U.S. patent application number 15/642361 was filed with the patent office on 2017-10-19 for method for matching queries with answer items in a knowledge base.
The applicant listed for this patent is NANOREP TECHNOLOGIES LTD.. Invention is credited to Omer Ben Nahum, Amit Ben Shahar.
Application Number | 20170300562 15/642361 |
Document ID | / |
Family ID | 51655228 |
Filed Date | 2017-10-19 |
United States Patent
Application |
20170300562 |
Kind Code |
A1 |
Ben Shahar; Amit ; et
al. |
October 19, 2017 |
METHOD FOR MATCHING QUERIES WITH ANSWER ITEMS IN A KNOWLEDGE
BASE
Abstract
The present invention includes an expert system in which a
search index furnishes answers to incoming queries provided in
natural language. A search index for a specific field contains
components that facilitate selecting a best fitting stored answer
to the incoming query. Furthermore, context of the incoming query
(e.g. location of the user, a current web page or service being
used/viewed by the user, the time, etc.) may be considered when
selecting a best fitting answer.
Inventors: |
Ben Shahar; Amit; (Petah
Tikva, IL) ; Ben Nahum; Omer; (Maccabim, IL) |
|
Applicant: |
Name |
City |
State |
Country |
Type |
NANOREP TECHNOLOGIES LTD. |
Herzliya Pituah |
|
IL |
|
|
Family ID: |
51655228 |
Appl. No.: |
15/642361 |
Filed: |
July 6, 2017 |
Related U.S. Patent Documents
|
|
|
|
|
|
Application
Number |
Filing Date |
Patent Number |
|
|
15438777 |
Feb 22, 2017 |
|
|
|
15642361 |
|
|
|
|
14311441 |
Jun 23, 2014 |
9639602 |
|
|
15438777 |
|
|
|
|
13757940 |
Feb 4, 2013 |
9110978 |
|
|
14311441 |
|
|
|
|
13019318 |
Feb 2, 2011 |
8407208 |
|
|
13757940 |
|
|
|
|
Current U.S.
Class: |
1/1 |
Current CPC
Class: |
G06F 16/3344 20190101;
G06F 16/319 20190101 |
International
Class: |
G06F 17/30 20060101
G06F017/30 |
Claims
1. A system for matching stored queries (matchSQs) to a user
natural language query (NLQ), said system comprising: a computing
platform including communication circuitry, processing circuitry
and computer executable code adapted to cause the computing
platform to: (a) receive digital data representing the user NLQ;
(b) search at least one knowledgebase for candidate matchSQs, which
candidate matchSQs include words corresponding to words in the user
NLQ; (c) score each match between the NLQ and each individual
candidate matchSQs by: a. calculating a first Similar Word Score
(SWS) between the NLQ and the individual candidate matchSQ by using
a defined formula to combine or multiply significance values of
words in the individual candidate matchSQ corresponding to words in
the NLQ; b. calculating a second SWS between the NLQ and the NLQ by
using the defined formula to combine or multiply significance
values of all words in the NLQ; c. calculating a third SWS between
the individual candidate matchSQ and the individual candidate
matchSQ by using the defined formula to combine or multiply
significance values of all words in the individual candidate
matchSQ; d. calculating a match score between the NLQ and the
individual candidate matchSQ by aggregating: (1) a difference
between the first SWS and the second SWS and (2) a difference
between the first SWS and the third SWS.
2. The system according to claim 1, further comprising
disqualifying matchSQs based on a comparison of context of the
candidate matchSQs to a context of the NLQ.
3. The system according to claim 1, wherein calculating a match
score between the NLQ and each individual candidate matchSQ further
includes factoring a similarity or dissimilarity in the context of
the NLQ and contexts of the candidate matchSQs.
4. The system according to claim 1, wherein the significance values
of words represent a frequency of use within one or more texts or
knowledgebases.
5. The system according to claim 1, further comprising calculating
a query specific significance value for words comprising the NLQ,
the query specific significance value of each given word
representing a ratio between a significance value of the word and a
sum of significance values of words comprising the NLQ.
6. The system according to claim 1, further comprising calculating
a query specific significance value for words comprising each
individual candidate matchSQ, the query specific significance value
of each given word representing a ratio between a significance
value of the word and a sum of significance values of words
comprising the individual candidate matchSQ.
7. A system for providing an automated response to a user natural
language query (NLQ) made in regard to a subject, said system
comprising: a computing platform including processing circuitry
associated with a tangible digital medium containing computer
executable code adapted to cause a processor to: (a) receive
digital data representing the user NLQ; (b) search at least one
knowledgebase for candidate matchSQs, which candidate matchSQs
include words corresponding to words in the user NLQ; (c) score
each match between the NLQ and each individual candidate matchSQs
by: i. calculating a first Similar Word Score (SWS) between the NLQ
and the individual candidate matchSQ by using a defined formula to
combine or multiply significance values of words in the individual
candidate matchSQ corresponding to words in the NLQ; ii.
calculating a second SWS between the NLQ and the NLQ by using the
defined formula to combine or multiply significance values of all
words in the NLQ; iii. calculating a third SWS between the
individual candidate matchSQ and the individual candidate matchSQ
by using the defined formula to combine or multiply significance
values of all words in the individual candidate matchSQ; iv.
calculating a match score between the NLQ and the individual
candidate matchSQ by aggregating: (1) a difference between the
first SWS and the second SWS and (2) a difference between the first
SWS and the third SWS; and communication circuitry adapted to send
the computer executable code to the processor.
8. The system according to claim 7, further comprising
disqualifying matchSQs based on a comparison of context of the
candidate matchSQs to a context of the NLQ.
9. The system according to claim 7, wherein calculating a match
score between the NLQ and each individual candidate matchSQ further
includes factoring a similarity or dissimilarity in the context of
the NLQ and contexts of the candidate matchSQs.
10. The system according to claim 7, wherein the significance
values of words represent a frequency of use within one or more
texts or knowledgebases.
11. The system according to claim 7, further comprising calculating
a query specific significance value for words comprising the NLQ,
the query specific significance value of each given word
representing a ratio between a significance value of the word and a
sum of significance values of words comprising the NLQ.
12. The system according to claim 7, further comprising calculating
a query specific significance value for words comprising each
individual candidate matchSQ, the query specific significance value
of each given word representing a ratio between a significance
value of the word and a sum of significance values of words
comprising the individual candidate matchSQ.
Description
PRIORITY CLAIMS
[0001] This application is a continuation of: U.S. patent
application Ser. No. 15/438,777, filed on Feb. 22, 2017 by the
inventors of the present application and titled: "Method for
Matching Queries with Answer Items in a Knowledge Base";
U.S. application Ser. No. 15/438,777 is a continuation of: U.S.
patent application Ser. No. 14/311,441, filed on Jun. 23, 2014 by
the inventors of the present application and titled: "Method for
Matching Queries with Answer Items in a Knowledge Base" and issued
as U.S. Pat. No. 9,639,602 on May 2, 2017; U.S. application Ser.
No. 14/311,441 is a continuation in part of: U.S. patent
application Ser. No. 13/757,940, filed on Feb. 4, 2013 by the
inventors of the present application and titled: "Method for
Matching Queries with Answer Items in a Knowledge Base" and issued
as U.S. Pat. No. 9,110,978 on Aug. 18, 2015; U.S. patent
application Ser. No. 13/757,940, is a continuation application of
U.S. patent application Ser. No. 13/019,318, filed on Feb. 2, 2011
by the inventors of the present application and titled: "Method for
Matching Queries with Answer Items in a Knowledge Base" and issued
as U.S. Pat. No. 8,407,208 on Mar. 26, 2013; each of the
aforementioned applications is hereby incorporated herein by
reference in its entirety.
FIELD OF THE INVENTION
[0002] The present invention relates to an expert system,
functional for providing answers to users presenting queries in
specified fields.
BACKGROUND OF THE INVENTION
[0003] Expert systems attempt to provide answers to problems in a
specific domain, with the intent of at least partially replacing
human experts. Expert system reasoning often tries to imitate human
reasoning. Since expert systems are to do with human reasoning and
communications, understanding and utilizing human words is a
crucial factor if a natural language is used for querying the
expert system. The two most significant components of an expert
system are the expert system shell and the knowledge base. The
expert system shell handles interpretation of user input into the
system in order to facilitate the reasoning. The knowledge base of
an expert system is the factual body consisting of information and
relationships between the elements of information arranged in such
a way as to fulfill the purpose of the expert system. The knowledge
base does not necessarily contain definitive rules to prioritize
between options or rule out others altogether.
[0004] The Internet offers a practical medium for users to
interconnect with an expert system. The convenience and
availability of Internet browsers enables most users to benefit
from the knowledge managed by specific expert systems almost
without limitations of location. Moreover, the ubiquitous access
users may now have to knowledge bases through personal hand held
communications end points, encourages the need for the
diversification in knowledge domains, so that one could derive
quick answers for an increasing amount of fields of knowledge.
SUMMARY OF THE INVENTION
[0005] The present invention includes an expert system in which a
search index furnishes answers to incoming queries provided in
natural language. A search index for a specific field contains
components that facilitate selecting a best fitting stored answer
to the incoming query. Furthermore, context of the incoming query
(e.g. location of the user, a current web page or service being
used/viewed by the user, the time, etc.) may be considered when
selecting a best fitting answer. A language specific storehouse of
weighted words and a private storehouse of weighted words
associated with a field-specific search index provide the basis for
evaluating the significance level of a natural language word of a
query. Again, context of the incoming query may be considered when
evaluating the significance level of a natural language word of a
query. Irrelevant portions of an incoming query may first be
deleted from the inquiry prior to processing. A procedure elects
candidates from a store of indexed answers to match the incoming
query to first form a list of candidates, based on the existence of
identical or similar words. Then, from the list of available
candidates, one that provides the best match is selected. The final
selection takes into account the influence of the excess of less
significant words. A candidate is derogated if it is found to
contain a larger number of less significant words as compared to
another candidate containing a smaller number of insignificant
words.
BRIEF DESCRIPTION OF THE DRAWINGS
[0006] The subject matter regarded as the invention is particularly
pointed out and distinctly claimed in the concluding portion of the
specification. The invention, however, both as to organization and
method of operation, together with objects, features, and
advantages thereof, may best be understood by reference to the
following detailed description of non-limiting exemplary
embodiments thereof when read with the accompanying drawings in
which:
[0007] FIG. 1A is a graphical presentation of an exemplary flow of
words arriving from a user to be subsequently given a significance
level;
[0008] FIG. 1B is a graphical presentation of an exemplary summary
of the contesting routine;
[0009] FIG. 2A is a graphical presentation of an exemplary usage of
a plurality of search indexes;
[0010] FIG. 2B is a graphical presentation of exemplary components
of the indexing component;
[0011] FIGS. 3A-3C are a series of graphical presentations of an
exemplary sequence of events carried out to produce similar word
scores of the comparison types;
[0012] FIG. 4 is a graphical presentation of a situation in which a
number of reference lists are associated with one indexed answer
item.
[0013] FIG. 5A is a flowchart including exemplary steps of
processing a natural language incoming query to select one or more
matching indexed answers, including the removal of irrelevant words
from the query, all in accordance with some embodiments of the
present invention.
[0014] FIG. 5B is a flowchart including exemplary steps of
processing a natural language incoming query to select one or more
matching indexed answers, all in accordance with some embodiments
of the present invention.
[0015] FIG. 6A is a flowchart including exemplary steps of
determining significance values of words of a query, including the
removal of irrelevant words from the query, all in accordance with
some embodiments of the present invention.
[0016] FIG. 6B is a flowchart including exemplary steps of
determining significance values of words of a query, including the
removal of irrelevant words from the query and factoring of
synonyms of each word in the query, all in accordance with some
embodiments of the present invention.
[0017] FIG. 6C is a flowchart including exemplary steps of
determining significance values of words of a query, in accordance
with some embodiments of the present invention.
[0018] FIG. 6D is a flowchart including exemplary steps of
determining significance values of words of a query, including
factoring of synonyms of each word in the query, all in accordance
with some embodiments of the present invention.
[0019] FIG. 7A is a flowchart including exemplary steps of
compiling lists of candidate matches of an IQ, in accordance with
some embodiments of the present invention.
[0020] FIG. 7B is a flowchart including exemplary steps of
compiling lists of candidate matches of an IQ, including factoring
of synonyms of each word in the query, all in accordance with some
embodiments of the present invention.
[0021] It will be appreciated that for simplicity and clarity of
illustration, elements shown in the figures have not necessarily
been drawn to scale. For example, the dimensions of some of the
elements may be exaggerated relative to other elements for clarity.
Further, where considered appropriate, reference numerals may be
repeated among the figures to indicate corresponding or analogous
elements.
[0022] It should be understood that the accompanying drawings are
presented solely to elucidate the following detailed description,
are therefore, exemplary in nature and do not include all the
possible permutations of the present invention.
DETAILED DESCRIPTION OF EMBODIMENTS OF THE INVENTION
[0023] In the following detailed description, numerous specific
details are set forth in order to provide a thorough understanding
of the invention. However, it will be understood by those skilled
in the art that the present invention may be practiced without
these specific details. In other instances, well-known methods,
procedures, components and circuits have not been described in
detail so as not to obscure the present invention.
[0024] Unless specifically stated otherwise, as apparent from the
following discussions, it is appreciated that throughout the
specification discussions utilizing terms such as "processing",
"computing", "calculating", "determining", or the like, refer to
the action and/or processes of a computer or computing system, or
similar electronic computing device, that manipulate and/or
transform data represented as physical, such as electronic,
quantities within the computing system's registers and/or memories
into other data similarly represented as physical quantities within
the computing system's memories, registers or other such
information storage, transmission or display devices. The term
server may refer to a single server or to a functionally associated
cluster of servers.
[0025] Embodiments of the present invention may include apparatuses
for performing the operations herein. This apparatus may be
specially constructed for the desired purposes, or it may comprise
a general purpose computer selectively activated or reconfigured by
a computer program stored in the computer. Such a computer program
may be stored in a computer readable storage medium, such as, but
is not limited to, memory cards (for example SD card), SIM cards,
any type of disk including floppy disks, optical disks, CD-ROMs,
magnetic-optical disks, read-only memories (ROMs), random access
memories (RAMs) electrically programmable read-only memories
(EPROMs), electrically erasable and programmable read only memories
(EEPROMs), magnetic or optical cards, or any other type of media
suitable for storing electronic instructions, and capable of being
coupled to a computer system bus.
[0026] The processes and displays presented herein are not
inherently related to any particular computer, communication device
or other apparatus. Various general purpose systems may be used
with programs in accordance with the teachings herein, or it may
prove convenient to construct a more specialized apparatus to
perform the desired method. The desired structure for a variety of
these systems will appear from the description below. In addition,
embodiments of the present invention are not described with
reference to any particular programming language or markup
language. It will be appreciated that a variety of programming
languages or techniques may be used to implement the teachings of
the inventions as described herein.
[0027] The present invention is a multi-field expert system,
meaning that implementing the invention will facilitate for a user
the extraction of data in a plurality of fields of knowledge. For
each of those fields of knowledge, a knowledge base including a
unique index is created that contains the knowledge of that
specific field, and provides a means to retrieve that knowledge in
response to natural language queries. Furthermore, each unique
index may include indications of context of each content contained
in the knowledge base to be considered when selecting a best
fitting answer to a given query.
[0028] In accordance with the present invention, queries are
presented as electronic signals representing natural language words
to a search index (SI). The queries are further analyzed and
responses provided, based on the existing knowledge base for that
specific SI and, optionally, also based on the context of the
incoming query. It is also possible to direct searches to multiple
fields at a time. An additional feature of the present invention is
that each incoming query (IQ) having been reciprocated by an answer
can and may be used to contribute to the performance of the
knowledge base. This aspect of the invention will be explained
later on. The analytic approach, in accordance with the present
invention, attempts to analyze an incoming query and find an
indexed answer item (IAI) that the expert system would be able to
further use in an optimal way to provide the best answers.
[0029] An IQ undergoes a process which, for reasons of convenience
of explanation, will be considered hereinafter as consisting of
several procedures, some of which are executed in an ordered
succession. The first procedure is a pre-processing procedure
(PPP), in which an incoming query is evaluated for some
administrative properties, for example language, general field of
commerce, technology, context etc, and for misspelling and
validity. Additionally, in the PPP, a private storehouse of
weighted words is assigned to the IQ. Further, irrelevant portions
of the query may be removed prior to further processing, i.e. words
irrelevant to the substance of the query (for example, superfluous
adjectives, articles (e.g. "a", "the", etc.) and other parts of
speech that do not add specific meaning to a sentence). According
to further embodiments, irrelevant portions of a query may also be
determined based on context (for example, the terms "know what I
mean" may be irrelevant when appearing together, whereas each term
may have relevance when appearing separately). Some of the
consequences of the PPP will be dealt with below.
[0030] Subsequently, in the analytic procedure performed on the IQ,
a significance level score is to be given to the fragments of the
IQ, this will be explained with reference to FIG. 1A, showing a
symbolic description of the way in which it is accomplished.
Incoming query 20 arrives, for example, by way of a communications
network running a TCP/IP communications protocol and is received by
a server connected to the network. Such a server invokes a program,
extracting information from the natural language (NL) query
arrived, which includes the words and possibly additional
information as to which of the one or more SIs are to be searched.
Assuming, by way of example, that the query is a sentence composed
of a succession of words making sense to a person versed in a
specific language and having practical knowledge in a field about
which the knowledge base is chartered to provide a service, the IQ
is first fragmented and in a typical case, information extracted.
The fragments form a succession of natural language words (NLWs)
lined up in a queue 22 to be submitted each to a contesting routine
24, in which each word is weighed by finding a match in a natural
language word storehouse 26 of weighted words. In the storehouse of
words, each entry, such as a word symbolized by NLW2, has an
associated numerical weight. Typically, the NLWs are processed in
the order in which they arrive. The role of contesting routine 24
is to determine the significance level (SL) score associated with
each incoming word, designated NLWT1, NLWT2 etc. The temporal
succession of words of original IQ 20 is then reordered to give a
new succession 28 referred to hereinafter generally as ordered
query word succession (OQWS), in which the NLW with the highest
significance level becomes associated with the highest position in
the new succession, as described graphically, a certain NLW of the
IQ becomes associated with SL score A, and a certain, other NLW
becomes associated with a SL score "least". Other NLWs occupy
intermediate positions in the OQWS as determined by their level of
significance (SL). It cannot be ruled out, statistically, that on
occasions the new order is the same as the order of arrival of
NLWs. The storehouses of weighted words are dynamic entities which
are constantly recompiled and their maintenance will be discussed
later on. The assignment of a significance level to each of the
incoming NLWs marks the completion of the contesting routine 24 as
relates to query 20. As an example, in the question: "where can I
find a library?" the word "I" is almost insignificant, the same can
be said about "can". The word "find" is however more indicative
because it relates to a fact that the person asking is searching
for something which has practical meaning. The most significant
word would be in this case "library". However, in other examples,
the same word may receive different weights. Context of the
incoming query may also be considered when assigning SL scores (For
example, in a query relating to geographical/locational question,
words representing physical locations may be assigned higher SL
scores). Contesting routine 24 which determines the SL score of
each NLW (in this case from the IQ), operates as follows:
[0031] Routine 24 references both NL storehouse of weighted words
26 which is private in the context of the current SI, and
storehouse 30 which is common to all SIs for the same language
(i.e. English). Each NLW storehouse (the SI private storehouse 26,
and the language storehouse 30) maintains a respective SL score for
each contained NLW. These scores are determined by distributing the
NLWs by their respective usage frequency on a logarithmic scale,
such that the most common word has the lowest SL score, and the
least common NLW gets the highest score. According to some
embodiments, each NLW storehouse (the SI private storehouse 26, and
the language storehouse 30) may maintain multiple SL scores for
each contained NLW, dependent on the context of the query. For
example, the word party may be uncommonly used between the hours of
8-5 and very frequent between the hours of 5-10 pm, or the word
team may be scarcely used by viewers of fashion related webpages
and commonly used by viewers of sports websites. In other words,
the SL scoring process may be dynamic and dependent on the context
of the incoming query. Typically resulting scores are normalized to
a scale of 0 to 100. In the routine, each NLWT obtains a SL score
by weighing both the SL score component of the private storehouse
26 and the SL score component obtained from the language storehouse
30, such that the significance given to each word in an IQ is a
value reflecting its combined relevance both with respect to the
present SI and field of knowledge and with respect to the language
generally. To demonstrate the difference in the two aspects, the
word `shoe` can be considered indicative of context in the scope of
the English language, but is very common and much less significant
in the scope of a footwear related knowledge-base.
[0032] According to further embodiments, synonyms may be considered
when determining SL scores. For this purpose, each NLW may be
automatically correlated to multiple NLWs having the same meaning
prior to assigning a SL score to the NLW. For example, the words
blouse and shirt may be considered equivalent when assigning SL
scores to either word. In this manner the word blouse may receive a
relatively low SL score even though it is uncommonly used, due to
the fact that the word shirt is very commonly used. According to
yet further embodiments, a degree of synonymity/similarity between
the synonyms may be factored when determining the SL score for one
of the synonyms and/or a commonality of the use of a certain
synonym in place of its synonymous pair may be factored into the
degree of synonymity. Further, in the later process of comparing
IQs to indexed answer items matches of NLWs to synonymous NLWs may
incur a match penalty which may be relative to the degree of
similarity between the significances of the pair of synonyms.
Accordingly, each recorded synonymous word pair may further include
a value representing the similarity between the words of the pair.
This value may be factored when determining an SL score for one of
the words of the pair as well as during matching of NLWs in an IQ
to indexed answer items. According to further embodiments, a value
representing the similarity between the words of a pair of synonyms
may also be context dependent, as some pairs of synonyms are more
or less similar in significance in different contexts.
[0033] Finally, after each word is assigned its SL score, the words
are reordered by their SL score and the SL scores are typically
normalized. In a preferred embodiment of the invention, the scores
are normalized, starting with the most significant NLW, which is
typically set to a 100 (or 1, depending on the scale). Thus the
second NLW is assigned, for example, to 75. The ratio between the
references is typically kept the same after normalization. For
example if NLWA was 50 before normalization, after normalization it
becomes 100, and the next NLW, SLWB was 25, is normalized to 50,
keeping the same ratio with the first NLW normalization factor. To
summarize the issue of associating each word of queue 22 with a SL
score, reference is made to FIG. 1B. NL word NLWT1 is contested
with private storehouse of words 26, receiving a SL score component
32. Concomitantly, language storehouse of words 30 is contested
with NLWT1, producing a respective SL score component 34. Both SL
score components are then combined to produce a SL score which is
now associated with NLWT1. The NLWs of a certain IQ are then
arranged in a OQWS 28, in the order of their respective SL scores.
One example of attaining such a combination is by performing
weighted averaging, while giving one factor a different weight than
the other. Finally their SL score is normalized, in the OQWS.
The Search Index (SI) Component, and the Searching and Finding Best
Match Procedure (SFBMP): Traversing IAIs
[0034] The next procedure, which follows in the analytical
sequence, is the selecting of the best matching indexed answer item
for the IQ out of a selected few. Before explaining the procedure,
the index component of a system of the invention is introduced.
First, reference is made to FIG. 2A. For each private storehouse of
weighted words a certain SI 36 exists which relates to this private
storehouse. Thus SI 36A relates to private storehouse of words 26A
of a certain field of knowledge, and SI 36B relates to private
storehouse of words 26B of another field of knowledge. Again,
according to some embodiments, each SI may be dynamic and dependent
on the context of an incoming query. Thus, SI 36A may be slightly
different in morning hours as opposed to evening hours or dependent
on the type of webpage the user was viewing when entering the query
or dependent on the current location of the user, etc.
[0035] As explained above, private storehouse of weighted words 26
contains a list of words each associated with a weight which
reflects its potential significance (possibly dependent on the
context of the query). More is described about the structure of SI
with reference to FIG. 2B. SI 36 includes a list of indexed
keywords 38, all of which keywords bear equivalence to a respective
word in a private word storehouse 26, thus IKW1, an indexed
key-word, is equivalent to NLW1 in the private storehouse 26 and
also equivalent to a NLW in the language storehouse 30, for example
NLW.alpha.. Each indexed keyword is associated with a list of
Indexed Keyword References. For example list of reference (LOR) 40
is associated with indexed word IKW1, and list of references (LOR)
42 associated with indexed word IKW2. An indexed keyword reference
(IKW REF) represents an association of an IKW and an IAI with a SL
score, typically calculated by contesting routine 24. A LOR may
sometimes contain only one reference. In each LOR, the order of
references is the order which quantitatively reflects the SL score
of the association between the IKW and the IAI, presented as a
numeric value, meaning the IKW REF that bears the strongest
association with an IAI will be positioned first in the LOR. For
example, LOR 42, contains several references, all associated with
indexed key word IKW2, and the first reference, designated 46 bears
the highest SL score of all IKW References in LOR 42. The term IKW
reference, stems from the fact that each IKW reference (IKW REF) in
index 36 refers to only one specific indexed answer item (IAI)
(one-to-one relationship). The mutual relationship is not
symmetrical, and each IAI can refer or receive reference to and
from a number of IKWs (one to many relationship), the only
restriction is that an IAI can refer to only one IKW REF in a
specific LOR. Each IAI points to an Internal List of References
(ILOR), containing all the references pointing to it, possibly from
a plurality of LORs. For example IAI2 designated item 68, contains
ILOR 78 that refers to two IKWs, i.e. IKWN REF1 designated item 70
of LOR 44 and to IKW2 REF2 designated item 72 of LOR 42.
Searching and Finding a Best Fitting Indexed Answer Item for an IQ
(SFBMP)--Making a Collection of IAIs, a List of Candidates
[0036] Once a new query is presented to the system, contesting
routine 24 constructs a OQWS, containing a weighted list of natural
words. Then, the SFBMP is invoked. For each NLW of the IQ, a search
is conducted finding the appropriate IKW in the appropriate SI.
NLWA is the first word in the OQWS, having a highest SL score. It
is associated with NLW1 in the private storehouse of weighted words
26 (and NLW.alpha. of the language storehouse 30), and with an
indexed keyword IKW 1. The SFBMP searches, proceeding to traverse
each of the LORs referenced by the IKWs referenced by the search
query, and for each of the IKW REFs in those LORs, it will in turn
traverse the ILOR contained in each IAI referenced by a traversed
IKW REF.
[0037] There are two traversal strategies available for determining
the order of traversal of each LOR, the sequential strategy, in
which traversing the REFs is in order of SL scores (LORs are sorted
by SL score), and another strategy, a "Similarity First" search
strategy, the traversal of the LOR starts from the IKW REF having a
SL score most similar to the weight in the NLW for the IQ, and
subsequently continues to the least similar one (possibly favoring
higher positive or negative differences between the SL similarity
making the search expansion asymmetrical). In addition, in a
preferred embodiment of the invention, both strategies may only
traverse a portion of the LOR, the portion size may be in
proportion to the SL for the contested NLW in the IQ, bringing
about an effective mechanism to limit traversal for low SL score
LORs, which are by definition longer and generally contain less
relevant IAI. For example: For NLWA with SL=100, 100% of the LOR's
references will be traversed, and for NLWB with SL=50, only 50% of
the LOR's references will be traversed.
[0038] To demonstrate a sequential search procedure, the following
example is given: Looking first in LOR 40 associated with IKW1, to
find its highest SL scoring IKW1 REF, i.e. IKW1 REF 1. This
demonstrates the fact that the order of NLWs in the OQWS dictates
the order of search inside the SI. The higher the SL scoring of a
NLW in the OQWS is, the earlier it will be visited. In another
example (also of a sequential strategy): NLWn is a NLW in the
private storehouse of weighted words 26, equivalent to IKWN in the
index. Associated with IKWN is LOR 44, and the highest SL scoring
IKWN reference in LOR 44, IKWN REF1 is automatically selected as a
first reference to be traversed in the search. This IKWN REF1
designated 70, points at IAI2, designated 68, at this point the
procedure will initiate traversal of IAI2's ILOR 78 until
completing all items referenced in the IAI's ILOR (in this example
the next item to be traversed will be IKW2 REF2 designated 72), and
then continue the traversal of the IKWN's LOR (in this example LOR
44) and the next item to be traversed will be IKWN REF2. When
finishing one NLW, the procedure continues to the next NLW, i.e.
NLWB in OQWS 28, etc.
[0039] During the search procedure, if an IKW points at an IAI
which was already traversed in the current SFBMP (typically by a
reference from a different LOR) it is not processed again. As the
procedure traverses index 36, for each traversed IAI, a match level
score (MLS) is calculated that is indicative of the perceived
similarity between each traversed IAI and the IQ. During the
traversal, a collection of relevant IAIs is accumulated, referred
to hereinafter elected IAIs. In order to obtain from a list of
elected IAIs the best matching IAI for the IQ, the calculated match
level score MLS will ultimately be used to determine the best
match, in response to the query.
SFBMP: Calculating the Match Level Score (MLS)
[0040] MLS calculation is accomplished in two steps. The first step
calculates a Similar Words Score (SWS) by matching two sets of NLWs
having each a respective SL score per contained NLW, for example a
set of NLWs in the OQWS and an indexed set in the elected IAI. To
calculate the SWS, for each NLW in the OQWS, a matching IKW REF is
searched for in the tested IAI's ILOR, and if a match is found, a
score is calculated. The meaning of a match in this case is either
an identical word, or an equivalent word or expression (such as a
synonym). The score is accomplished by combining the SL score of
the IKW REF with the SL score of the NLW in both the OQWS and the
respectively equivalent IKW REF, for example by multiplying the
respective SL scores, and optionally applying a penalty or bonus
for a high difference or high similarity between the two SL scores,
respectively. The combined scores from all NLWs and equivalent
IKWs, respectively, are accumulated to create the SWS. An example
is given below.
[0041] In addition to the MLS calculation accomplished by comparing
the elected IAI and the OQWS and obtaining the pertinent SWS score,
another type of SWS score is obtained. This other type is obtained
by computing the self matching for each elected IAI, or in other
words using own components for each respective elected IAI only and
in a similar way by self matching regarding the OQWS. These two SWS
sub-types are referred to hereinafter in general as
Perfect-Match-Scores (PMSs). As can be seen in FIG. 3A, SWSs 94, 96
and 98 are generated in a procedure that exploits the difference
between a OQWS 100 of an IQ, and the elected IAIs obtained during
the run of the SFBMP procedure, using both the OQWS and the
respective IAIs, IAI 102, IAI 104 and IAI 106. The two sub-types of
PMSs typically generate the highest possible SWS for each
respective NLW set, accordingly. PMS 110 results from self
comparing of the OQWS words and PMS 112, 114 and 116 results from
the self comparing of IAI 102 to IAI 106, respectively. In FIG. 3B,
the computation of SWSs calculated for the comparison of an elected
IAI with the OQWS, and referred to hereinafter as CSWS, is
pictorially explained. OQWS 212 contains a list of NLWs arranged in
the order of the SL scores. Several IAIs found as described above
due to their containment of indexed natural language words (INLWs),
IAI 1, IAI2 and IAI3, all contribute, in an order of their SL
scores, combining to form SWS. Thus, the respective SL scores of
NLW 214 of OQWS 212 are combined with INLW 214A in accordance with
a calculation rule 218. A combination computation score (CCS) 220
is obtained. Then, NLW 224 is found no match in IAI1, but NLW 226
combined as before (same rule 218) with INLW 226A of IAI 1, to
yield CCS 230, the order of computation dictated by a traversal
strategy discussed above. The CCSs are summed up to produce a CSWS,
and the number of CCSs produced or used for calculating the CSWS is
predetermined or dynamically determined or not limited at all.
After the traversal of IAI1 is completed, or in parallel, as in
FIG. 3C the INLWs of IAI2 are traversed and combined with
equivalents, if available, in an order dictated by the according to
traversal strategy. CSWS2 is produced by summing up the CCSs.
[0042] Finally, in the second step, the MLS calculation includes
the combining of the difference between (a) the SWS, and (b) each
of the two PMSs generated as described above. The combining is not
necessarily balanced, meaning, for example, that the difference
between the CSWS and the PMS may receive a higher factor in the
calculation of the final MLS and vice versa.
[0043] For example: [0044] 1. PMS for a OQWS with 2 NLWs having SL
scores 100 and 50 respectively is calculated as follows (such a
typical calculation multiplying each score by itself as if a match
was found):
[0044] (100*100+50*50)=12,500 [0045] 2. PMS for IAI with 2 NLWs
with SL scores 100 and 70 respectively is:
[0045] (100*100+70*70)=14,900 [0046] 3. CSWS (assuming that only
the 1 NLW appears in both items, and that NLW is the one that
scored 100 in both sets of keywords, each of the SWS), is
calculated as follows:
[0046] (100*100)=10,000. [0047] 4. The difference between the CSWS
and the OQWS PMS is:
[0047] (12,500*10,000)=2,500, [0048] and the difference with the
IAI PMS is:
[0048] (14,900-10,000)=4,900.
[0049] In this example the IAI PMS difference factor is 0.3,
smaller than unity, resulting in a final MLS score of:
(2,500*1+4,900*0.3)=3,970.
[0050] Using the CSWS and the PMS results is a good estimation of
similarity between the matched items. However, The reason that a
simple SWS matching does not provide a good enough indication of
similarity between an IQ and an IAI lies in the nature of NL
queries, which dictates that any added NL word to a query may alter
its meaning, and thus should yield different results.
[0051] For example: Assuming a SI containing 2 IAIs: "Definition of
Algorithm", as IAI1 and "What is the best known algorithm for image
processing", as IAI2. Matching both IAIs with the IQ "What is an
algorithm", a person can easily deduce that the better match is
IAI1: "Definition of Algorithm". Using only the SWS for comparison,
will generate a higher score for IAI2 because it contains all of
the NL words from the Query, while IAI1 includes only 1 matching
word, and thus using only the SWSs will result in a bad match.
[0052] However, a quick look at the difference between the PMS and
the CSWS will show that due to the excess of words in IAI2 with
respect to the number of words in IQ1, the PMS will be much higher
and thus result in a higher difference, while at the same time, the
PMS for the IQ will be lower and much closer to the PMS of IAI1 and
will result in a very small difference--prioritizing it as the
better match.
[0053] Using difference based scoring allows detecting the
difference of meaning between long queries to short items in the
index SI and vice-versa, effectively and quantitatively describing
the difference between general and specific queries, and general
and specific Knowledge-base entries (IAIs in the present
disclosure). The apriori removal of irrelevant portions of a query,
as described above, may help to prevent distortions this process
may cause when a query includes additional NLWs unrelated to the
query. For example, the query "how in god's name can I get a pizza
delivered to this location?" may be converted to "how can I get a
pizza delivered to this location?", prior to processing, such that
the appearance of the words "in god's name" in the query and not in
the IAI will not reduce the match level score.
[0054] Typically, PMS scores for IAIs can be stored in the index to
avoid re-calculation during each comparison.
[0055] According to further embodiments, a MLS may be influenced by
similarities and dissimilarities in contexts of the relevant IQ and
IAI, such that similarities in contexts may result in an increase
in the MLS (a bonus) and dissimilarities in contexts may result in
a decrease in the MLS (a penalty). According to further
embodiments, different types of contexts may result in different
bonuses/penalties. For example, similarity or dissimilarity in a
location context (e.g. US) may be more significant (and therefore
applied more weight) than similarity or dissimilarity in a gender
context (e.g. male).
Limiting Index Traversal
[0056] As the MLS for each traversed IAI is calculated during the
traversal, and considering that traversal of the index generally
starts from highest SL scoring NLW in the OQWS and either highest
or best matching SL in the respective IKW LOR (depending on the
used search strategy), it can be assumed that traversal follows the
order of probability for finding good matches, meaning that the
longer traversal continues, the less probable it is to find a
better matching IAI than was already traversed. For that reason, in
another aspect of the invention, a limitation of traversal is
applied, that aims to stop the traversal of IKW LOR when the
perceived resulting MLS degradation rises beyond a certain
threshold to shorten the search process and decrease resource
consumption. The traversal limitation is achieved typically by
using the best MLS calculated in the course of a current traversal
or current search operation (that may include the traversal of a
plurality of LORs), and monitoring for degradation beyond a
specific level, for example, when the calculated MLS for a
traversed TAT is at least 2 times higher from the lowest (and best)
MLS score observed during current LOR traversal.
Plurality of Reference Lists for a Single IAI
[0057] In the example of selecting a best matching indexed answer
item for an IQ described above, each IAI had one single list of
references associated. Each successive reference in that list has a
decreased or equal SL with respect to the preceding one. In yet
another aspect of the invention, a IAI may be associated with a
plurality of reference lists. To explain the consequences of such a
possibility reference is made to FIG. 4. IAIa designated 292, is
associated with several different reference lists, such as
reference lists 294 and 296 respectively. Thus, for a specific
reference list, different PMS and SWS scores may be obtained,
indicating different best matches for different IQs.
[0058] The reason for the existence of multiple reference lists is
associated with several aspects of linguistic diversity, for
example, usage of similar but different words to indicate a
specific entity, spelling mistakes, or different variants of a same
language. Comparison between quality of total scores associated
with different reference list can be used to correct spelling
mistakes, or unify spelling in the pre-processing procedure
mentioned above.
Context Based Filtering
[0059] According to embodiments, context of an IQ and/or indexed
answers items (IAIs) may be factored when performing the analysis
and calculations described herein. According to some embodiments,
some IAIs may be associated with a specific context (e.g. a
specific geographical location such as US). In such embodiments,
the process of matching IQs to IAIs may include factoring a
similarity or dissimilarity in contexts of the IAIs in relation to
the examined IQ, such that similarity in context between an IQ to
an IAI may result in a bonus to the match score whereas a
dissimilarity may result in a penalty/reduction in the match score.
According to further embodiments, dissimilarity in context, in some
case, may result in immediate disqualification of the IAI as a
response to the IQ. For example, if an IQ is associated with the
context "US", all IAIs having a geographical context other than US
(e.g. Europe) may immediately be disqualified.
[0060] According to some embodiments, a context of an IQ or IAI may
include: [0061] a. a user's activity on a website and/or a
topic/context of one or more pages of a website the user has
visited, including, but not limited to, pages visited, products
viewed, products bought, services bought or investigated, options
selected, login information, interactions on the site, and so on;
[0062] b. data relating to the user, including, but not limited to,
geographic location (country, city, etc.) possibly determined based
on IP analysis or other, IP address, user type (paying, free, VIP,
premium, etc.), personal information and demographics (e.g.:
gender, age, ethnicity, income, etc.), and so on; [0063] c. data
relating to the device the user is using, including, but not
limited to, device type (mobile phone, tablet, desktop), device
properties (screen size, hardware capabilities, installed software,
device settings), device manufacturer and price, and so on; [0064]
d. Data relating to the current web page the IQ was entered in,
such as, but not limited to, HTML content, textual content, script
content, URL arguments, images, hidden content, forms, and so on;
and e. Contextual data derived from the language of the IQ, such as
the choice of vocabulary and grammatical nuances.
[0065] According to some embodiments, determining a context of an
IQ and/or IAI for calculation of match scores and/or SL scores, may
include an analysis of the above mentioned parameters, which may
include comparing data from multiple sources (e.g. based on IP
address and user input), weighting different types of data,
factoring the nature of the IQ, use of context models, and/or any
other known form of contextual analysis. According to further
embodiments, such an analysis may further include extraction of
data from external sources based on collected contextual data (e.g.
extracting ethnicity based on a surname from a name database,
extracting economic status based on IP address from an appropriate
database, etc.).
[0066] According to some embodiments, an IQ and/or IAI may be
associated with multiple contexts (e.g. geo context, age context
and gender context). In such embodiments, the system may factor
more than one or all of the contexts when determining SL scores and
match scores. Further, in embodiments where dissimilarity in
context results in disqualification of IAIs, a dissimilarity in one
of the contexts may be sufficient to disqualify an IAI. Yet
further, different types of contexts may be treated differently,
such that dissimilarity in one may result in disqualification
whereas dissimilarity in another may only result in a penalty. For
example, in a case where an IQ has contexts of US and female, an
IAI with contexts of Europe and female may be disqualified whereas
an IAI with a context of US and male may only incur a penalty.
[0067] According to some embodiments, IAIs may have multiple
versions for different contexts (e.g. different version for Europe
and the US or different versions for users sending their IQ from
different websites, etc.). In such embodiments, there may also be a
generic IAI for IQs having contexts not matching any of the context
specific versions.
[0068] According to some embodiments, contexts of IQs may be
divided into "Mandatory" contexts and "Optional" contexts. In such
embodiments, a "Mandatory" context may indicate that IAIs lacking
the context of the IQ (with or without a conflicting context)--are
not returned as results (such that only users from the specified
context will see the IAI). An "Optional" context may indicate that
the system will favor results from a matching context, but will not
enforce it, such that searches lacking the specified context (eg.
users from a location other than a location specific context of an
IAI) may still be exposed to the information.
[0069] According to some embodiments, contexts of IQs may be
divided into "Explicit" contexts and "`Implicit" contexts. In such
embodiments, an "Implicit" context may be derived from an IQ and/or
IAI based on the text of the IQ or IAI, whereas an "Explicit"
context is derived by the system from a computational
surrounding/environment of a user sending a query, irrespective of
the text of the query (e.g. IP address, current website being used,
user profile information, etc).
[0070] According to some embodiments, `Mandatory` & `Explicit`
contexts may be used in user-type contexts (e.g. free/paying),
security-level contexts, and other critical contexts. In other
words, `Mandatory` & `Explicit` contexts may be used when a
proprietor of the system wants to deny users not having the
relevant contexts access to the specific IAIs, e.g. if the user
hasn't paid for them. Another example of a case when an explicit
context may be mandatory may be a commercial offer available only
to certain users--e.g. an offer valid only in a specific location.
In this case, a user from another location will not be presented
with the answer including this offer even if his query specifically
recites the relevant context (e.g. a user from Finland inquiring
"how much does a cost in Kansas'?"--in this case this user may not
be presented with an IAI describing a special offer available in
Kansas, although the context (geo location) is clearly implicit to
the query, as this IAI may have a mandatory explicit context of
Kansas).
[0071] According to some embodiments, `Optional` and `Implicit`
contexts may be used in less critical contexts, such as
product-page, gender, etc. where a user is likely to hint on the
relevant context in the search query, and there is no strict
limitation on what information may or may not be displayed to a
specific user.
Adding New IAIs to the Index and Removing Others
[0072] The process of adding IAIs into the index is a relatively
straight-forward process, having understood the scoring and
matching process of the invention. The basis for adding IAIs into
the index is the creation of the IAI's ILOR.
[0073] To add a new IAI to the index, first an ILOR must be created
with IKW REFs that is to be used to access the IAI. Typically, the
ILOR is created by converting an IQ's OQWS to an IKW REF, by
converting each NLW in the IQ's OQWS to an IKW REF with the same SL
score that was calculated for the NLW in the OQWS. Each of the IKW
REFs in the ILOR is then inserted into the appropriate IKW LOR in
the index, at its appropriate sorted position. The addition of a
new IAI item mandates an update of the SL score of the affected
NLWs in both the private storehouse 26 and the language storehouse
30, as the occurrence frequency of the NLWs changes (additional
occurrences have been presented). This change in SL score changes
the potential SL scores and SWS and MLS scores of the entire search
operation and requires a maintenance process to ensure consistency
throughout the SI. Of course, context may also be considered when
adding a new IAI.
[0074] Removing IAIs from the SI is the simple process of removing
the IKW REFs from their respective LOR. This change also affects
the frequency of the referenced NLWs and maintenance is
required.
Maintenance of the SI and Language Storehouses
[0075] As the occurrence frequency of NLWs changes (due to addition
of new IAIs, removal of IAIs etc.), the change normally affects the
SL score of IKW REFs in the SI and possibly other indexes as well,
as the NLW SL score of each IKW REF changes in accordance with the
NLW SL score in both the private storehouse 26 and the language
storehouse 30. Therefore, a subsequent re-calculation and
re-positioning of every affected IKW REF should be applied. In
addition, the SL score of NLWs and IKWs themselves is changed.
[0076] In another aspect of the invention, re-calculation and
re-factoring of the various index entities (i.e NLW, IKW REF etc.),
which will be referred to hereinafter as the maintenance process,
is managed in a way to only perform maintenance on IKW LORs that
their associated NLW SL score has been modified beyond a specified
factor, effectively limiting the amount of maintenance performed on
the index considerably while still performing maintenance on items
that are affected the most. Concurrently, the language and private
storehouses of words are recompiled as a part of the maintenance
process.
[0077] The present invention can be practiced by employing
conventional tools, methodology and components. Accordingly, the
details of such tools, component and methodology are not set forth
herein in detail. In the previous descriptions, numerous specific
details are set forth, in order to provide a thorough understanding
of the present invention. However, it should be recognized that the
present invention might be practiced without resorting to the
details specifically set forth.
[0078] In the description and claims of embodiments of the present
invention, each of the words, "comprise" "include" and "have", and
forms thereof, are not necessarily limited to members in a list
with which the words may be associated.
[0079] Only exemplary embodiments of the present invention and but
a few examples of its versatility are shown and described in the
present disclosure. It is to be understood that the present
invention is capable of use in various other combinations and
environments and is capable of changes or modifications within the
scope of the inventive concept as expressed herein.
[0080] While certain features of the invention have been
illustrated and described herein, many modifications,
substitutions, changes, and equivalents will now occur to those
skilled in the art. It is, therefore, to be understood that the
appended claims are intended to cover all such modifications and
changes as fall within the true spirit of the invention.
* * * * *