U.S. patent application number 15/244965 was filed with the patent office on 2017-05-11 for method and apparatus for extracting entity names and their relations.
The applicant listed for this patent is INTEL CORPORATION. Invention is credited to Yimin Zhang, Joe F. Zhou.
Application Number | 20170132529 15/244965 |
Document ID | / |
Family ID | 4574706 |
Filed Date | 2017-05-11 |
United States Patent
Application |
20170132529 |
Kind Code |
A1 |
Zhang; Yimin ; et
al. |
May 11, 2017 |
Method and Apparatus for Extracting Entity Names and Their
Relations
Abstract
According to one embodiment of the invention, a method includes
generating a person-name Information Gain (IG)-Tree and a relation
IG-Tree from annotated data. The method also includes tagging and
partial parsing of an input document. The names of the persons are
extracted within the input document using the person-name IG-tree.
Additionally, names of organizations are extracted within the input
document. The method also includes extracting entity names that are
not names of persons and organizations within the input document.
Further, the relations between the identified entity names are
extracted using the relation-IG-tree.
Inventors: |
Zhang; Yimin; (Beijing,
CN) ; Zhou; Joe F.; (Beijing, CN) |
|
Applicant: |
Name |
City |
State |
Country |
Type |
INTEL CORPORATION |
Santa Clara |
CA |
US |
|
|
Family ID: |
4574706 |
Appl. No.: |
15/244965 |
Filed: |
August 23, 2016 |
Related U.S. Patent Documents
|
|
|
|
|
|
Application
Number |
Filing Date |
Patent Number |
|
|
14293898 |
Jun 2, 2014 |
9430742 |
|
|
15244965 |
|
|
|
|
10019879 |
Oct 18, 2003 |
8745093 |
|
|
PCT/CN2000/000293 |
Sep 28, 2000 |
|
|
|
14293898 |
|
|
|
|
Current U.S.
Class: |
1/1 |
Current CPC
Class: |
G06F 16/313 20190101;
G06F 16/3344 20190101; G06N 20/00 20190101; G06F 16/94 20190101;
G06F 16/81 20190101 |
International
Class: |
G06N 99/00 20060101
G06N099/00; G06F 17/30 20060101 G06F017/30 |
Claims
1-30. (canceled)
31. A method comprising: generating a number of Information-Gain
(IG)-Trees based on a memory learning technique and the extracted
training sets; and extracting entity names and relations between
entity names based on the IG-Trees.
32. The method of claim 31, further comprising: receiving annotated
data; parsing, at least partially, the annotated data, wherein
parsing includes identifying syntactic structure of sentences
within the annotated data; and extracting training sets from the
parsed annotated data, wherein the training sets are based on
features including one or more of local context features, global
context features, surface linguistic features, and deep linguistic
features.
33. The method of claim 31, wherein the number of IG-Trees is
generated based on raw data that has been annotated.
34. The method of claim 33, wherein the number of IG-Trees is
generated based on a number of features of the annotated data.
35. The method of claim 31, wherein the number of IG-Trees is
selected from a group consisting of a person-name IG-Tree, an
entity-name IG-Tree, a noun phrase IG-Tree and a relation
IG-Tree.
36. A machine-readable medium comprising instructions which, when
executed by a machine, cause the machine to perform operations
comprising: generating a number of Information-Gain (IG)-Trees
based on a memory-learning technique and the extracted training
sets; and extracting entity names and relations between entity
names based on the IG-Trees.
37. The machine-readable medium of claim 36, wherein the operations
further comprise: receiving annotated data; parsing, at least
partially, the annotated data, wherein parsing includes identifying
syntactic structure of sentences within the annotated data; and
extracting training sets from the parsed annotated data, wherein
the training sets are based on features including one or more of
local context features, global context features, surface linguistic
features, and deep linguistic features.
38. The machine-readable medium of claim 36, wherein the number of
IG-Trees is generated based on raw data that has been
annotated.
39. The machine-readable medium of claim 37, wherein the number of
IG-Trees is generated based on a number of features of the
annotated data.
40. The machine-readable medium of claim 36, wherein the number of
IG-Trees is selected from a group consisting of a person-name
IG-Tree, an entity-name IG-Tree, a noun phrase IG-Tree and a
relation IG-Tree.
41. A system having a memory to store instructions, and a
processing device to execute the instructions, wherein the
instructions cause the processing device perform operations
comprising: generating a number of Information-Gain (IG)-Trees
based on a memory-learning technique and the extracted training
sets; and extracting entity names and relations between entity
names based on the IG-Trees.
42. The system of claim 41, wherein the operations further
comprise: receiving annotated data; parsing, at least partially,
the annotated data, wherein parsing includes identifying syntactic
structure of sentences within the annotated data; and extracting
training sets from the parsed annotated data, wherein the training
sets are based on features including one or more of local context
features, global context features, surface linguistic features, and
deep linguistic features.
43. The system of claim 41, wherein the number of IG-Trees is
generated based on raw data that has been annotated.
44. The system of claim 42, wherein the number of IG-Trees is
generated based on a number of features of the annotated data.
45. The system of claim 41, wherein the number of IG-Trees is
selected from a group consisting of a person-name IG-Tree, an
entity-name IG-Tree, a noun phrase IG-Tree and a relation IG-Tree.
Description
FIELD
[0001] The invention relates to the field of information
extraction. More specifically, the invention related to a method
and apparatus for extracting entity names and their relations.
BACKGROUND
[0002] Due to an increased knowledge base, the number of documents
across different subject matter areas continues to grow. For
example, with the advent of the Internet and the World Wide Web
(WWW), the documents on the different web sites on the Internet
continues to grow as the number of networks and servers connected
thereto continue to increase on a global scale. Accordingly, the
fields of information retrieval, document summarization,
information filtering and/or routing as well as topic tracking
and/or detection systems continue to grow in order to track and
service the vast amount of information.
[0003] In the field of information extraction, work has been done
to automatically learn patterns from a training corpus in order to
extract entity names and their relations from a given document. A
training corpus is defined to include writings, documents, or works
for a given subject matter. Moreover, an entity name is defined to
include, but is not limited to, proper names. Examples of entity
names include a person's name, a organization's name and a
product's name. Currently, tools for the extraction of entity names
include man-made rules and keyword sets to identify entity names.
Disadvantageously, building rules is often complex, error-prone and
time-consuming and usually requires a through understanding and
detailed knowledge of the system internals of a given language.
[0004] Another technique currently employed in the extraction of
entity names includes a statistical method. However, the training
of such a system requires vast amounts of human annotated data in
order to provide an accurate statistical analysis. Moreover, this
statistical method for the extraction of entity names is limited in
that only local context information can be employed during the
training of this method.
BRIEF DESCRIPTION OF THE DRAWINGS
[0005] Embodiments of the invention may be best understood by
referring to the following description and accompanying drawings
which illustrate such embodiments. In the drawings:
[0006] FIG. 1 illustrates a computer that provides for entity name
extractions and relations, according to embodiments of the present
invention;
[0007] FIG. 2 is a flowchart illustrating a method of training for
subsequent extraction of entity names and determination of
relations of such entity names, according to embodiments of the
present invention; and
[0008] FIG. 3 is a flowchart illustrating a method for the
extraction of entity names and relations there between based on the
training data generated by learning unit 110, according to
embodiments of the present invention.
DETAILED DESCRIPTION
[0009] In the following description, numerous specific details are
set forth to provide a thorough understanding of the invention.
However, it is understood that the invention may be practiced
without these specific details. In other instances, well-known
structures and techniques have not been shown in detail in order
not to obscure embodiments of the present invention.
[0010] FIG. 1 illustrates a computer that provides for entity name
extractions and relations, according to embodiments of the present
invention. As shown, computer 102 includes processing unit 104,
memory 102, annotation database 106, raw data database 108,
learning unit 110 and extracting unit 112, which are coupled
together. In one embodiment, learning unit 110 and extracting unit
112 are software programs that can reside in memory 102 and
processing unit 104 during their executing on processing unit
104.
[0011] In particular, memory 102 includes a machine-readable medium
on which is stored a set of instructions (i.e., software) embodying
any one, or all, of the methodologies described herein. Software
can reside, completely or at least partially, within memory 102
and/or within processing unit 104. For the purposes of this
specification, the term "machine-readable medium" shall be taken to
include any mechanism that provides (i.e., stores and/or transmits)
information in a form readable by a machine (e.g., a computer). For
example, a machine-readable medium includes read only memory (ROM);
random access memory (RAM); magnetic disk storage media; optical
storage media; flash memory devices; electrical, optical,
acoustical or other form of propagated signals (e.g., carrier
waves, infrared signals, digital signals, etc.); etc.
[0012] In conjunction with computer 102 of FIG. 1, a method of
operation of computer 102 in accordance with embodiments of the
present invention will now be described with reference to the flow
chart shown in FIG. 2. In particular, FIG. 2 is a flowchart
illustrating a method for the generation of training data employed
for subsequent extraction of entity names and determination of
relations of such entity names, according to embodiments of the
present invention. Learning unit 110 generates method 200 of FIG.
2. Moreover, as will be described below, method 200 employs a
memory-based learning technique in the generation of the
Information Gain (IG)-Trees, which are subsequently employed in the
extraction of entity names and relations there between for a given
document or set of documents.
[0013] FIG. 2 illustrates method 200 that commences with the
receipt of annotated training data by learning process 100, at
process block 202. In an embodiment, such annotated training data
is generated by a data analyst or an expert in a given subject
matter domain to which the annotated training data relates. In one
such embodiment, this person is presented with raw data that
includes different entity names and annotates such entity names as
well as the relations between such entity names. In one embodiment,
an entity name is defined to include proper names for persons,
locations, product and organizations. For example, an organization
entity name is INTEL.RTM.. However, embodiments of the present
invention are not so limited, as any other type of entity having a
proper name can be defined to have an entity name. For example, the
name of a pet could also be considered an entity name.
[0014] Accordingly, a data analyst or domain expert receives and
defines annotations for these entity names. For example, a given
sentence in the raw data for the organization entity name of
INTEL.RTM. could be annotated: "INTEL.RTM. is a leading designer
and manufacturer of computer processors." In an embodiment,
Extensible Markup Language (XML) tags are employed for the
insertion of the annotations into the raw data. These annotations
allow for the tracking of the different classes of entities as well
as assignments of unique identifications for the different entity
names within the raw data. Returning to the INTEL.RTM. example
described above, the following XML tag can be inserted into the raw
data: [0015] (1) "<ENT ID=1 Class=ORGANIZATION> INTEL.RTM.
</ENT> is a leading designer and manufacturer of computer
processors."
[0016] As shown, the "<ENT" and "</ENT>" denote the XML
tag, which includes a class category assignment and an
identification (ID) assignment. In particular, the class category
assignment and the ID assignment for INTEL.RTM. are organization
and one, respectively. Accordingly, subsequent processing by
learning unit 110 can incorporate or ignore the XML annotation tags
by using masking techniques to recognize or not recognize such
tags, as is known in the art. The above-described annotation
technique is by way of example and not by way of limitation, as
other techniques can be employed in the annotation of the data.
[0017] Additionally, a data analyst or domain expert also annotates
the relationship among and between the different entity names. For
example, for the entity names of John Smith and INTEL.RTM., the
relationship could be: "John Smith is a senior vice president of
INTEL.RTM.." Accordingly, the relationship between the person
entity name of "John Smith" and the organization entity name of
INTEL.RTM. could be employee/employer relationship.
[0018] In an embodiment, Extensible Markup Language (XML) tags are
employed for the insertion of the annotations regarding the
relationship among the entity names, which were provided by the
data analyst or domain expert. These annotations allow for the
tracking of the relationships among the different classes of
entities. Returning to the John Smith/INTEL.RTM. example described
above, the following XML tag can be inserted into the raw data:
[0019] (2) "John Smith is a senior vice president of INTEL.RTM.
<REL TYPE=Employee-OF ID=1 ID=2>."
[0020] As shown, the "<" and ">" denote the XML tag, which
includes the ID numbers for the entity names as well as the
relationship there between. In particular, the ID numbers for
INTEL.RTM. and John Smith are one and two respectively.
Additionally, the relationship type between the two entity names is
"Employee-OF." Similar to the annotations for the entity names,
subsequent processing by learning unit 110 can incorporate or
ignore the XML annotation tags for the relationships between the
entity names by using masking techniques to recognize or not
recognize such tags, as is known in the art. The above-described
annotation technique is by way of example and not by way of
limitation, as other techniques can be employed in the annotation
of the data. In an embodiment, this raw data that has been
annotated regarding the entity names and their relations is stored
in annotation database 106.
[0021] The above-described annotated data was described in terms of
a manual entry of the annotations and relations among the entity
names by a data analyst or a domain expert. However, embodiments of
the present invention are not so limited, as other sources can be
employed in generating of the annotated data. For example, such
data can be obtained via a database and/or automated computer
applications.
[0022] Returning to FIG. 2, in one embodiment, this annotated data
outputted from process block 202 is segmented, at process block
204. In particular, in certain languages, such as Chinese, there is
no space between words. Therefore, in such embodiments, the word
boundaries are defined, thereby allowing the words within the
annotated data to be identified and segmented (i.e., separated).
Accordingly, this segmentation provides a space between each of the
words in the annotated data. In an embodiment, a software
application can receive the annotated data and can automatically
output segmented annotated data, as there are different software
applications available to provide this functionality for different
languages. However, embodiments of the present invention are not so
limited, as other processes can be employed for the segmentation of
the annotated data. For example, in an embodiment, a data analyst
or a domain expert can analyze and segment the words within the
annotated data.
[0023] In an embodiment, learning unit 110 also tags this annotated
data, at process block 206. In particular, tagging is defined to
include the marking of the words in the annotated data as their
given type. For example, the words in the annotated data can be
tagged as verbs, nouns, proper nouns, pronouns, adverbs, adjective,
etc. In an embodiment, software applications, as are known in the
art, are employed in the tagging of this data.
[0024] Moreover, learning unit 110 partially parses this annotated
data, at process block 208. In particular, partial parsing
identifies the syntactic structure of the sentences within the
annotated data. For example, learning unit 110 identifies subjects,
main verbs and objects of the sentences. However, embodiments of
the present invention are not so limited, as other syntactic
structures of a sentence can also be identified during this partial
parsing process. For example, the prepositional phrases,
subordinate clauses or other structures can be identified in the
partial parsing process.
[0025] Learning unit 110 then generates a number of training sets
from the parsed annotated data, at process block 210. In one such
embodiment, four training sets are extracted from the parsed
annotated data. A first training set includes the features of the
first and last word in a person's name in the training data.
Learning unit 110 also generates a second training set, which
includes the features for the entity names in the training data.
Moreover, learning unit 110 generates a third training set, which
includes the features for the noun phrases in the training data.
Noun phrases are defined to include those phrases in the sentences
that include one to a number of noun phrases therein. Learning unit
110 also generates a fourth training set, which includes the
features of the relations between entity names in the training
data. The above-described training sets are by way of example and
not by way of limitation, as other types of training sets can be
incorporated into embodiments of the present invention.
[0026] In one embodiment, a feature of a given word or set of words
is defined to include (1) local context features, (2) global
context features, (3) surface linguistic features and (4) deep
linguistic features, as are known in the art. For example, local
context features of a word or set of words include the features
extracted from the local context of the word or set of words. An
example of a local context feature could include the n-th word
before and/or after the word or word sequence. Another example of a
local context feature could include the nearest verb before and/or
after the word or word sequence. Such features are described in
more detail below in conjunction with the Information Gain
(IG)-Trees, which are generated from these training sets.
[0027] At process block 212, learning unit 110 generates one to a
number of Information Gain (IG)-Trees employing a memory-based
learning technique, as is known in the art. In particular,
memory-based learning is a type of learning that is supervised and
inductive and is based on examples. Memory-based learning entails a
classification based supervised learning approach. Embodiments of
memory-based learning include similarity-based, example-based,
analogical, case-based, instance-based and lazy learning depending
on the context. In one embodiment, learning unit 110 generates (1)
a person name IG-Tree, (2) a entity name IG-Tree, (3) a noun phrase
IG-Tree and (4) a relation-IG-Tree based on the person name
training set, the entity name training set, the noun phrase
training set and the relation training set, respectively. The
above-described IG-Trees are by way of example and not by way of
limitation, as other types of IG-Trees can be incorporated into
embodiments of the present invention.
[0028] With regard to the person-name IG-Tree and the noun-phrase
IG-Tree, examples of features included therein will now be
described. In an embodiment, context features are defined in terms
of the character types of words in proximity to the word or word
sequence. Examples of a character type for a word includes, but is
not limited to, a surname, a number, types of ASCII characters and
normal words. A context feature for a given word could, therefore,
include the character type of the n-th previous word or the
character type of the n-th word subsequent to the word or word
sequence. Further examples of context features for a given word or
word sequence could include the number of characters in the n-th
previous and/or next word. Moreover, the context features could
include the actual n-th previous or next word in relationship to
the given word or word sequence. In an embodiment, another set of
relevant context features includes the tags for the previous and/or
next n-th word.
[0029] In one embodiment, an additional context feature is the
dependency relation for the context words. Examples of such
dependency relations includes, but are not limited to, an
adjective/noun relation wherein the adjective is modifying the
noun, (2) verb/noun wherein the verb is the predicate for a noun
acting as an object, (3) noun/verb wherein the noun is the subject
and the verb is the predicate, (4) preposition/object, (5)
adverb/head wherein the head can be a verb or adjective, (6)
noun/noun, wherein a noun is modifying a noun and (7) verb/verb,
wherein the second verb is the object of the first verb.
[0030] In one embodiment, a dependency relation includes the first
verb appearing before and/or after the given word or word sequence
in the same sentence. In an embodiment, a dependency relation
includes the syntactical relationship between a verb and the given
word or word sequence. For example, for a given sentence, there
could be an object/verb relationship or a subject/verb
relationship. Moreover, in one embodiment, another dependency
relations are the semantic features for a given syntactical
relationship between a verb and the word or word sequence. In an
embodiment, another context feature for a given word or word
sequence includes bigram and mutual information features, as are
known in the art.
[0031] The class labels associated with each word for a person-name
IG-Tree include (1) the left boundary of a person name, (2) the
right boundary of a person name and (3) whether or not the word is
in a person name. The class labels associated with each word for a
noun-phrase IG-Tree include (1) the left boundary of the noun
phrase, (2) the right boundary of the noun phrase and (3) whether
or not the word is in a noun phrase. Accordingly, in the extracting
process that is described below, these features help in determining
a word's or word sequence's class, thereby allowing the extracting
process to determine the boundary of a noun phrase/person name in a
given document or set of documents.
[0032] With regard to the entity-name IG-Tree, examples of features
included therein will now be described. In an embodiment, a local
context feature included in the entity-name IG-Tree could be the
relationship of the word or word sequence with a verb in the same
sentence. For example, one such feature could be the first verb
appearing after the entity name in the same sentence. Another
example could be the first verb appearing before the entity name in
the same sentence.
[0033] In one embodiment, another local context feature included in
the entity-name IG-Tree could be the syntactical relationship of
the word or word sequence with a given verb. Such syntactical
relationships could be object/verb or a subject/verb. For example,
if a given sentence is "I eat rice" and the current word is "rice",
the object/verb of such a word would be "eat" while the
subject/verb would be "I." Another example of a syntactical
relationship could be the preposition word before this given entity
name. Moreover, another local context feature could include the
semantic features. An example of a semantic feature could be the
semantic category of the headword of a given entity name. Another
example of a semantic feature could be the semantic category of the
verb for a given object/verb or subject/verb syntactical
relationship.
[0034] Global context features provide a broader view of a word or
word sequence with regard to the entire document. For example, a
global context feature included in the entity-name IG-Tree could be
a set of first verbs in a same sentence that appear after a word or
word sequence for the entire document or corpus. In a similar
example, another global context feature included in the entity-name
IG-Tree could be a set of first verbs in a same sentence that
appear before a word or word sequence for the entire document or
corpus.
[0035] The class labels associated with each entity name for an
entity-name IG-Tree include their types. Examples of an entity name
type include, but are not limited to, person, organization and
product. Accordingly, in the extracting process that is described
below, these features help in determining an entity name's class
label, thereby allowing the extracting process to determine the
boundary of a entity name in a given document or set of
documents.
[0036] With regard to the relation IG-Tree, examples of features
included therein will now be described. In an embodiment, one
feature includes the syntactic position for this entity name. For
example, a syntactic position could be the head of the subject or
the modifier of the subject for a given sentence. Another feature
that can be included is a set of words or entity names in the
subject, the modifier of the subject or the head of the subject.
Moreover, another feature could be the semantic category of the
head of the subject or the set of semantic categories of terms in
the modifier of the subject.
[0037] In one embodiment, features that can be included in the
relation IG-Tree can include a set of words or entity names in the
object, in the modifier of the object or in the head of the object.
Other features that can be included are the semantic categories of
the head of the object as well as the set of semantic categories of
terms in the modifier of the object. Additional features can
include the main verb of the sentence and the semantic category of
this verb.
[0038] Another set of features to be included in the relation
IG-Tree relates to the prepositions in the different sentences. For
example, one set of such features could be the first preposition in
a given sentence as well as a set of the words or entity names in
the object of the first preposition in this sentence. This example
is applicable for any preposition phrase within a sentence. For
example, another set of features could be the second preposition in
a given sentence as well as a set of the words or entity names in
the object of the second preposition phrase in this sentence.
[0039] The class labels for the relation IG-Tree include the
relationship name as well as the entity names that define the
relationship. For example, for an employee/company relationship,
the relationship could be "employee of" while the entity names
could include the employee and the organization. Accordingly, in
the extracting process that is described below, these features help
in determining a class label for a given relationship between two
entity names.
[0040] The above-described features included in the different
IG-Trees are by way of example and not by way of limitation. For
example, embodiments of the present invention can incorporate other
types of features into the different IG-Trees. Moreover,
embodiments of the present invention can incorporate the features
described for a given IG-Tree into other IG-Trees.
[0041] Accordingly, learning unit 110 illustrated in FIG. 2
generates a number of IG-Trees based on the annotated data
received. In conjunction with computer 102 of FIG. 1, a method of
operation of computer 102 in accordance with embodiments of the
present invention will now be described with reference to the flow
chart shown in FIG. 3. In particular, FIG. 3 is a flowchart
illustrating a method for the extraction of entity names and
relations there between based on the IG-Trees generated by learning
unit 110, according to embodiments of the present invention.
Extracting unit 112 generates method 300 of FIG. 3.
[0042] Method 300 commences with the receipt of an input document
by extracting unit 112, at process block 302. In an embodiment,
extracting unit 112 segments the input document, at process block
304. As described above in conjunction with the segmenting of the
annotated data, in certain languages, such as Chinese, there is no
space between words. Therefore, in such embodiments, the word
boundaries are defined, thereby allowing the words within the input
document to be identified and segmented (i.e., separated).
Accordingly, this segmentation provides a space between each of the
words in the input document. In an embodiment, a software
application can receive the input document and can automatically
output a document that has been segmented, as there are different
software applications available to provide this functionality for
different languages. However, embodiments of the present invention
are not so limited, as other processes can be employed for the
segmentation of the input document. For example, in an embodiment,
a data analyst or a domain expert can analyze and segment the words
within the input document.
[0043] In one embodiment, extracting unit 112 tags the input
document, at process block 306. In particular, tagging is defined
to include the marking of the words in the input document as their
given type. For example, the words in the input document can be
tagged as verbs, nouns, proper nouns, pronouns, adverbs, adjective,
etc. In an embodiment, software applications, as are known in the
art, are employed in the tagging, of this data.
[0044] Additionally, in an embodiment, extracting unit 112
partially parses the input document. Partial parsing identifies the
syntactic structure of the sentences within the input document. For
example, extracting unit 112 identifies subjects, main verbs and
objects of the sentences. However, embodiments of the present
invention are not so limited, as other syntactic structures of a
sentence can also be identified during this partial parsing
process. For example, the prepositional phrases, subordinate
clauses or other structures can be identified in the partial
parsing process.
[0045] Extracting unit 112 also identifies the people names within
the input document using the person-name IG-Tree generated by
learning unit 110, at process block 310. Such technique of
extracting or identifying given words from a document using
IG-Trees is known in the art. In one such embodiment, the input
document has been segmented, tagged and/or partially parsed prior
to the identification of the people names.
[0046] Moreover, extracting unit 112 identifies the organizational
names within the input document using simple pattern matching
rules, as are known in the art. For example, if a given noun phrase
includes the terms "corp." or "ltd.", such pattern matching rules
assume that this noun phrase is an organization name. In one such
embodiment, the input document has been segmented, tagged and/or
partially parsed prior to the identification of the people
names.
[0047] Additionally extracting unit 112 identifies the remaining
entity names, such as product names, that have not already been
identified using simple pattern matching rules, as are known in the
art, at process block 314. In one such embodiment, the input
document has been segmented, tagged and/or partially parsed prior
to the identification of the people names. Moreover, extracting
unit 112 also identifies the noun phrases in the input document
using the noun-phrase IG-Tree that was generated by learning unit
110, at process block 316.
[0048] Further, in an embodiment, extracting unit 112 performs an
additional partial parsing procedure using the entity names and
noun phrases that were previously extracted, at process block 318.
In one embodiment, extracting unit 112 classifies the noun phrases
that were extracted into different categories, at process block
320. Examples of the different categories include person, location,
organization, etc. Accordingly, this classification of noun phrases
can be compared to the previously identified entity names to allow
for the identification of those entity names that might be missed
in prior identifications.
[0049] Additionally, extracting unit 112 identifies relations
between and among the identified (extracted) entity names using the
relation IG-tree generated by learning unit 110. Accordingly,
extracting unit 112 can extract the entity names and the relations
there between using IG-Trees employing on a memory-based learning
technique.
[0050] Although the present invention has been described with
reference to specific exemplary embodiments, it will be evident
that various modifications and changes may be made to these
embodiments without departing from the broader spirit and scope of
the invention. Accordingly, the specification and drawings are to
be regarded in an illustrative rather than a restrictive sense.
* * * * *