U.S. patent application number 13/605231 was filed with the patent office on 2012-12-27 for providing question and answers with deferred type evaluation using text with limited structure.
This patent application is currently assigned to International Business Machines Corporation. Invention is credited to Pablo A. Duboue, James J. Fan, David A. Ferrucci, James W. Murdock, IV, Christopher A. Welty, Wlodek W. Zadrozny.
Application Number | 20120330934 13/605231 |
Document ID | / |
Family ID | 45871692 |
Filed Date | 2012-12-27 |
United States Patent
Application |
20120330934 |
Kind Code |
A1 |
Duboue; Pablo A. ; et
al. |
December 27, 2012 |
PROVIDING QUESTION AND ANSWERS WITH DEFERRED TYPE EVALUATION USING
TEXT WITH LIMITED STRUCTURE
Abstract
A system, method and computer program product for conducting
questions and answers with deferred type evaluation based on any
corpus of data. The method includes processing a query including
waiting until a "Type" (i.e. a descriptor) is determined AND a
candidate answer is provided. Then, a search is conducted to look
(search) for evidence that the candidate answer has the required
Lexical Answer Type (e.g., as determined by a matching function
that can leverage a parser, a semantic interpreter and/or a simple
pattern matcher). Prior to or during candidate answer evaluation, a
process is provided for extracting and storing collections of
entity-type pairs from semi-structured text documents. During QA
processing and candidate answer scoring, a process is implemented
to match the query LAT against the lexical type of each provided
candidate answer and generate a score judging a degree of
match.
Inventors: |
Duboue; Pablo A.; (Montreal,
CA) ; Fan; James J.; (Parsippany, NJ) ;
Ferrucci; David A.; (Yorktown Heights, NY) ; Murdock,
IV; James W.; (Millwood, NY) ; Welty; Christopher
A.; (Hawthorne, NY) ; Zadrozny; Wlodek W.;
(Tarrytown, NY) |
Assignee: |
International Business Machines
Corporation
Armonk
NY
|
Family ID: |
45871692 |
Appl. No.: |
13/605231 |
Filed: |
September 6, 2012 |
Related U.S. Patent Documents
|
|
|
|
|
|
Application
Number |
Filing Date |
Patent Number |
|
|
13239165 |
Sep 21, 2011 |
|
|
|
13605231 |
|
|
|
|
61386017 |
Sep 24, 2010 |
|
|
|
Current U.S.
Class: |
707/722 ;
707/736; 707/E17.014 |
Current CPC
Class: |
G06F 16/334
20190101 |
Class at
Publication: |
707/722 ;
707/736; 707/E17.014 |
International
Class: |
G06F 17/30 20060101
G06F017/30 |
Claims
1. A system for generating answers to questions comprising: a
memory storage device; one or more processing devices, each
operatively connected to said memory device and configured for
performing a method comprising: determining a lexical answer type
(LAT) associated with an input query received at an automated
questions and answer (QA) system; obtaining one or more candidate
answers to the input query using a data source having
semi-structured content; determining a lexical type (LT) for each
said one or more obtained candidate answers from said
semi-structured content; comparing, at said one or more processing
devices, said query LAT with said candidate answer LT; and
generating a score representing a degree of match between said
query LAT with said candidate answer LT, said score indicative of a
quality of said obtained candidate answer.
2. The system of claim 1, wherein said processor device is further
configured to: identify, in said semi-structured content, one or
more entities and associated lexical type information; and store,
in a data storage device in communication with said QA system,
entity-type structures, each entity-type structure representing
said one or more entities and associated lexical type information,
wherein said determining a lexical type (LT) includes accessing
said stored entity-type structures to identify a lexical type (LT)
from a type associated with said one or more entities stored in
said entity-type data structures.
3. The system of claim 2, wherein said identifying and storing are
performed off-line as a pre-processing step.
4. The system of claim 2, wherein said identifying and storing are
performed on-demand by a QA system at a time the input query is
received.
5. The system of claim 2, wherein said processor device is
configured to parse each respective said query LAT and said
candidate answer LT to obtain respective terms or phrases for
each.
6. The system of claim 1, wherein said comparing comprises matching
individual terms of respective query LAT and candidate answer LT,
or matching entire phrases of each respective query LAT and
candidate answer LT.
7. The system of claim 5, wherein processing device is further
configured to: determine which terms of respective said query LAT
and candidate answer LT are to be used for said term matching;
obtain individual scores based on degree of match between terms of
respective said LAT and candidate answer LT; and combine obtained
scores of said matches determined for said phrase matching.
8. A system for providing content to a database used by an
automatic question and answering (QA) system comprising: a memory
storage device; one or more processing devices, each operatively
connected to said memory device and configured for performing a
method to: automatically identify semi-structured text data from a
data source; automatically identify one or more entity-type
relations from said semi-structured text data, said entity-type
relation including one or more entities associated with a type;
automatically extract said identified entity-type relations; and
store said extracted entity-type relations as entity-type data
structures in said database, wherein a processing device is
configured to perform said automatic identifying of semi-structured
text and entity-type relations, said extracting and said
storing.
9. The system as claimed in claim 1, wherein said semi-structured
text comprises item-delimited markup, said automatically
identifying of semi-structured text data comprising: parsing
content of said data source to identify said item-delimiting
markup, said item delimited mark-up specifying said type
information and entities forming an entity-type data structure.
10. The system as claimed in claim 9, wherein said item-delimiting
markup includes a title, a header, a recitation of the word "list"
of entities of a specified type, bullet markers, parentheses, a
hypertext link, a Uniform Resource Locator, a table in said data
source.
Description
CROSS-REFERENCE TO RELATED APPLICATIONS
[0001] The present invention relates to and claims the benefit of
the filing date of commonly-owned, co-pending U.S. patent
application Ser. No. 13/239,165 filed Sep. 21, 2011 which claims
the benefit of United States relates to and claims the benefit of
the filing date of commonly-owned, co-pending U.S. Provisional
Patent Application No. 61/386,017, filed Sep. 24, 2010, the entire
contents and disclosure of which is incorporated by reference as if
fully set forth herein.
BACKGROUND
[0002] 1. Field of the Invention
[0003] The invention relates generally to information retrieval
systems, and more particularly, the invention relates to a novel
query/answer system and method for open domains implementing a
deferred type evaluation of candidate answers using text with
limited structure.
[0004] 2. Description of the Related Art
[0005] An introduction to the current issues and approaches of
question answering (QA) can be found in the web-based reference
http://en.wikipedia.org/wiki/Question_answering. Generally, QA is a
type of information retrieval. Given a collection of documents
(such as the World Wide Web or a local collection) the system
should be able to retrieve answers to questions posed in natural
language. QA is regarded as requiring more complex natural language
processing (NLP) techniques than other types of information
retrieval such as document retrieval, and it is sometimes regarded
as the next step beyond search engines.
[0006] QA research attempts to deal with a wide range of question
types including: fact, list, definition, How, Why, hypothetical,
semantically-constrained, and cross-lingual questions. Search
collections vary from small local document collections, to internal
organization documents, to compiled newswire reports, to the World
Wide Web.
[0007] Closed-domain QA deals with questions under a specific
domain, for example medicine or automotive maintenance, and can be
seen as an easier task because NLP systems can exploit
domain-specific knowledge frequently formalized in ontologies.
Open-domain QA deals with questions about nearly everything, and
can only rely on general ontologies and world knowledge. On the
other hand, these systems usually have much more data available
from which to extract the answer.
[0008] Alternatively, closed-domain QA might refer to a situation
where only a limited type of questions are accepted, such as
questions asking for descriptive rather than procedural
information.
[0009] Access to information is currently dominated by two
paradigms. First, a database query that answers questions about
what is in a collection of structured records. Second, a search
that delivers a collection of document links in response to a query
against a collection of unstructured data, for example, text or
html.
[0010] A major unsolved problem in such information query paradigms
is the lack of a computer program capable of accurately answering
factual questions based on information included in a collection of
documents that can be either structured, unstructured, or both.
Such factual questions can be either broad, such as "what are the
risks of vitamin K deficiency?", or narrow, such as "when and where
was Hillary Clinton's father born?"
[0011] It is a challenge to understand the query, to find
appropriate documents that might contain the answer, and to extract
the correct answer to be delivered to the user. There is a need to
further advance the methodologies for answering open-domain
questions.
SUMMARY
[0012] The present invention addresses the need described above by
providing a dynamic infrastructure and methodology for conducting
question answering with deferred type evaluation using text with
limited structure.
[0013] An aspect of the invention provides a system implementing
machine processing for answering questions employing a processing
step in which semi-structured information, for example information
with limited structure, is extracted from the knowledge and
database sources and re-represented in a form suitable for machine
processing.
[0014] Particularly, a system and method is provided for extracting
answer-typing information from sources with limited structure and
using that extracted type information for scoring candidate
answers.
[0015] Thus, in one aspect, there is provided a system and method
for providing content to a database used by an automatic QA system.
The method includes automatically identifying semi-structured text
data from a data source; automatically identifying one or more
entity-type relations from said semi-structured text data, said
entity-type relation including one or more entities associated with
a type; automatically extracting said identified entity-type
relations; and, storing said extracted entity-type relations as
entity-type data structures in said database, wherein a processing
device is configured to perform said automatic identifying of
semi-structured text and entity-type relations, said extracting and
said storing.
[0016] Further to this aspect, the semi-structured text comprises
item-delimited markup, said automatically identifying of
semi-structured text data comprising parsing content of said data
source to identify said item-delimiting markup, said item delimited
mark-up specifying said type information and entities forming an
entity-type data structure.
[0017] Further, the item-delimiting markup includes a title, a
header, a recitation of the word "list" of entities of a specified
type, bullet markers, parentheses, a hypertext link, a Uniform
Resource Locator, or a table in said data source.
[0018] In a further aspect, there is provided computer-implemented
system and method for automatically generating answers to questions
comprising the steps of: determining a lexical answer type (LAT)
associated with an input query; obtaining one or more candidate
answers to the input query using a data source having
semi-structured content; determining a lexical type (LT) for each
the one or more obtained candidate answer from the semi-structured
content; comparing the query LAT with the candidate answer LT; and
generating a score representing a degree of match between the
compared query LAT with the candidate answer LT, the score
indicative of a quality of the obtained candidate answer, wherein a
processing device automatically performs one or more of the
determining a query LAT, computing candidate answers, determining a
LT, comparing and generating.
[0019] In this further aspect, the computer-implemented method
further comprises: identifying, in the semi-structured content, one
or more entities and associated lexical type information; and,
storing, in a data storage device in communication with a QA
system, entity-type structures, each entity-type structure
representing the one or more entities and associated lexical type
information, wherein said determining a lexical type includes
accessing said stored entity-type structures to identify a lexical
type (LT) from a type associated with said one or more entities
stored in said entity-type data structures.
[0020] In this further aspect, the comparing comprises parsing each
respective the query LAT and the candidate answer LT to obtain
respective terms or phrases for each; the comparing further
comprising one or more of: matching individual terms of respective
query LAT and candidate answer LT, or matching entire phrases of
each respective query LAT and candidate answer LT.
[0021] A computer program product is provided for performing
operations. The computer program product includes a storage medium
readable by a processing circuit and storing instructions run by
the processing circuit for running a method(s). The method(s) are
the same as listed above.
BRIEF DESCRIPTION OF THE DRAWINGS
[0022] The objects, features and advantages of the invention are
understood within the context of the Detailed Description, as set
forth below. The Detailed Description is understood within the
context of the accompanying drawings, which form a material part of
this disclosure, wherein:
[0023] FIG. 1 shows a high level logical architecture 10 of a
question/answering method in which the present invention may be
employed;
[0024] FIG. 2 is a flow diagram of a method for conducting
questions and answers with deferred type evaluation in a QA
system;
[0025] FIG. 3 is a flow diagram illustrating score production
performed at step 132 of the flow diagram of FIG. 2;
[0026] FIG. 4 depicts the method for automatically identifying
pages, documents and other content including text of limited
structure ("semi-structured") that can be implemented at step 132b
of FIG. 3;
[0027] FIG. 5 illustrates a LATTE process 300 applying phase and
term matching functions that can be implemented at step 132c of
FIG. 3;
[0028] FIG. 6 depicts a block diagram of the phrase and term
processing functions 345 of FIG. 5
[0029] FIG. 7 illustrates an example flow of control 500 for the
LATTE process and score production among the components depicted in
FIG. 6;
[0030] FIG. 8 illustrates an example flow control for the LATTE
process of FIGS. 5-6
[0031] FIG. 9 depicts an aspect of a UIMA framework implementation
for providing one type of analysis engine for processing CAS data
structures; and
[0032] FIG. 10 illustrates an exemplary hardware configuration to
run method steps described in FIGS. 2-8 in one embodiment.
DETAILED DESCRIPTION
[0033] Commonly-owned, co-pending U.S. patent application Ser. No.
12/126,642, titled "SYSTEM AND METHOD FOR PROVIDING QUESTION AND
ANSWERS WITH DEFERRED TYPE EVALUATION", incorporated by reference
herein, describes a QA system and method in which answers are
automatically generated for questions. More particularly, it
involves comparing the lexical types determined from the question
to the lexical types associated with each candidate answer that is
derived from a search. The lexical answer type requested by the
question is referred to herein as a "lexical answer type" or "LAT."
The lexical types that are associated with each candidate answer is
referred to herein as a lexical type or "LT".
[0034] The QA system utilizes the system and methodology described
below with respect to FIG. 1. Evidence gathering and answer scoring
functions 50 are programmed processing components that receive (1)
a "lexical answer type" (LAT), e.g., a text string describing some
concept and (2) "candidate answer" input(s), e.g., a text string
describing some entity. The programmed processing components
generate an output including a judgment (e.g., a score)
representing to what degree the entity is an instance of the
concept. As described in FIGS. 2 and 3 below, this is used as part
of answer scoring functions, for example, based on computed
candidate answer lexical types (LTs).
[0035] As described below with respect to FIG. 3, after processing
an input query to determine a LAT and searching and obtaining one
or more candidate answers, there is performed for each candidate
answer received the steps of matching the candidate against
instances in a database which results in generating an output data
structure, including the matched instances 132a; looking
(searching) for evidence that the candidate answer has the required
LAT and retrieving LT(s) associated with those instances in the
knowledge base (KB) 132b; and attempting to match LAT(s) with LT(s)
(e.g., as determined by a matching function that using a parser, a
semantic interpreter and/or a simple pattern matcher) and producing
a score representing the degree of match 132c. More particularly,
the candidate answer LT and query LAT(s) are represented as lexical
strings. Production of the score, referred to as a "TyCor" (Type
Coercion) score, is comprised of the three steps: (1) candidate
answer to instance matching, (2) instance to type association
extraction, and (3) LAT to type matching. The score reflects the
degree to which the candidate may be "coerced" to the LAT, where
higher scores indicate a better coercion.
[0036] In one embodiment, as will be described herein with respect
to FIG. 4, the present disclosure extends and complements the
effectiveness of the system and method described in co-pending U.S.
patent application Ser. No. 12/126,642 by automatically providing a
source of information that associates "entities", e.g., candidate
answers for questions, to lexical types. Programmed components
build or populate a repository of information, e.g., a database or
knowledge base (KB) that can be used to accomplish the task of
computing one or more lexical types (LT) for each candidate answer
by facilitating automatic retrieval of "types" associated with
answer instances (answer-typing) in a KB as described with respect
to step 132b in FIG. 3. That is, given an instance (e.g., a word
such as a noun) the method automatically evaluates the LT specified
where the answer-typing data exists in a form that has a limited
amount of explicit structure, i.e. semi-structured. In one aspect,
the system and method produces a knowledge base of instances and
types used in matching.
[0037] Furthermore, as will be described herein with respect to
FIGS. 5-7, a system and method is provided that can be used to
automatically perform the task of "matching" query LATs to
candidate answer LTs such as when performing an answer scoring
function in the QA system of FIG. 1 that can be implemented at step
132c in FIG. 3. For candidate answer to instance matching, the
candidate answer is matched against an instance or instances within
the knowledge resource (e.g., a database or KB, or web page), where
the form the instance takes depends on the knowledge resource, and
lexical types (LT) associated with those instances are retrieved
from the knowledge base (KB).
[0038] The system and method for extracting and using typing
information from sources with limited structure and using that
extracted type information for answering questions is now
described. In practice, the method is generally performed during
the build of the KB, typically as off-line processes; however, can
be performed during real-time running of QA invocations, such as
described in U.S. patent application Ser. No. 12/126,642.
[0039] The system and method utilizes machine processing for
answering questions that employs special processing steps in which
information with limited structure is automatically extracted from
the various data sources and databases and re-represented in a form
suitable for machine (automated) processing. The results may then
be used in question answering as specified in commonly-owned,
co-pending U.S. patent application Ser. No. 12/126,642, whereby
given an input query LAT, an output is a judgment whether an entity
is an instance of a concept, e.g., by evaluating whether a thing,
e.g., noun, or a word, or entity, is of or has the Lexical Answer
Type specified.
[0040] An embodiment of the process for extracting collections of
entity-type pairs from semi-structured text is now described with
reference to FIG. 4. As shown in FIG. 4, the computer-implemented
method at 202 includes automatically identifying pages (e.g.,
documents, lists, and other data content) having limited structure
(i.e., "semi-structured") that can be stored, accessed or obtained
from a data source or database, etc. The data sources may include,
but are not limited to, e.g., public or private databases or
knowledge bases, Intranets, the Internet, web-pages, which can be
searched and/or crawled for content, e.g., using a search engine,
Web-based crawler, or other search tools. Further sources can
include a memory storage device such as a hard disk drive including
content such as, for example, a knowledge base, which can be
searched and iterated through to identify stored semi-structured
content to be processed.
[0041] The process for extracting collections of entity-type pairs
from semi-structured text may be done either as a pre-processing
step, or on-demand at the time that the question is received as
input to the system.
[0042] One kind of limited structure the computer-implemented
method identifies is pages or documents that are explicitly labeled
as having a list (i.e., a labeled list of elements). For example,
Wikipedia.RTM. (a registered service mark of Wikimedia Foundation,
Inc.) pages that are titled "List of" followed by a noun phrase).
Lists refer to lists of instances of a specified "type". For
example, http://en.wikipedia.org/wiki/List_of_poets presents a list
of poets; there is an entry in this list for Robert Frost, which
implies that Robert Frost is a poet.
[0043] Another type of limited structure is a tag, such as a
category in Wikipedia.RTM. or a Uniform Resource locator (URL) such
as a social bookmark (e.g., from http://delicious.com/). For
example, the Wikipedia page for Robert Frost has an explicit
Wikipedia category of "English-language poets" which is a
subcategory of "Poets by language" which is a subcategory of
"Poets." Thus one might conclude that Robert Frost is a poet.
Documents of this sort are typically designed to facilitate
comprehension by a human reader, unlike fully structured text
sources which are designed for machine processing.
[0044] Next, as shown in FIG. 4, the computer-implemented method at
205 includes automatically identifying item-delimiting markup. That
is, the lists (input) have a markup, e.g., that my be found using a
parsing function, for example, including, but not limited to, a
title, header or recitation of the word "list", bullet markers,
parentheses, a hypertext link which may be an indicator of a list,
at table column/row, or any marker which can be identified.
[0045] The identified list elements (content) may be tagged,
flagged or otherwise identified in the knowledge base or memory
e.g., in the disk, volatile or non-volatile memory storage, for
subsequent access of the identified contents, e.g., during QA
processing in one embodiment or prior to QA processing as a
preprocessing step in another embodiment. This includes discerning
what might not be desired list content, i.e. not part of list
entry. For example, in the list of US Poets, the content that would
include the year (e.g., 1852) would not be relevant if the list is
of poets, and thus may not be output or flagged.
[0046] The next automated step 210 shown in FIG. 4 includes
extracting Entity Lexical-type Pairs. That is, besides identifying
pages or documents having delimited items (e.g., a bulleted list of
contents) the identifying step further includes identifying the
lexical type of the delimited items that may be obtained from the
content, e.g., a title or header, in front of the list items. In
one embodiment, each lexical type and corresponding identified
item(s) are then paired and the pair is stored in the
knowledge-base using its internal mechanism for storing pairs,
e.g., in a relational data-base management system or a more
specialized combination of arrays and/or linked list structures.
The output of the processing of FIG. 4 is stored in the KB 21 of
FIG. 1.
[0047] One publicly-available example of such a list page that is
identified in step 202, 212 is
http://en.wikipedia.org/wiki/List_of_poets_from_the_United_States.
That page (as of July, 2010) includes on it a bulleted list 213
with poet names and some years in which they lived in parentheses.
This list page may be identified by a search engine or crawler, via
a web or Internet-search, or a search at a local networked or
attached hard disk drive providing an information base having
documents with semi-structured text.
[0048] Thus, in the example, as shown in FIG. 4 at 212, a page may
be identified as having content such as a title "List of Poets from
the United States" 218. Further, the identified list page may
include delimited mark-up 215 which may be searched and identified,
for example, bullets, as shown in front of each of the items
(entities) in the list 213. In the example, as shown in FIG. 4, at
215, those entities that can be recognized as items of the list 213
are marked in italicized text. In this example, a candidate answer
LT is assigned "poets from the United States", for example, from
the "list of" title 218.
[0049] At 220 in FIG. 4, implemented functions perform identifying
and extracting of pairs of the entities and their lexical types,
inferred from the fact that instances in the list are instances of
the lexically expressed "type" derived from the title of the list.
In this case, the instances include "Robert Frost" and others, and
the lexically expressed type is "poets from the United States." In
one embodiment, the LT may be stored in or flagged in a "Type"
field with the identified and extracted candidate answer instances
(entities or items) relating to the type or contents are stored as
linked structures in an "Entity" field. These fields are encoded in
a knowledge-base, which may be implemented using a relational
database management system or a more specialized combination of
arrays and/or linked list structures. The knowledge base thus
stores the identified, extracted entity-type relations including a
lexical type associated with one or more candidate answer entities
as a searchable content structure 220.
[0050] In some cases, precise extraction of the lexical type
requires distinguishing between the lexical type from the
organizational information. For example, "List of novels by point
of view" provides elements with lexical type "novel" and organizes
those elements by point of view. In contrast, a "List of novels by
Henry James" would provide elements with lexical type "novel by
Henry James" and does not give any information about the
organization of the list. On the other hand, it may not be
necessary to make such precise judgments, depending on the specific
LAT to Type Matching (LATTE) configuration used; for example the
"headword" passage matcher (see below), that will recognize "novel"
as the headword in either case.
[0051] Thus, example processing elements implemented by programmed
methods stored in a non-transitory storage medium, e.g., a memory
storage device, for handling "list of" pages include functionality
for: 1. Identifying the "type" that the list is enumerating.; 2.
Identifying the elements (entities) of the list (using the text of
the list elements, hyperlinks on the elements, titles of the
documents that those hyperlinks point to, etc.); and, 3. Inferring
that each element (identified in #2) is of the type (identified in
#1).
[0052] In FIG. 4, the first step in the automatic extraction of
entity/lexical-type pairs method, in one embodiment, implements a
search tool to simply select all of the text following "List of . .
. " documents, e.g., from an encyclopedic database source. That is,
step 202 above is the identification of list bearing pages within
the resource which process can include filtering pages whose title
starts with "List of" or, can involve an automatic classifier. In
one embodiment, the classifier is trained using statistical machine
learning: the learning process is given a set of examples of list
bearing pages and examples of non-list bearing pages and one or
more algorithms that generate features and the classifier
identifies some combination of those features that suggest that a
page is or is not list bearing. Features that are useful for this
purpose include structural elements (e.g., tables or bulleted
lists) and lexical elements (e.g., the presence of the word "list"
on the page).
[0053] In one embodiment, higher precision extraction requires
additional logic. For example, many list names provide additional
information about the organization of the list.
[0054] In one embodiment, step 205 in FIG. 4, is the abstraction of
the mark-up contained in the original pages into two elements
related to the extraction task: headers and bullets. Headers refer
to section markers, while "bullets", for example, designate a set
of mark-up elements that are used within the page to delimit list
elements. Other delimiters include but are not limited to: bullets,
table rows, etc.
[0055] The second step further addresses many different ways that
list membership can be indicated in semi-structured text (lists,
tables, etc.). For example, many lists include not only the
elements but also context and commentary regarding the elements
along with functionality implemented for separating the elements
from the other text. Furthermore, associating the extracted
elements with instances in a fixed knowledge-base may be
non-trivial. In some cases, the semi-structured content explicitly
provides that information (e.g., the entries in Wikipedia.RTM.
"List of" pages are usually hyperlinks, and the URLs that they link
do can be interpreted as unique identifiers for knowledge-base
instances). In other cases, an explicit instance lookup step may be
required, using the same mechanism that implements step 132a in
FIG. 3. That step is applied to candidate answers, and it is also
applied to extracted elements from a list.
[0056] In a further embodiment, explicit lookup may not be needed,
as the system may map names of instances to lexical types. The
process further addresses cases in which the list name includes a
conjunction. For example, any element of a "List of cities and
towns along the Ohio River" is a city or a town that is near the
Ohio River. In one embodiment, a list title in which the lexical
type is conjoined is split into two distinct lexical types and a
distinct entry in the knowledge-base is added for each. In another
embodiment, the entire lexical type including the conjunction may
be included in the knowledge-base and special logic for handing the
conjunction would be employed at phrase matching time (as described
herein below with respect to FIG. 8 processing at 350).
[0057] In FIG. 4, step 210 is the extraction of the actual
entity/lexical-type pairs. The lexical type is extracted from the
title of the page (e.g., by dropping the "List of" at the beginning
of the title, if any). Each entity is extracted with the aid of the
abstract bullet mark-up added in the prior step. For each bullet,
instances of a hyperlink closer to the bullet are searched for and,
if there are no hyperlinks between the bullet and the next one, no
entity is extracted for that bullet. The target for that hyperlink
will be the entity of the extracted entity/lexical-type pair. In
the event of missing hyperlinks, the technique is expanded to
extract entities as the most salient phrase (usually the subject)
in the bullet text, as determined by an automatic salience
annotator.
[0058] The three-step process shown in FIG. 4 for handling "list
of" pages is a specialization of a more general process for
extracting entity-type pairs from sources with limited
structure:
[0059] A more general formalism applies to sources that may not
include "list of" pages. The process steps employed in such cases
would include: 1. Identifying text that the limited structure
implies is a type; 2. Identifying text that the limited structure
implies is an entity; 3. Inferring that entities (identified in #2)
are instances of types (identified in #1). For example, one could
identify a category in a Wikipedia document from the existence of
the string "[[Category:" before the text and the string "]]" after
the text; the category may be a useful lexical type or it may
require normalization or inference to produce a lexical type. For
Wikipedia.RTM. categories, the entity that has the type is the one
that the page is about; its name is the title of the page. Thus one
extracts the title from the page in step #2 and associates it with
the type in step #3 by adding the title and the extracted lexical
type into a knowledge base. For example, the Wikipedia.RTM. page
with title "Robert Frost" has the following text in its source:
"[[Category:Sonneteers]]". In step #1, the word "Sonneteers" is
extracted as a lexical type using the pattern described above. In
step #2, the string "Robert Frost" is extracted from the document
title. In step #3, the pair ("Sonneteers", "Robert Frost") is added
to the knowledge base.
[0060] The results of the extraction process may further comprise
pairs of strings corresponding to entities and types. Pairs of this
sort may be used to answer questions using deferred type
evaluation. One step in deferred type evaluation is matching the
lexical answer type (LAT) to the known types of some candidate
answer. That step then becomes relatively easy if the candidate
answer is an entry in a fully-structured knowledge-base, because
knowledge-base entries (by definition) have formal, unambiguous
types whose relationships are known.
[0061] In one embodiment, matching the LAT to the type of some
candidate answer accounts for cases where the entities and types
were taken directly from text. For example, given a question asking
for a person, and a candidate answer that appears on a list of
poets, one can only conclude that the candidate has a valid type if
one can determine that all poets are people. This may be possible
using some dictionary or thesaurus resource, which is more reliable
if the LAT and the known types are disambiguated to specific word
senses. Logic for using resources such as dictionaries or thesauri
to determine whether terms like person and poet are consistent with
each other is encoded in Primitive Term Matchers, described
below.
[0062] In a further aspect, the function or process for matching
that can be implemented at 132c in FIG. 3, is now described in
connection with the process 300 described in connection with FIG.
5.
[0063] In one aspect, the evidence gathering and answer scoring
module 50 of QA system 10 is programmed to receive inputs including
a question LAT 302, obtained from processing of the query, and an
input including the candidate answer lexical type 312 obtained, for
example, from the KB. As shown in FIG. 5, in one embodiment, a text
processing function 325 is invoked, and, in one embodiment, one or
more computer-implemented processes for "LAT to Type Matching"
(LATTE) may be performed including: 1. Matching individual terms
(e.g., matching "country" to "nation"); and 2. Matching entire
phrases (e.g., matching "country with a national currency" to
"nation with its own currency"). The latter (phrase matchers) are
composed of the former (term matchers) plus control logic
implemented to determine which terms to try to match and how to
combine the scores of the term matches into a score for the phrase
match.
[0064] As shown in FIG. 5, in one embodiment, this text processing
325 may be implemented by an off the shelf text processor including
functionality for breaking the received question LAT 302 and
candidate lexical type 312 phrases into tokens, and determining the
grammatical relationships amongst the tokenized words or phrases,
or grammatical roles therebetween.
[0065] More particularly, the text processing component 325 divides
each of the inputs into distinct terms (e.g., via a tokenization
process). In one embodiment, the component further applies
syntactic and/or semantic analysis as required by the phrase
matcher or term matchers using established state-of-the-art
natural-language processing technologies; for example, it may use a
parser to identify the head word of its input, for use by the
head-word passage matcher. For example, breaking down a phrase
grammatically leaves a root word and word(s) that modify or relate
to the word (modifier (e.g., adjective) of a noun phrase including
a head-word (the noun)).
[0066] In one aspect, as referred to herein and shown in FIG. 5,
this text processing applied to produce both question LAT results
330 and the processed candidate answer lexical type (LT) results
340. These results 330, 340 may be organized as linked nodes in a
database which may be then subject to one or more matching
components embodied as programmed phrase and term matcher
components 345 in FIG. 5. These components generate a matching
score 346 indicating a degree of match.
[0067] FIG. 6 depicts a block diagram of an example phrase and term
matching methodology 345. These functions are embodied as (1) a
phrase match and (2) a term matcher. A computer-programmed phrase
matcher component 350 implements processing to receive an input
pair of phrases (each of which consists of one or more terms) and
produce as output a judgment regarding the degree of match between
the two phrases. A term matcher 355 takes as input of pair of terms
and produces as output a judgment regarding the degree of match
between the two terms.
[0068] In one aspect, the phrase matching component 350 provides
functionality to decide which pairs of terms to compare using the
term matcher 355 and how to combine the conclusions of the term
matcher into a conclusion regarding the degree of match of the
phrases. The term matcher is a "delegate" of the phrase matcher.
The term "delegate" refers to a relationship among two functional
units in which one of the units invokes the other to provide an
implementation of some of its functionality.
[0069] In view of FIG. 6, one example of a phrase matcher 350 is a
headword phrase matcher 351, which treats each phrase as a headword
plus a collection of modifiers and attempts to match headwords to
headwords and modifiers to modifiers. That is, a headword phrase
matcher 351 component applies its specified term matcher to the
grammatical head-word of the LAT and the grammatical head-word of
the candidate answer lexical type. For example, such a matcher
configured with a "string-equals" term matcher would give a high
match score to "European nation" and "nation on Earth" because both
phrases have the head-word "nation."
[0070] In one embodiment, shown in FIG. 6, there are two broad
sub-classes of term-matching components implemented: 1. Primitive
term matchers 355a employ a single, specific matching strategy and
provide a score based on the degree of match; and, 2. Aggregate
term matchers 355b combine multiple matchers (which may be
primitive or aggregate).
[0071] A Primitive term matcher 355a employs a strategy to
determine the extent to which the input pair(s) of terms match each
other. A simple example of a primitive term matcher is a
"text-equals" primitive term matcher 356 which considers a pair of
terms with identical text to match and any other pair of terms to
not match. For example, text-equals primitive term matcher 356
provides a score of 1.0 to strings that are identical and 0.0 to
strings that are not identical. A more complex example of a
primitive term matcher is the "geopolitical" term matcher 366,
which applies only to pairs of terms that are both geopolitical
entities and gives high scores when the terms are equivalent (e.g.,
"U.S." and "America") and/or closely related (e.g., "U.S." and
"Texas"). A more complex example of a primitive term matcher 355a
is a "thesaurus synonym" term matcher (not shown), which provides a
high score to terms that are synonyms in a known thesaurus; such a
matcher may be more precise if it uses the surrounding context to
disambiguate the terms. Another example of a primitive term matcher
355a is a "string-edit-distance" term matcher (not shown), which
gives a high score to terms that have approximately the same
letters (e.g., elephant .about.=elephand), which can be very useful
in contexts where minor spelling errors are common.
[0072] In FIG. 6, an aggregate term matcher 355b employs one or
more "delegate" term matchers each of which may be primitive 355a
or aggregate 355b. If a delegate term matcher is an aggregate, it
will also have delegates, which will also be primitive or
aggregate. Each aggregate term matcher invokes one or more of its
delegates according to its control logic. The aggregate term
matcher implements functionality for combining the score of those
delegates according to its combination logic. Each primitive term
matcher employs internal logic for computing a score and returning
the score.
[0073] In one embodiment, the matchers that are combined together
by a single aggregate term matcher 355b are delegates of that
matcher. Each aggregate term matcher implements a control flow as
described in FIG. 7 among its delegates and some strategy for
combining the results of the delegates into a single score.
[0074] Aggregate term matcher combining functionality includes flow
control among its delegates and implements strategy for combining
the results of applying the delegate term matchers into a final
conclusion regarding the degree of match between the pair of terms.
In one embodiment, the combined results of the delegates generate a
single score. For example, an aggregate term matcher 355b runs all
of its delegates and then returns the sum of all of the scores of
all the delegates. An example of an aggregate term matcher 355b is
a maximum score aggregate term matcher 367, which takes an input
pair of terms, applies each of its delegates to that pair of terms,
and returns the maximum score across all of the delegates. In
another embodiment, an aggregate term matcher 355b includes a
product of scores aggregate term matcher 368 which takes an input
pair of terms, applies each of its delegates to that pair of terms,
and multiplies together all of the scores of all of the delegates.
In one embodiment, an aggregate term matchers may use a statistical
model derived from machine learning to combine the scores of the
delegates into a score for the aggregate. In one embodiment,
logistic regression is the machine learning method that takes
labeled training instances with numerical features and produces a
statistical model that can be used to classify instances with
numerical features; it does so by assigning a numerical weight to
each feature, and then computing a score by multiplying the
numerical feature scores by the weights.
[0075] FIG. 7 illustrates an example flow of control 500 for the
LATTE process and score production among the components depicted in
FIG. 6. At 505, after receiving as input text strings representing
the phrases to be matched, the Phrase matcher 350a1 at 510
implements logic to choose pairs of terms, one from each of the two
input phrases--the phrases comprising: question terms and passage
terms to match. At 510, after the phrase matcher 350a1 determines
which terms to invoke, and if any terms are delegate terms, will
determine whether a delegate term is aggregated (i.e., includes
multiple terms). At 510, if a delegate term is aggregated, the
phrase matcher invokes its delegate term aggregate matcher
implementing functions 355a1; otherwise, will invoke a primitive
term matcher 355b. At 512, aggregate term matcher functions 355a1
selects one of its delegates to match that pair of terms; that
delegate is either a primitive 355b or the lower-level aggregate
matcher implementing functionality 355a1.
[0076] At 514, primitive term matcher 355b uses its single, atomic
term matching logic to compute a score indicating how well the pair
of terms match. In one embodiment, atomic algorithms may determine
whether a pair of terms mean the same thing (e.g., using a
dictionary with synonyms). The result is returned to whichever
component invoked it; that component including a higher level
aggregate matcher functionality 355a2, or the phrase matcher
functionality 350a2.
[0077] At 516, the aggregate term matcher functions 355a2 checks to
see if it has any additional delegate term matchers to try to apply
to the given pair of terms in which case it returns to either the
lower level aggregate matcher functions 355a1 or a higher level
aggregate matcher functionality 355a3.
[0078] At 518, with no more delegate term matchers to apply, the
aggregate term matcher functions 355a3 computes a final match score
for the pair of terms by combining results from all of its
delegates. That result is returned to whatever component invoked
the aggregate term matcher, either the higher-level aggregate
functionality 355a2 or the phrase matcher functions 350a2.
[0079] At 520, the phrase matcher functions 350a2 checks to see if
there are any other pairs of terms that it needs to match in order
to determine how well the phrase matches. Thus the phrase matcher
350a2 invokes phrase matcher functions 350a1; otherwise, will
invoke further phrase matcher functionality 350a3. That is, at 525,
with no more pairs of terms to match, the phrase matcher functions
350a3 computes a final score for the two input phrases by combining
results from each call to its delegate term matcher. That result is
returned as the final output of the phrase matching process which
ends at 530.
[0080] Utilizing matching process as described in FIG. 7, FIG. 8
illustrates an example flow control for the LATTE process of FIGS.
5-6 applied for a non-limiting example of an initial query or
question "What American poet published his first book in 1913 while
living in England?" A question analysis component identifies
"American poet" as a LAT in this question according to common QA
system processing such as described in U.S. patent application Ser.
No. 12/126,642. A candidate answer generation component of the
question answering system identifies, for the example, a search
candidate answer: e.g., Robert Frost. In one embodiment, the method
determines whether this candidate answer is an instance of the LAT
(i.e., is "Robert Frost" an "American poet") using semi-structured
resources such as obtained in the real-time or off-line manner
described above with respect to FIG. 4.
[0081] In the flow control for the example matching process of FIG.
7 applied for the non-limiting example, when the LATTE process (LAT
to lexical type matching process) is automatically run as part of
the QA system process, a specific question LAT 302 and a specific
candidate answer lexical type 312 is available. In the example, the
question LAT that was extracted is "American poet" 368. The example
candidate answer is "Robert Frost." The pairs extracted in FIG. 4
include a pair relating "Robert Frost" to the lexical type "poets
from the United States," which is a candidate-answer lexical type
369. In the example embodiment described with respect to processing
of FIG. 6, the question LAT 302 "American poet" 368 and the
candidate answer lexical type 312 "poets from the United States"
369 are received as inputs to text processing component 325.
[0082] In FIG. 8, text processing 325 by a text processor provides
outputs (e.g., splitting into tokens, labeling with grammatical
roles) including versions of the received question LAT 330 and the
candidate answer lexical type 340. The phrase matcher 350 will then
receive these inputs and compare them against each other. It does
so by applying its specified term matcher to the terms it selects
from its inputs. The algorithm for selecting terms is specific to
any specific instantiation of a phrase matcher; for example, the
head-word phrase matcher selects the head-word of each phrase to
match and also attempts to match each modifier of the head-word in
the first input phrase to each modifier of the second phrase.
[0083] In this non-limiting example, the text processing 325
includes performing an analysis to identify head-word and provides
lemma forms for terms and grammatical relationships among those
terms; a natural-language parser, provides all of that information.
For the example processing of FIG. 6, the results 330 of processing
the example question LAT include, for example, the identified
head-word for the example question LAT--a root or lemmatized
head-word "Poet" 370 (e.g., "poets" is lemmatized to head-word
"poet"), and further detecting that "American" 371 is acting as
noun-modifying adjective as indicated by connector 373 labeled
"nadj" indicating a grammatical relations amongst the terms, e.g.
an adjective modifying the noun "poet"). Likewise, text processing
results 340 of the candidate answer lexical type 312 includes: the
lemmatized result head-word "Poet" 342 (e.g., "poets" is lemmatized
to head-word "poet"). The processing further includes automatically
detecting the prepositional phrase "from the United States" 347
where it is determined the preposition "from", object of
preposition "United States", and determiner "the" for the processed
candidate answer lexical type 312 (typically this information would
be provided by a natural-language parser). The root or head-word
and all grammatical relationships may be represented in software
program as nodes interconnected by labeled connectors, e.g.,
connector 343a labeled "prep" pointing head-word 342 to the
preposition "from"; the connector 343b labeled "obj" pointing the
preposition "from" to the object "United States"; and, the
connector 343c labeled "det" pointing the object "United States" to
the determiner "the" indicating a grammatical relations amongst the
phrase terms, including determiner "the" 344.
[0084] The phrase matcher 350 then automatically applies a
specified term matcher to compare processed results (terms) 330 in
the question's lexical answer type 302 to results (terms) 340 in a
lexical type of the candidate answer 312.
[0085] First, the phrase matcher 350 determines which terms in the
question's lexical answer type to attempt to match to terms in the
candidate answer's lexical type. For example, the phrase matcher
determines that the head-word (i.e., the root node for each graph,
in both cases, labeled "poet") is in the same logical position;
this graph is derived from the output of the natural-language
parser. The phrase matcher 350 also determines that the
noun-adjective modifier 371 (e.g., "American") fills a comparable
role to the object of the preposition 345 in the processed
candidate answer lexical type 345 (e.g., "United States"). The
aggregate term matcher 355 is then responsible for determining if
those terms do actually match. In this example, two delegate
primitive term matchers 355 are used: a "text-equals" term matcher
356 that receives the inputs and implements functions to determine
and conclude that "poet" 370 from the question LAT and "poet" 342
from the candidate answer lexical type are exactly equal and a
specialized geopolitical matcher 366 that determines from the input
that "American" 371 and "United States" 345 are consistent with
each other.
[0086] More particularly, a programmed processing system is
configured with a headword phrase matcher 351 that is configured
with a term matcher, e.g., the maximum score aggregate term matcher
355a. The maximum score aggregate term matcher 355a is configured
with two delegate term matchers: the text-equals primitive matcher
356 and the geopolitical term matcher 366. This phrase matcher 351
receives as input two phrases: "American poet" 368 and "poet from
the United States" 369. Prior to executing the phrase matcher, text
analysis has been run that has identified "poet" as the headword
342, 370 of each of these phrases. It also indicates grammatical
relationships among the terms, e.g., that "American" modifies poet
in the first phrase and that "United States" modifies poet in the
second. The phrase matcher takes pairs of terms from each results
330, 340 and uses the results of text analysis to decide to compare
"poet" in the first phrase to "poet" in the second phrase (because
each is the headword of its phrase). Consequently, it invokes the
aggregate term matcher 355b on this pair of terms. The aggregate
term matcher 355b then invokes the text-equals primitive term
matcher, which observes that the strings are identical and gives
the pair of terms a high score. The aggregate term matcher also
invokes the geopolitical primitive term matcher, which does not
provide an opinion because it only applies to geopolitical
entities. The aggregate term matcher 355b then computes the maximum
of these results, which is the high score from the text-equals
primitive term matcher. Thus it has a high score for the pair
("poet", "poet"). Next, the phrase matcher 351 decides to compare
"American" to "United States" because both are modifiers of the
headword. It does so by invoking the aggregate term matcher 355b.
The aggregate term matcher invokes the text-equals primitive term
matcher 356, which observes that the strings are not identical and
gives the pair of terms a 0 score. The aggregate term matcher also
invokes the geopolitical primitive term matcher 366 which uses a
knowledge-base of geopolitical entities which asserts that
"American" and "United States" refer to the same place; the
geopolitical primitive term matcher 366 thus returns a high score
for this pair of terms. The aggregate term matcher takes these two
results (a 0 score from text-equals and a high score from
geopolitical) and takes the maximum among them, which is a high
score.
[0087] It returns this high score to the phrase matcher, which now
has high scores for both the headword ("poet") and the modifier
("American") in the first phrase matching corresponding terms in
the second phrase. It combines these and returns a conclusion
indicating that the two phrases match very well.
[0088] Having found the question LAT and the candidate answer
lexical type match, and the system concludes that the candidate
answer "Robert Frost," meets the type requirements for this
question.
[0089] Thus, the two components: the process for extracting
collections of entity-type pairs from semi-structured text of FIG.
4; and, the process for matching in FIGS. 5-7 are tightly
integrated and complement each other when used in QA system, and
further also be used in radically different devices.
[0090] FIG. 1 shows a system diagram described in U.S. patent
application Ser. No. 12/126,642 depicting a high-level logical
architecture 10 and methodology in which the system and method for
deferred type evaluation using text with limited structure is
employed in one embodiment.
[0091] Generally, as shown in FIG. 1, the high level logical
architecture 10 includes a Query Analysis module 20 implementing
functions for receiving and analyzing a user query or question. The
term "user" may refer to a person or persons interacting with the
system, or refers to a computer system 22 generating a query by
mechanical means, and where the term "user query" refers to such a
mechanically generated query and context 19'. A candidate answer
generation module 30 is provided to implement a search for
candidate answers by traversing structured, semi structured and
unstructured sources contained in a Primary Sources module 11 and
in an Answer Source Knowledge Base (KB) module 21 containing
collections of relations and lists extracted from primary sources.
All the sources of information can be locally stored or distributed
over a network, including the Internet.
[0092] The Candidate Answer generation module 30 of architecture 10
generates a plurality of output data structures containing
candidate answers based upon the analysis of retrieved data. In
FIG. 1, an Evidence Gathering and answer scoring module 50
interfaces with the primary sources 11 and knowledge base 21 for
concurrently analyzing the evidence based on passages having
candidate answers, and scores each of candidate answers, in one
embodiment, as parallel processing operations. In one embodiment,
the architecture may be employed utilizing the Common Analysis
System (CAS) candidate answer structures as is described in
commonly-owned, issued U.S. Pat. No. 7,139,752, the whole contents
and disclosure of which is incorporated by reference as if fully
set forth herein.
[0093] As depicted in FIG. 1, the Evidence Gathering and Answer
Scoring module 50 comprises a Candidate Answer Scoring module 40
for analyzing a retrieved passage and scoring each of candidate
answers of a retrieved passage. The Answer Source Knowledge Base
(KB) 21 may comprise one or more databases of structured or
semi-structured sources (pre-computed or otherwise) comprising
collections of relations (e.g., Typed Lists). In an example
implementation, the Answer Source knowledge base may comprise a
database stored in a memory storage system, e.g., a hard drive.
[0094] An Answer Ranking module 60 provides functionality for
ranking candidate answers and determining a response 99 returned to
a user via a user's computer display interface (not shown) or a
computer system 22, where the response may be an answer, or an
elaboration of a prior answer or request for clarification in
response to a question--when a high quality answer to the question
is not found. A machine learning implementation is further provided
where the "answer ranking" module 60 includes a trained model
component (not shown) produced using a machine learning techniques
from prior data.
[0095] The processing depicted in FIG. 1, may be local, on a
server, or server cluster, within an enterprise, or alternately,
may be distributed with or integral with or otherwise operate in
conjunction with a public or privately available search engine in
order to enhance the question answer functionality in the manner as
described. Thus, the method may be provided as a computer program
product comprising instructions executable by a processing device,
or as a service deploying the computer program product. The
architecture employs a search engine (e.g., a document retrieval
system) as a part of Candidate Answer Generation module 30 which
may be dedicated to searching the Internet, a publicly available
database, a web-site (e.g., IMDB.com) or, a privately available
database. Databases can be stored in any storage system,
non-volatile memory storage systems, e.g., a hard drive or flash
memory, and can be distributed over the network or not.
[0096] As mentioned, the system and method of FIG. 1 makes use of
the Common Analysis System (CAS), a subsystem of the Unstructured
Information Management Architecture (UIMA) that handles data
exchanges between the various UIMA components, such as analysis
engines and unstructured information management applications. CAS
supports data modeling via a type system independent of programming
language, provides data access through a powerful indexing
mechanism, and provides support for creating annotations on text
data, such as described in
(http://www.research.ibm.com/journal/sj/433/gotz.html) incorporated
by reference as if set forth herein. It should be noted that the
CAS allows for multiple definitions of the linkage between a
document and its annotations, as is useful for the analysis of
images, video, or other non-textual modalities (as taught in the
herein incorporated reference U.S. Pat. No. 7,139,752).
[0097] In one embodiment, the UIMA may be provided as middleware
for the effective management and interchange of unstructured
information over a wide array of information sources. The
architecture generally includes a search engine, data storage,
analysis engines containing pipelined document annotators and
various adapters. The UIMA system, method and computer program may
be used to generate answers to input queries. The method includes
inputting a document and operating at least one text analysis
engine that comprises a plurality of coupled annotators for
tokenizing document data and for identifying and annotating a
particular type of semantic content. Thus it can be used to analyze
a question and to extract entities as possible answers to a
question from a collection of documents.
[0098] As further shown in greater detail in the architecture
diagram of FIG. 1, the "Query Analysis" module 20 receives an input
that comprises the query 19 entered, for example, by a user via
their web-based browser device. An input query 19 may comprise a
text string. The query analysis block 20 includes additionally a
Lexical Answer Type (LAT) block 200 that implements functions and
programming interfaces to provide additional constraints on the
answer type (LAT). The computation in the block 20 comprises but is
not limited to the Lexical Answer Type. The LAT block 200 includes
certain functions/sub-functions (not shown) to determine the
LAT.
[0099] As mentioned above, a LAT of the question/query is the type
(i.e. the descriptor) of the referent of the entity that is a valid
answer to the question. In practice, LAT is the descriptor of the
answer detected by a natural language understanding module
comprising a collection of patterns and/or a parser with a semantic
interpreter. With reference to the Lexical Answer Type (LAT) block
200, in the query analysis module 20 of FIG. 1, the LAT represents
the question terms that identify the semantic type of the correct
answer. In one embodiment, as known, a LAT may be detected in a
question through pattern LAT detection rules. These rules are
implemented and can be encoded manually or learned by machine
automatically through association rule learning. In this case, the
natural language understanding model can be limited to
implementation the rules.
[0100] FIG. 2 is a flow diagram of a computer programmed method for
conducting questions and answers with deferred type evaluation in
one embodiment as described in Ser. No. 12/126,642. Generally, in
the method of "deferred type evaluation" depicted in FIG. 2, a
first processing step 112 represents the step of receiving, at a
processing device, an input query, and generating a data structure,
e.g., a CAS structure, including a question string and context for
input to a Lexical Answer Type (LAT) processing unit block 200
(FIG. 1) where, as indicated at step 115, the Query is analyzed and
lexical answer type (LAT) is computed. As a result of processing in
the LAT processing component, as run at step 115, there is
generated an output data structure, e.g., a CAS structure,
including the computed LAT and possibly additional terms from the
original question.
[0101] As result of processing in the LAT block 200 then, as
typified at step 120, there is generated an output data structure,
e.g., a CAS structure, including the computed original query
(terms, weights) (as described in co-pending U.S. patent
application Ser. No. 12/152,441 the whole contents and disclosure
of which is incorporated by reference as if fully set forth
herein.
[0102] Returning to FIG. 2, at processing step 122, there is
performed the step of searching for candidate answer documents, and
returning the results.
[0103] As a result of processing in a candidate answer generation
module, as typified at step 122, there is generated an output data
structure, e.g., a CAS structure, including all of the documents
found from the data corpus (e.g., primary sources and knowledge
base).
[0104] In FIG. 2, step 128, there is depicted the step of analyzing
each document for a candidate answer to produce a set of candidate
answers which may be output as a CAS structure using LAT (the
lexical answer type).
[0105] For the example questions discussed herein, as a result of
processing in the candidate answer generation module 30, as
typified at step 132, FIG. 2, those candidate answers that are
found in the document is checked against the query LAT requirement
and will be returned as answer(s) at step 133 (based on their
scores).
[0106] FIG. 3 described herein above, is a flow diagram
illustrating score production performed at step 132 of the flow
diagram of FIG. 2.
[0107] In particular, as shown in FIG. 3, step 132 implements the
following steps: for each candidate answer received, there is
performed matching the candidate against instances in the database
(step 132a) which results in generating an output data structure,
e.g., a CAS structure, including the matched instances; retrieving
lexical types (LT) associated with those instances in the knowledge
base (KB) (step 132b); and, at step 132c attempting to match LAT(s)
with lexical types (LT), producing a score representing the degree
of match.
[0108] More particularly, the candidate answer LT and query LAT(s)
are represented as lexical strings. Production of the score,
referred to herein as the "TyCor" (Type Coercion) score, is
comprised of three steps: candidate to instance matching, instance
to type association extraction, and LAT to type matching. The score
reflects the degree to which the candidate may be "coerced" to the
LAT, where higher scores indicate a better coercion.
[0109] In candidate to instance matching, the candidate is matched
against an instance or instances within the knowledge resource,
where the form the instance takes depends on the knowledge
resource. With a structured knowledge base, instances may be
entities, with an encyclopedic source such as Wikipedia instances
may be entries in the encyclopedia, with lexical resources such as
the WordNet.RTM. lexical database (A trademark of the Trustees of
Princeton University) instances may be synset entries (sets of
synonyms), and with unstructured document (or webpage) collections,
instances may be any terms or phrases occurring within the text. If
multiple instances are found, a rollup using an aggregation
function is employed to combine the scores from all candidates. If
no suitable instance is found, a score of 0 is returned.
[0110] Next, instance association information is extracted from the
resource. This information associates each instance with a type or
set of types. Depending on the resource, this may take different
forms; in a knowledge base, this corresponds to particular
relations of interest that relate instances to types, with an
encyclopedic source, this could be lexical category information
which assigns a lexical type to an entity, with lexical resources
such as WordNet.RTM., this is a set of lexical relations, such as
hyponymy, over synsets (e.g. "artist" is a "person"), and with
unstructured document collections this could be co-occurrence or
proximity to other terms and phrases representing type.
[0111] Then, each LAT is then attempted to match against each type.
A lexical manifestation of the type is used. For example, with
encyclopedias, this could be the string representing the category,
with a lexical resource such as WordNet.RTM., this could be the set
of strings contained within the synset. The matching is performed
by using string matching or additional lexical resources such as
Wordnet.RTM. to check for synonymy or hyponymy between the LAT and
type. Special logic may be implemented for types of interest; for
example person matcher logic may be activated which requires not a
strict match, synonym, or hyponym relation, but rather that both
LAT and type are hyponyms of the term "person". In this way, "he"
and "painter", for example, would be given a positive score even
though they are not strictly synonyms or hyponyms. Finally, the set
of pairs of scores scoring the degree of match may be resolved to a
single final score via an aggregation function.
[0112] Thus, in an implementation set forth in steps 132a-132c of
FIG. 3, for the example question described herein, each candidate
answer type (LT) in the document is automatically checked against
the LAT requirement. This may be performed by the Candidate Answer
Scoring block 40, shown in FIG. 1, as part of the evidence
gathering and answer scoring module 50, and particularly, a
Candidate Answer Type Analysis module 400 that produces a
probability measure that Candidate Answer is of the correct type
based, e.g., on a grammatical and semantic analysis of the document
with which the Candidate Answer appears. In one embodiment, this
processing entails using an automated scoring function that
compares candidate answer lexical types (LTs) to the query LAT and
producing a score for each candidate answer. The a scoring function
can be expressed as a weighted combination of different typing
scores, and, in one embodiment it may be expressed as
TyCorScore=0.2.*TyCorWordNet+0.5*TyCorKB+0.4*TyCorDoc
[0113] This expresses the preferences for more organized sources
such as knowledge bases (KB), followed by type matching in a
retrieved document, and synonyms being least preferred way of
matching types.
[0114] Other combinations of scores are possible, and the optimal
scoring function can be learned as described in the co-pending U.S.
patent application Ser. No. 12/152,411 entitled SYSTEM AND METHOD
FOR PROVIDING ANSWERS TO QUESTIONS, filed May 14, 2008, the content
and disclosure of which is incorporated by reference as if fully
set forth herein.
[0115] The scoring function itself is a mathematical expression,
that--in one embodiment--could be based on the logistic regression
function (a composition of linear expressions with the exponential
function), and may be applied to a much larger number of typing
scores.
[0116] The output of the "Candidate Answer Scoring" module 40 is a
CAS structure having a list of answers with their scores given by
the processing modules in the answer scoring modules included in
the Candidate Answer Scoring block 40 of the evidence gathering and
answer scoring module 50. In one embodiment, these candidate
answers are provided with TyCor matching score as described herein
above.
[0117] Finally, returning to FIG. 2, at step 133, the top candidate
answers (based on their TyCor scores) are returned.
[0118] In one embodiment, the above-described modules of FIGS. 1,
4, 5-7 can be represented as functional components in UIMA is
preferably embodied as a combination of hardware and software for
developing applications that integrate search and analytics over a
combination of structured and unstructured information. The
software program that employs UIMA components to implement end-user
capability is generally referred to as the application, the
application program, or the software application.
[0119] The UIMA high-level architecture, one embodiment of which is
illustrated in FIG. 1, defines the roles, interfaces and
communications of large-grained components that cooperate to
implement UIM applications. These include components capable of
analyzing unstructured source artifacts, such as documents
containing textual data and/or image data, integrating and
accessing structured sources and storing, indexing and searching
for artifacts based on discovered semantic content.
[0120] Although not shown, a non-limiting embodiment of the UIMA
high-level architecture includes a Semantic Search Engine, a
Document Store, at least one Text Analysis Engine (TAE), at least
one Structured Knowledge Source Adapter, a Collection Processing
Manager, at least one Collection Analysis Engine, all interfacing
with Application logic. In one example embodiment, the UIMA
operates to access both structured information and unstructured
information to generate candidate answers and an answer in the
manner as discussed herein. The unstructured information may be
considered to be a collection of documents, and can be in the form
of text, graphics, static and dynamic images, audio and various
combinations thereof.
[0121] Aspects of the UIMA are further shown in FIG. 7, where there
is illustrated a Analysis Engine (AE) 600 that can be a component
part of the Text Analysis Engine (TAE). Included in the AE 600 is a
Common Analysis System (CAS) 610, an annotator 620 and a controller
630. A second embodiment of a TAE (not shown) includes an aggregate
Analysis Engine composed of two or more component analysis engines
as well as the CAS, and implements the same external interface as
the AE 600.
Common Analysis System 610
[0122] The Common Analysis System (CAS) 610 is provided as the
common facility that all Annotators 620 use for accessing and
modifying analysis structures. Thus, the CAS 610 enables
coordination between annotators 620 and facilitates annotator 620
reuse within different applications and different types of
architectures (e.g. loosely vs. tightly coupled). The CAS 610 can
be considered to constrain operation of the various annotators.
[0123] The CAS 610 principally provides for data modeling, data
creation and data retrieval functions. Data modeling preferably
defines a tree hierarchy of (data) types, as shown in the example
Table 1 provided below. The types have attributes or properties
referred to as features. In preferred embodiments, there are a
small number of built-in (predefined) types, such as integer (int),
floating point (float) and string; UIMA also includes the
predefined data type "Annotation". The data model is defined in the
annotator descriptor, and shared with other annotators. In the
Table 1, some "Types" that are considered extended from prior art
unstructured information management applications to accommodate
question answering in the preferred embodiment of the invention
include:
TABLE-US-00001 TABLE 1 TYPE's PARENT TYPE (or feature) (or feature
type) Query Record Top Query Query Record Query Context Query
Record Candidate Answer Record Annotation Candidate Answer
Candidate Answer Record Feature: CandidateAnswerScore Float
QueryLexical Answer Type Annotation CandidateAnswer LT Annotation
Feature: TyCorScore Float
[0124] In Table 1, for example, all of the question answering types
(list in the left column) are new types and extend either another
new type or an existing type (shown in the right column). For
example, both Query and Query Context are kinds of Query Record, a
new type; while Candidate Answer Record extends the UIMA type
Annotation, but adds a new feature CandidateAnswerScore which is a
Float. In addition, Table 1 describes the query LAT as having a
UIMA Annotation type; CandidateAnswerLT is also an Annotation, but
with an additional featue TyCorScore of type Float.
[0125] CAS 610 data structures may be referred to as "feature
structures." To create a feature structure, the type must be
specified (see TABLE 1). Annotations (and--feature structures) are
stored in indexes.
[0126] The CAS 610 may be considered to be a collection of methods
(implemented as a class, for example, in Java or C++) that
implements an expressive object-based data structure as an abstract
data type. Preferably, the CAS 610 design is largely based on a TAE
Feature-Property Structure, that provides user-defined objects,
properties and values for flexibility, a static type hierarchy for
efficiency, and methods to access the stored data through the use
of one or more iterators.
[0127] The abstract data model implemented through the CAS 610
provides the UIMA 100 with, among other features: platform
independence (i.e., the type system is defined declaratively,
independently of a programming language); performance advantages
(e.g., when coupling annotators 610 written in different
programming languages through a common data model); flow
composition by input/output specifications for annotators 610 (that
includes declarative specifications that allow type checking and
error detection, as well as support for annotators (TAE) as
services models); and support for third generation searching
procedures through semantic indexing, search and retrieval (i.e.
semantic types are declarative, not key-word based).
[0128] The CAS 610 provides the annotator 620 with a facility for
efficiently building and searching an analysis structure. The
analysis structure is a data structure that is mainly composed of
meta-data descriptive of sub-sequences of the text of the original
document. An exemplary type of meta-data in an analysis structure
is the annotation. An annotation is an object, with its own
properties, that is used to annotate a sequence of text. There are
an arbitrary number of types of annotations. For example,
annotations may label sequences of text in terms of their role in
the document's structure (e.g., word, sentence, paragraph etc), or
to describe them in terms of their grammatical role (e.g., noun,
noun phrase, verb, adjective etc.). There is essentially no limit
on the number of, or application of, annotations. Other examples
include annotating segments of text to identify them as proper
names, locations, military targets, times, events, equipment,
conditions, temporal conditions, relations, biological relations,
family relations or other items of significance or interest.
[0129] Typically an Annotator's 620 function is to analyze text, as
well as an existing analysis structure, to discover new instances
of the set of annotations that it is designed to recognize, and
then to add these annotations to the analysis structure for input
to further processing by other annotators 620.
[0130] In addition to the annotations, the CAS 610 of FIG. 7 may
store the original document text, as well as related documents that
may be produced by the annotators 620 (e.g., translations and/or
summaries of the original document). Preferably, the CAS 610
includes extensions that facilitate the export of different aspects
of the analysis structure (for example, a set of annotations) in an
established format, such as XML.
[0131] More particularly, the CAS 610 is that portion of the TAE
that defines and stores annotations of text. The CAS API is used
both by the application and the annotators 620 to create and access
annotations. The CAS API includes, preferably, at least three
distinct interfaces. A Type system controls creation of new types
and provides information about the relationship between types
(inheritance) and types and features. One non-limiting example of
type definitions is provided in TABLE 1. A Structure Access
Interface handles the creation of new structures and the accessing
and setting of values. A Structure Query Interface deals with the
retrieval of existing structures.
[0132] The Type system provides a classification of entities known
to the system, similar to a class hierarchy in object-oriented
programming. Types correspond to classes, and features correspond
to member variables. Preferably, the Type system interface provides
the following functionality: add a new type by providing a name for
the new type and specifying the place in the hierarchy where it
should be attached; add a new feature by providing a name for the
new feature and giving the type that the feature should be attached
to, as well as the value type; and query existing types and
features, and the relations among them, such as "which type(s)
inherit from this type".
[0133] Preferably, the Type system provides a small number of
built-in types. As was mentioned above, the basic types are int,
float and string. In a Java implementation, these correspond to the
Java int, float and string types, respectively. Arrays of
annotations and basic data types are also supported. The built-in
types have special API support in the Structure Access
Interface.
[0134] The Structure Access Interface permits the creation of new
structures, as well as accessing and setting the values of existing
structures. Preferably, this provides for creating a new structure
of a given type; getting and setting the value of a feature on a
given structure; and accessing methods for built-in types. Feature
definitions are provided for domains, each feature having a
range.
[0135] In an alternative environment, modules of FIGS. 1, 2 can be
represented as functional components in GATE (General Architecture
for Text Engineering) (see:
http://gate.ac.uk/releases/gate-2.0alpha2-build484/doc/userguide.html).
Gate employs components which are reusable software chunks with
well-defined interfaces that are conceptually separate from GATE
itself. All component sets are user-extensible and together are
called CREOLE--a Collection of REusable Objects for Language
Engineering. The GATE framework is a backplane into which plug
CREOLE components. The user gives the system a list of URLs to
search when it starts up, and components at those locations are
loaded by the system. In one embodiment, only their configuration
data is loaded to begin with; the actual classes are loaded when
the user requests the instantiation of a resource.). GATE
components are one of three types of specialized Java Beans: 1)
Resource: The top-level interface, which describes all components.
What all components share in common is that they can be loaded at
runtime, and that the set of components is extendable by clients.
They have Features, which are represented externally to the system
as "meta-data" in a format such as RDF, plain XML, or Java
properties. Resources may all be Java beans in one embodiment. 2)
ProcessingResource: Is a resource that is runnable, may be invoked
remotely (via RMI), and lives in class files. In order to load a PR
(Processing Resource) the system knows where to find the class or
jar files (which will also include the metadata); 3)
LanguageResource: Is a resource that consists of data, accessed via
a Java abstraction layer. They live in relational databases; and,
VisualResource: Is a visual Java bean, component of GUIs, including
of the main GATE gui. Like PRs these components live in .class or
.jar files.
[0136] In describing the GATE processing model any resource whose
primary characteristics are algorithmic, such as parsers,
generators and so on, is modelled as a Processing Resource. A PR is
a Resource that implements the Java Runnable interface. The GATE
Visualisation Model implements resources whose task is to display
and edit other resources are modelled as Visual Resources. The
Corpus Model in GATE is a Java Set whose members are documents.
Both Corpora and Documents are types of Language Resources (LR)
with all LRs having a Feature Map (a Java Map) associated with them
that stored attribute/value information about the resource.
FeatureMaps are also used to associate arbitrary information with
ranges of documents (e.g. pieces of text) via an annotation model.
Documents have a DocumentContent which is a text at present (future
versions may add support for audiovisual content) and one or more
AnnotationSets which are Java Sets.
[0137] As UIMA, GATE can be used as a basis for implementing
natural language dialog systems and multimodal dialog systems
having the disclosed question answering system as one of the main
submodules. The references, incorporated herein by reference above
(U.S. Pat. Nos. 6,829,603 and 6,983,252, and 7,136,909) enable one
skilled in the art to build such an implementation.
[0138] FIG. 10 illustrates an exemplary hardware configuration of a
computing system 400 in which the present system and method may be
employed. The hardware configuration preferably has at least one
processor or central processing unit (CPU) 411. The CPUs 411 are
interconnected via a system bus 412 to a random access memory (RAM)
414, read-only memory (ROM) 416, input/output (I/O) adapter 418
(for connecting peripheral devices such as disk units 421 and tape
drives 440 to the bus 412), user interface adapter 422 (for
connecting a keyboard 424, mouse 426, speaker 428, microphone 432,
and/or other user interface device to the bus 412), a communication
adapter 434 for connecting the system 400 to a data processing
network, the Internet, an Intranet, a local area network (LAN),
etc., and a display adapter 436 for connecting the bus 412 to a
display device 438 and/or printer 439 (e.g., a digital printer of
the like).
[0139] As will be appreciated by one skilled in the art, aspects of
the present invention may be embodied as a system, method or
computer program product. Accordingly, aspects of the present
invention may take the form of an entirely hardware embodiment, an
entirely software embodiment (including firmware, resident
software, micro-code, etc.) or an embodiment combining software and
hardware aspects that may all generally be referred to herein as a
"circuit," "module" or "system." Furthermore, aspects of the
present invention may take the form of a computer program product
embodied in one or more computer readable medium(s) having computer
readable program code embodied thereon.
[0140] Any combination of one or more computer readable medium(s)
may be utilized. The computer readable medium may be a computer
readable signal medium or a computer readable storage medium. A
computer readable storage medium may be, for example, but not
limited to, an electronic, magnetic, optical, electromagnetic,
infrared, or semiconductor system, apparatus, or device, or any
suitable combination of the foregoing. More specific examples (a
non-exhaustive list) of the computer readable storage medium would
include the following: an electrical connection having one or more
wires, a portable computer diskette, a hard disk, a random access
memory (RAM), a read-only memory (ROM), an erasable programmable
read-only memory (EPROM or Flash memory), an optical fiber, a
portable compact disc read-only memory (CD-ROM), an optical storage
device, a magnetic storage device, or any suitable combination of
the foregoing. In the context of this document, a computer readable
storage medium may be any tangible medium that can contain, or
store a program for use by or in connection with a system,
apparatus, or device running an instruction.
[0141] A computer readable signal medium may include a propagated
data signal with computer readable program code embodied therein,
for example, in baseband or as part of a carrier wave. Such a
propagated signal may take any of a variety of forms, including,
but not limited to, electro-magnetic, optical, or any suitable
combination thereof. A computer readable signal medium may be any
computer readable medium that is not a computer readable storage
medium and that can communicate, propagate, or transport a program
for use by or in connection with a system, apparatus, or device
running an instruction.
[0142] Program code embodied on a computer readable medium may be
transmitted using any appropriate medium, including but not limited
to wireless, wireline, optical fiber cable, RF, etc. or any
suitable combination of the foregoing.
[0143] Computer program code for carrying out operations for
aspects of the present invention may be written in any combination
of one or more programming languages, including an object oriented
programming language such as Java, Smalltalk, C++ or the like and
conventional procedural programming languages, such as the "C"
programming language or similar programming languages. The program
code may run entirely on the user's computer, partly on the user's
computer, as a stand-alone software package, partly on the user's
computer and partly on a remote computer or entirely on the remote
computer or server. In the latter scenario, the remote computer may
be connected to the user's computer through any type of network,
including a local area network (LAN) or a wide area network (WAN),
or the connection may be made to an external computer (for example,
through the Internet using an Internet Service Provider).
[0144] Aspects of the present invention are described below with
reference to flowchart illustrations and/or block diagrams of
methods, apparatus (systems) and computer program products
according to embodiments of the invention. It will be understood
that each block of the flowchart illustrations and/or block
diagrams, and combinations of blocks in the flowchart illustrations
and/or block diagrams, can be implemented by computer program
instructions. These computer program instructions may be provided
to a processor of a general purpose computer, special purpose
computer, or other programmable data processing apparatus to
produce a machine, such that the instructions, which run via the
processor of the computer or other programmable data processing
apparatus, create means for implementing the functions/acts
specified in the flowchart and/or block diagram block or blocks.
These computer program instructions may also be stored in a
computer readable medium that can direct a computer, other
programmable data processing apparatus, or other devices to
function in a particular manner, such that the instructions stored
in the computer readable medium produce an article of manufacture
including instructions which implement the function/act specified
in the flowchart and/or block diagram block or blocks.
[0145] The computer program instructions may also be loaded onto a
computer, other programmable data processing apparatus, or other
devices to cause a series of operational steps to be performed on
the computer, other programmable apparatus or other devices to
produce a computer implemented process such that the instructions
which run on the computer or other programmable apparatus provide
processes for implementing the functions/acts specified in the
flowchart and/or block diagram block or blocks.
[0146] The flowchart and block diagrams in the Figures illustrate
the architecture, functionality, and operation of possible
implementations of systems, methods and computer program products
according to various embodiments of the present invention. In this
regard, each block in the flowchart or block diagrams may represent
a module, segment, or portion of code, which comprises one or more
operable instructions for implementing the specified logical
function(s). It should also be noted that, in some alternative
implementations, the functions noted in the block may occur out of
the order noted in the figures. For example, two blocks shown in
succession may, in fact, be run substantially concurrently, or the
blocks may sometimes be run in the reverse order, depending upon
the functionality involved. It will also be noted that each block
of the block diagrams and/or flowchart illustration, and
combinations of blocks in the block diagrams and/or flowchart
illustration, can be implemented by special purpose hardware-based
systems that perform the specified functions or acts, or
combinations of special purpose hardware and computer
instructions.
[0147] The embodiments described above are illustrative examples
and it should not be construed that the present invention is
limited to these particular embodiments. Thus, various changes and
modifications may be effected by one skilled in the art without
departing from the spirit or scope of the invention as defined in
the appended claims.
* * * * *
References