml?rev=1406113&r1=1406112&r2=1406113&view=diff
==============================================================================
--- uima/sandbox/trunk/TextMarker/uima-docbook-textmarker/src/docbook/workbench/tools.textmarker.workbench.textruler.xml (original)
+++ uima/sandbox/trunk/TextMarker/uima-docbook-textmarker/src/docbook/workbench/tools.textmarker.workbench.textruler.xml Tue Nov 6 12:35:47 2012
@@ -5,44 +5,54 @@
<!ENTITY % uimaents SYSTEM "../../target/docbook-shared/entities.ent" >
%uimaents;
]>
-<!-- Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements.
- See the NOTICE file distributed with this work for additional information regarding copyright ownership.
- The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not
- use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0
- Unless required by applicable law or agreed to in writing, software distributed under the License is
- distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- See the License for the specific language governing permissions and limitations under the License. -->
+<!--
+Licensed to the Apache Software Foundation (ASF) under one
+or more contributor license agreements. See the NOTICE file
+distributed with this work for additional information
+regarding copyright ownership. The ASF licenses this file
+to you under the Apache License, Version 2.0 (the
+"License"); you may not use this file except in compliance
+with the License. You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing,
+software distributed under the License is distributed on an
+"AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+KIND, either express or implied. See the License for the
+specific language governing permissions and limitations
+under the License.
+-->
<section id="section.ugr.tools.tm.workbench.textruler">
<title>TextRuler</title>
- <para> Using the knowledge engineering approach, a knowledge engineer normally writes
- handcrafted rules to create a domain dependent information extraction application, often
- supported by a gold standard. When starting the engineering process for the acquisition of the
- extraction knowledge for possibly new slot or more general for new concepts, machine learning
- methods are often able to offer support in an iterative engineering process. This section
- gives a conceptual overview of the process model for the semi-automatic development of
- rule-based information extraction applications.
+ <para> Using the knowledge engineering approach, a knowledge engineer normally writes handcrafted
+ rules to create a domain dependent information extraction application, often supported by a gold
+ standard. When starting the engineering process for the acquisition of the extraction knowledge
+ for possibly new slot or more general for new concepts, machine learning methods are often able
+ to offer support in an iterative engineering process. This section gives a conceptual overview
+ of the process model for the semi-automatic development of rule-based information extraction
+ applications.
</para>
<para> First, a suitable set of documents that contain the text fragments with interesting
patterns needs to be selected and annotated with the target concepts. Then, the knowledge
engineer chooses and configures the methods for automatic rule acquisition to the best of his
- knowledge for the learning task: Lambda expressions based on tokens and linguistic features,
- for example, differ in their application domain from wrappers that process generated HTML
- pages.
+ knowledge for the learning task: Lambda expressions based on tokens and linguistic features, for
+ example, differ in their application domain from wrappers that process generated HTML pages.
</para>
<para> Furthermore, parameters like the window size defining relevant features need to be set to
an appropriate level. Before the annotated training documents form the input of the learning
task, they are enriched with features generated by the partial rule set of the developed
- application. The result of the methods, that is the learned rules, are proposed to the
- knowledge engineer for the extraction of the target concept.
+ application. The result of the methods, that is the learned rules, are proposed to the knowledge
+ engineer for the extraction of the target concept.
</para>
<para> The knowledge engineer has different options to proceed: If the quality, amount or
- generality of the presented rules is not sufficient, then additional training documents need
- to be annotated or additional rules have to be handcrafted to provide more features in general
- or more appropriate features. Rules or rule sets of high quality can be modified, combined or
- generalized and transfered to the rule set of the application in order to support the
- extraction task of the target concept. In the case that the methods did not learn reasonable
- rules at all, the knowledge engineer proceeds with writing handcrafted rules.
+ generality of the presented rules is not sufficient, then additional training documents need to
+ be annotated or additional rules have to be handcrafted to provide more features in general or
+ more appropriate features. Rules or rule sets of high quality can be modified, combined or
+ generalized and transfered to the rule set of the application in order to support the extraction
+ task of the target concept. In the case that the methods did not learn reasonable rules at all,
+ the knowledge engineer proceeds with writing handcrafted rules.
</para>
<para> Having gathered enough extraction knowledge for the current concept, the semi-automatic
process is iterated and the focus is moved to the next concept until the development of the
@@ -50,103 +60,196 @@
</para>
<section id="ugr.tools.tm.textruler.learner">
<title>Available Learners</title>
- <para> Overview ||Name||Strategy||Document||Slots||Status |BWI (1) |Boosting, Top Down
- |Struct, Semi |Single, Boundary |Planning |LP2 (2) |Bottom Up Cover |All |Single, Boundary
- |Prototype |RAPIER (3) |Top Down/Bottom Up Compr. |Semi |Single |Experimental |WHISK (4)
- |Top Down Cover |All |Multi |Prototype |WIEN (5) |CSP |Struct |Multi, Rows |Prototype
- </para>
- <para> * Strategy: The used strategy of the learning methods are commonly coverage algorithms.
- * Document: The type of the document may be ''free'' like in newspapers, ''semi'' or
- ''struct'' like HTML pages. * Slots: The slots refer to a single annotation that represents
- the goal of the learning task. Some rule are able to create several annotation at once in
- the same context (multi-slot). However, only single slots are supported by the current
- implementations. * Status: The current status of the implementation in the TextRuler
- framework.
- </para>
- <para> Publications
- </para>
- <para> (1) Dayne Freitag and Nicholas Kushmerick. Boosted Wrapper Induction. In AAAI/IAAI,
- pages 577â583, 2000.
- </para>
- <para> (2) F. Ciravegna. (LP)2, Rule Induction for Information Extraction Using Linguistic
- Constraints. Technical Report CS-03-07, Department of Computer Science, University of
- Sheffield, Sheffield, 2003.
- </para>
- <para> (3) Mary Elaine Califf and Raymond J. Mooney. Bottom-up Relational Learning of Pattern
- Matching Rules for Information Extraction. Journal of Machine Learning Research, 4:177â210,
- 2003.
- </para>
- <para> (4) Stephen Soderland, Claire Cardie, and Raymond Mooney. Learning Information
- Extraction Rules for Semi-Structured and Free Text. In Machine Learning, volume 34, pages
- 233â272, 1999.
- </para>
- <para> (5) N. Kushmerick, D. Weld, and B. Doorenbos. Wrapper Induction for Information
- Extraction. In Proc. IJC Artificial Intelligence, 1997.
- </para>
- <para> BWI BWI (Boosted Wrapper Induction) uses boosting techniques to improve the performance
- of simple pattern matching single-slot boundary wrappers (boundary detectors). Two sets of
- detectors are learned: the "fore" and the "aft" detectors. Weighted by their confidences and
- combined with a slot length histogram derived from the training data they can classify a
- given pair of boundaries within a document. BWI can be used for structured, semi-structured
- and free text. The patterns are token-based with special wildcards for more general rules.
- </para>
- <para> Implementations No implementations are yet available.
- </para>
- <para> Parameters No parameters are yet available.
- </para>
- <para> LP2 This method operates on all three kinds of documents. It learns separate rules for
- the beginning and the end of a single slot. So called tagging rules insert boundary SGML
- tags and additionally induced correction rules shift misplaced tags to their correct
- positions in order to improve precision. The learning strategy is a bottom-up covering
- algorithm. It starts by creating a specific seed instance with a window of w tokens to the
- left and right of the target boundary and searches for the best generalization. Other
- linguistic NLP-features can be used in order to generalize over the flat word sequence.
- </para>
- <para> Implementations LP2 (naive): LP2 (optimized):
- </para>
- <para> Parameters Context Window Size (to the left and right): Best Rules List Size: Minimum
- Covered Positives per Rule: Maximum Error Threshold: Contextual Rules List Size:
- </para>
- <para> RAPIER RAPIER induces single slot extraction rules for semi-structured documents. The
- rules consist of three patterns: a pre-filler, a filler and a post-filler pattern. Each can
- hold several constraints on tokens and their according POS-tag- and semantic information.
- The algorithm uses a bottom-up compression strategy, starting with a most specific seed rule
- for each training instance. This initial rule base is compressed by randomly selecting rule
- pairs and search for the best generalization. Considering two rules, the least general
- generalization (LGG) of the slot fillers are created and specialized by adding rule items to
- the pre- and post-filler until the new rules operate well on the training set. The best of
- the k rules (k-beam search) is added to the rule base and all empirically subsumed rules are
- removed.
- </para>
- <para> Implementations RAPIER:
- </para>
- <para> Parameters Maximum Compression Fail Count: Internal Rules List Size: Rule Pairs for
- Generalizing: Maximum 'No improvement' Count: Maximum Noise Threshold: Minimum Covered
- Positives Per Rule: PosTag Root Type: Use All 3 GenSets at Specialization:
- </para>
- <para> WHISK WHISK is a multi-slot method that operates on all three kinds of documents and
- learns single- or multi-slot rules looking similar to regular expressions. The top-down
- covering algorithm begins with the most general rule and specializes it by adding single
- rule terms until the rule makes no errors on the training set. Domain specific classes or
- linguistic information obtained by a syntactic analyzer can be used as additional features.
- The exact definition of a rule term (e.g. a token) and of a problem instance (e.g. a whole
- document or a single sentence) depends on the operating domain and document type.
- </para>
- <para> Implementations WHISK (token): WHISK (generic):
- </para>
- <para> Parameters Window Size: Maximum Error Threshold: PosTag Root Type:
- </para>
- <para> WIEN WIEN is the only method listed here that operates on highly structured texts only.
- It induces so called wrappers that anchor the slots by their structured context around them.
- The HLRT (head left right tail) wrapper class for example can determine and extract several
- multi-slot-templates by first separating the important information block from unimportant
- head and tail portions and then extracting multiple data rows from table like data
- structures from the remaining document. Inducing a wrapper is done by solving a CSP for all
- possible pattern combinations from the training data.
- </para>
- <para> Implementations WIEN:
- </para>
- <para> Parameters No parameters are available.
- </para>
+ <para>
+ The available learners are based on the following publications:
+ <orderedlist numeration="arabic">
+ <!--
+ <listitem>
+ <para> Dayne Freitag and Nicholas Kushmerick. Boosted Wrapper Induction. In AAAI/IAAI,
+ pages 577-583, 2000.</para>
+ </listitem>
+ -->
+ <listitem>
+ <para> F. Ciravegna. (LP)2, Rule Induction for Information Extraction Using Linguistic
+ Constraints. Technical Report CS-03-07, Department of Computer Science, University of
+ Sheffield, Sheffield, 2003.</para>
+ </listitem>
+ <listitem>
+ <para> Mary Elaine Califf and Raymond J. Mooney. Bottom-up Relational Learning of Pattern
+ Matching Rules for Information Extraction. Journal of Machine Learning Research,
+ 4:177-210, 2003.</para>
+ </listitem>
+ <listitem>
+ <para> Stephen Soderland, Claire Cardie, and Raymond Mooney. Learning Information
+ Extraction Rules for Semi-Structured and Free Text. In Machine Learning, volume 34,
+ pages 233-272, 1999.</para>
+ </listitem>
+ <listitem>
+ <para> N. Kushmerick, D. Weld, and B. Doorenbos. Wrapper Induction for Information
+ Extraction. In Proc. IJC Artificial Intelligence, 1997.</para>
+ </listitem>
+ </orderedlist>
+ </para>
+ <para>
+ Each available learner has several features. Their meaning is explained here:
+ <itemizedlist>
+ <listitem>
+ <para> Strategy: The used strategy of the learning methods are commonly coverage
+ algorithms.</para>
+ </listitem>
+ <listitem>
+ <para>
+ Document: The type of the document may be
+ <quote>free</quote>
+ like in newspapers,
+ <quote>semi</quote>
+ or
+ <quote>struct</quote>
+ like in HTML pages.
+ </para>
+ </listitem>
+ <listitem>
+ <para> Slots: The slots refer to a single annotation that represents the goal of the
+ learning task. Some rule are able to create several annotation at once in the same
+ context (multi-slot). However, only single slots are supported by the current
+ implementations.</para>
+ </listitem>
+ <listitem>
+ <para> Status: The current status of the implementation in the TextRuler framework.</para>
+ </listitem>
+ </itemizedlist>
+ </para>
+ <para>
+ The following table gives an overview:
+ <table id="table.ugr.tools.tm.workbench.textruler.available_learners" frame="all">
+ <title>Overview of available learners</title>
+ <tgroup cols="6" colsep="1" rowsep="1">
+ <colspec colname="c1" colwidth="1*" />
+ <colspec colname="c2" colwidth="1*" />
+ <colspec colname="c3" colwidth="1*" />
+ <colspec colname="c4" colwidth="1*" />
+ <colspec colname="c5" colwidth="1*" />
+ <colspec colname="c6" colwidth="1*" />
+ <thead>
+ <row>
+ <entry align="center">Name</entry>
+ <entry align="center">Strategy</entry>
+ <entry align="center">Document</entry>
+ <entry align="center">Slots</entry>
+ <entry align="center">Status</entry>
+ <entry align="center">Publication</entry>
+ </row>
+ </thead>
+ <tbody>
+ <!--
+ <row>
+ <entry>BWI</entry>
+ <entry>Boosting, Top Down</entry>
+ <entry>Struct, Semi</entry>
+ <entry>Single, Boundary</entry>
+ <entry>Planning</entry>
+ <entry>1</entry>
+ </row>
+ -->
+ <row>
+ <entry>LP2</entry>
+ <entry>Bottom Up Cover</entry>
+ <entry>All</entry>
+ <entry>Single, Boundary</entry>
+ <entry>Prototype</entry>
+ <entry>2</entry>
+ </row>
+ <row>
+ <entry>RAPIER</entry>
+ <entry>Top Down/Bottom Up Compr.</entry>
+ <entry>Semi</entry>
+ <entry>Single</entry>
+ <entry>Experimental</entry>
+ <entry>3</entry>
+ </row>
+ <row>
+ <entry>WHISK</entry>
+ <entry>Top Down Cover</entry>
+ <entry>All</entry>
+ <entry>Multi</entry>
+ <entry>Prototype</entry>
+ <entry>4</entry>
+ </row>
+ <row>
+ <entry>WIEN</entry>
+ <entry>CSP</entry>
+ <entry>Struct</entry>
+ <entry>Multi, Rows</entry>
+ <entry>Prototype</entry>
+ <entry>5</entry>
+ </row>
+ </tbody>
+ </tgroup>
+ </table>
+ </para>
+ <!--
+ <section id="section.ugr.tools.tm.workbench.textruler.bwi">
+ <title>BWI (Boosted Wrapper Induction)</title>
+ <para> BWI uses boosting techniques to improve the performance of simple pattern matching
+ single-slot boundary wrappers (boundary detectors). Two sets of detectors are learned: the
+ "fore" and the "aft" detectors. Weighted by their confidences and combined with a slot
+ length histogram derived from the training data they can classify a given pair of boundaries
+ within a document. BWI can be used for structured, semi-structured and free text. The
+ patterns are token-based with special wildcards for more general rules. </para>
+ <para> Implementations No implementations are yet available. </para>
+ <para> Parameters No parameters are yet available. </para>
+ </section>
+ -->
+ <section id="section.ugr.tools.tm.workbench.textruler.lp2">
+ <title>LP2</title>
+ <para>LP2 This method operates on all three kinds of documents. It learns separate rules for
+ the beginning and the end of a single slot. So called tagging rules insert boundary SGML
+ tags and additionally induced correction rules shift misplaced tags to their correct
+ positions in order to improve precision. The learning strategy is a bottom-up covering
+ algorithm. It starts by creating a specific seed instance with a window of w tokens to the
+ left and right of the target boundary and searches for the best generalization. Other
+ linguistic NLP-features can be used in order to generalize over the flat word sequence.
+ </para>
+ <para> Parameters Context Window Size (to the left and right): Best Rules List Size: Minimum
+ Covered Positives per Rule: Maximum Error Threshold: Contextual Rules List Size: </para>
+ </section>
+ <section id="section.ugr.tools.tm.workbench.textruler.rapier">
+ <title>RAPIER</title>
+ <para>RAPIER induces single slot extraction rules for semi-structured documents. The rules
+ consist of three patterns: a pre-filler, a filler and a post-filler pattern. Each can hold
+ several constraints on tokens and their according POS-tag- and semantic information. The
+ algorithm uses a bottom-up compression strategy, starting with a most specific seed rule for
+ each training instance. This initial rule base is compressed by randomly selecting rule
+ pairs and search for the best generalization. Considering two rules, the least general
+ generalization (LGG) of the slot fillers are created and specialized by adding rule items to
+ the pre- and post-filler until the new rules operate well on the training set. The best of
+ the k rules (k-beam search) is added to the rule base and all empirically subsumed rules are
+ removed. </para>
+ <para> Parameters Maximum Compression Fail Count: Internal Rules List Size: Rule Pairs for
+ Generalizing: Maximum 'No improvement' Count: Maximum Noise Threshold: Minimum Covered
+ Positives Per Rule: PosTag Root Type: Use All 3 GenSets at Specialization: </para>
+ </section>
+ <section id="section.ugr.tools.tm.workbench.textruler.whisk">
+ <title>WHISK</title>
+ <para> WHISK is a multi-slot method that operates on all three kinds of documents and learns
+ single- or multi-slot rules looking similar to regular expressions. The top-down covering
+ algorithm begins with the most general rule and specializes it by adding single rule terms
+ until the rule makes no errors on the training set. Domain specific classes or linguistic
+ information obtained by a syntactic analyzer can be used as additional features. The exact
+ definition of a rule term (e.g. a token) and of a problem instance (e.g. a whole document or
+ a single sentence) depends on the operating domain and document type. </para>
+ <para> Parameters Window Size: Maximum Error Threshold: PosTag Root Type. </para>
+ </section>
+ <section id="section.ugr.tools.tm.workbench.textruler.wien">
+ <title>WIEN </title>
+ <para> WIEN is the only method listed here that operates on highly structured texts only. It
+ induces so called wrappers that anchor the slots by their structured context around them.
+ The HLRT (head left right tail) wrapper class for example can determine and extract several
+ multi-slot-templates by first separating the important information block from unimportant
+ head and tail portions and then extracting multiple data rows from table like data
+ structures from the remaining document. Inducing a wrapper is done by solving a CSP for all
+ possible pattern combinations from the training data. </para>
+ <para> Parameters No parameters are available. </para>
+ </section>
</section>
</section>
\ No newline at end of file