Semantic analysis system for ranking search results

Abstract

A computer-implemented semantic analysis system for ranking search results has instructions including the steps: receive a query about a first document authored by a person; read in data identifying the first document; read in a second document about the person authored after the first document; identify indicator terms in the second document related to whether or not there is a significant discrepancy between the second document and the first document; read in weighting factors for the identified indicator terms; determine an aggregate weighting factor that decreases with an increasing number of indicator terms found; and calculate a score of how much of a discrepancy there is between the second document and the first document using the weighting factors. When multiple first documents are returned for multiple persons, present the links in order from high score to low score.

Description

COPYRIGHT AND TRADEMARK NOTICE

A portion of the disclosure of this patent document contains material to which a claim for copyright is made. The copyright owner has no objection to the facsimile reproduction by anyone of the patent document or the patent disclosure, as it appears in the Patent and Trademark Office patent file or records, but reserves all other copyright rights whatsoever.

BACKGROUND OF THE INVENTION

There is need for a computer-based semantic analysis system for ranking search results wherein the search query seeks to find first documents authored by persons with content about themselves that have a significant discrepancy with second documents that have content about said persons. The second documents may be authored by the persons themselves or by other persons that are different than said persons.

SUMMARY OF THE INVENTION

The summary of the invention is provided as a guide to understanding the invention. It does not necessarily describe the most generic embodiment of the invention or the broadest range of alternative embodiments.

FIG. 1 is a schematic of a computer-based semantic analysis system 100 for ranking search results. FIG. 2 is a flow chart of a method 200 for semantic ranking of search results. FIG. 3 is a flow chart of a method 300 for determining a semantic ranking score. Referring to FIGS. 1, 2, and 3, the semantic analysis system comprises:

• • a) an input device 102 for reading in data;
  • b) a microprocessor 104 for processing said data;
  • c) an output device 106 for outputting said processed data; and
  • d) a permanent memory 108 comprising computer readable instructions to physically cause said microprocessor to carry out a method 200 (FIG. 2) comprising the steps:
    • i) receive 112, 202 via said input device, a search query (e.g. from a user 110);
    • ii) responsive to said search query, read in 113, 204 data (e.g. a link 176, such as a hyperlink) identifying a first document 114 authored by a person 111 in a set of persons wherein:
      • 1) said first document comprises textual content 116 about said person;
      • 2) said first document was authored at a first time; and
      • 3) said first document is computer accessible by a link 176;
    • iii) read in 122, 206 a second document 124 wherein said second document was authored after said first time and wherein said second document comprises textual content 126 about said person;
    • iv) read in 132, 208 a set of indicator terms 134 that are associated with a discrepancy between said textual content of said second document and said textual content of said first document;
    • v) determine 136, 210 which 138, 142 of said indicator terms are present in said second document;
    • vi) determine 140, 212 a score 152 indicating how much of a discrepancy there is between said textual content of said second document and said textual content of said first document according to the steps 300:
      • 1) read in 144, 302 an indicator weighting factor 146, 148, 154 for each of said indicator terms present 156, 158 in said second document (note: δ_i has a numerical value of 1 when an indictor term is present and a numerical value of 0 when an indicator term is not present);
      • 2) determine 304 a first sum 162 of all of said indicator weighting factors for all of said indicator terms present in said second document;
      • 3) determine 306 a second sum equal to said first sum plus a constant D 164;
      • 4) determine 308 a count N 166 of the number of said indicator terms present in said second document; and
      • 5) determine 310 said score by multiplying said second sum by an aggregate weighting factor C_N 168 that has a value of 1 when N is 1 and decreases with increasing N over a range of N;
    • vii) compare 170, 214 (FIG. 2) said score 152 to a threshold 172; and
    • viii) provide 174 by said output device, said link 176 to said first document when said score is greater than or equal 178 to said threshold.

BRIEF DESCRIPTION OF DRAWINGS

FIG. 1 is a schematic of a semantic analysis system for ranking search results.

FIG. 2 is a flow chart of a method for semantic ranking of search results.

FIG. 3 is a flow chart of a method for determining a semantic ranking score.

FIG. 4 is a flow chart of a method for determining an aggregate weighting factor.

FIG. 5 is a flow chart of a method for automatically adjusting a threshold.

FIG. 6 is a flow chart of a method for automatically adjusting weighting factors.

FIG. 7 is a table presented on a screen of results of the method for automatically adjusting weighting factors.

FIG. 8 is a graph presented on a screen of the values of aggregate weighting factors C_N versus N produced by the method for automatically adjusting weighting factors.

FIG. 9 is a graph presented on a screen illustrating the accuracy of the method for semantic ranking of search results with and without an aggregate weighting factor C_N that decreases with increasing N over a range of N.

DETAILED DESCRIPTION

The detailed description describes non-limiting exemplary embodiments. Any individual features may be combined with other features as required by different applications for at least the benefits described herein.

As used herein, the term “about” means plus or minus 10% of a given value unless specifically indicated otherwise.

As used herein, a “computer-based system” comprises an input device for receiving data, an output device for outputting data in tangible form (e.g. printing or displaying on a computer screen), a permanent memory for storing data as well as computer code, and a microprocessor or other type of digital processor for executing computer code wherein said computer code resident in said permanent memory will physically cause said microprocessor to read in data via said input device, process said data by said microprocessor and output said processed data via said output device.

An input device may be any device adapted to receive input and, if needed, convert said input to computer usable form (e.g. digital encoding). Input devices include keyboards, cameras, microphones and digital inputs for receiving data from other computer-based systems.

An output device may be any device adapted to output data. Output devices include computer screens, speakers, printers and digital output devices for sending data to other computer-based systems.

As used herein, a formula or equation is only an expression of an algorithm or method. For example, the formula “Y=A+B” is an expression of the prose algorithm “Set the value of Y equal to the value of A plus the value of B”. A person of ordinary skill in the art of computer science will understand that said prose algorithms can be expressed as a series of one or more steps of a computer coded algorithm. As used herein, therefore, an equation or formula is not an expression of an abstract idea, a law of nature or any other fundamental truth.

Ranking Multiple Search Results

Multiple search results can be ranked according to the method illustrated in FIG. 2. The method for ranking search results comprises the steps:

• • a) execute 216 said steps ii to vii (i.e. items 204 to 214) for each person in said set of persons to determine a score for each first document authored by said each of said persons, wherein each of said first documents is accessible by a link;
  • b) rank 218 said links to said first documents in order of highest score to lowest score; and
  • c) output 220 by said output device, said links to said first documents in said order of said rank.

As used herein, “ranking from high score to low score” also encompasses ranking from low score to high score. For example, if one determines a score equal to the inverse of the score described herein, then ranking would be from low inverse score to high inverse score.

Determining an Aggregate Weighting Factor

FIG. 4 is a flow chart of a method 400 for determining an aggregate weighting factor, C_N. The method comprises the steps:

• • a) determine 402 a first number by subtracting 1 from N;
  • b) determine 404 a second number by multiplying said first number by a parameter A and then adding 1; and
  • c) determine 406 said C_N by taking the inverse of said second number.The parameter A may have a value of about 0.07. The parameter A may have values such that C_N decreases from 1 to 0.8 or less over a range of N of 1 to 2.

The constant D as described above may have a value less than zero. D may have a value of about −0.00138.

Automatically Adjusting a Threshold

FIG. 5 is a flow chart of a method 500 for automatically adjusting a threshold, V′₀. The method comprises the steps:

• • a) determine 502 a numerator equal to a square of said score;
  • b) determine 504 a denominator equal to 1 plus said square of said score;
  • c) determine 506, 180 (FIG. 1) a probability of there being a significant discrepancy between said content of said second document and said content of said first document by taking a ratio of said numerator to said denominator;
  • d) determine 508 a plurality of said probabilities for a plurality of first documents authored by a plurality of said persons in said set of persons wherein:
    • i) each of said plurality of said probabilities is based on a plurality of second documents each associated with each one of said plurality of said persons in said set of persons; and
    • ii) each of said second documents comprises content;
  • e) determine 510 an average of said plurality of said probabilities; and
  • f) automatically adjust 512 said threshold such that said average of said plurality of said probabilities has a desired value.A suitable value of said desired value is in the range of 0.8 or greater.

Automatically Adjusting Weighting Factors

FIG. 6 is a flow chart of a method 600 for automatically adjusting the indicator weighting factors and the aggregate weighting factor. The method comprises the steps:

• • a) read in 602 for each of said first documents associated with said plurality of persons, an adjudication of whether or not there is a significant discrepancy between each of said contents of said second documents and each of said contents of said first documents;
  • b) determine 604 a plurality of residuals between each of said probabilities in said plurality of probabilities and each of said adjudications;
  • c) determine 606 a sum of squares of said residuals;
  • d) automatically adjust 608 said indicator weighting factors and said aggregate weighting factor such that said sum of squares of said residuals has a minimized value.

As used herein, a “minimized value” is a value that is reduced from an initial level to a subsequent lower level. It does not necessarily mean that a given value is at its lowest possible level.

As used herein, an “aggregate weighting factor” may include multiple values of the aggregate weighting factor for different values of N over a range of N. Hence multiple values of the aggregate weighting factor for a range of values of N may be automatically adjusted so that the sum of squares of the residuals has a minimized value.

As used herein, an “adjudication” is an independent assessment of whether or not there is a significant discrepancy between the content of a second document and the content of a first document. The independent assessment may be performed automatically at least in part by a computer-based system, performed at least in part by a person, or a combination of both. For the purposes of computation, an adjudication may have a numerical value of 1 when there is a significant discrepancy and a numerical value of 0 when there is not a significant discrepancy. Alternatively, an adjudication may have a fractional value between 0 and 1 when the adjudication is an independent estimate of the probability of there being a significant discrepancy.

As used herein, a probability may be expressed in any units, such as fractional (i.e. 0 to 1) or percentage (i.e. 0% to 100%). Thus, any particular units (e.g. fractional) recited herein encompass all units (e.g. percentage).

EXAMPLES

It has been discovered that the methods described herein are effective at determining if there is a likelihood of a significant discrepancy between the content of a first document authored by a person at a first time and a second document with content about the person authored at a later time. The first document might be an application form. The application form may be one or more of a job application, an application for citizenship, an insurance claim, a posting on a dating web site, an application for security clearance, a loan application, a patent application or any document authored by a person that is to be submitted to a third party for an approval.

A person submitting an application form to a third party for approval may be motivated to withhold information or falsify information in order to secure said approval. The third party, therefore, has need to find out if the information in the application form is consistent with other information about the person that might be in a second document either authored by the same person or another person. Second documents may include social media postings by the same person, a blog posting by another person or any other document in any other medium. If the application form is for an insurance claim related to a loss event suffered by the person (e.g. an on-the-job injury), then it may be desirable for the second document to be authored after the first document or at least after the loss event since the insurance claim attests to a change in health of the person.

A third party may have a list of indicator terms that, if found in a second document, indicate that there is a discrepancy between the content of the second document and the content of the first document. The loss terms may be specific to the type of application form and the contents of the application form. Referring to FIG. 1, for example, if a person 111 (e.g. initials KJ) submits an application form 114 for a workers' compensation insurance claim and the content 116 of the application form states that the person suffered a back injury (e.g. “I hurt my back”), but another person 115 (e.g. initials PJ) states in a blog posting or other social media posting “After we went to the gym, KJ (i.e. the person) and I played a great game of baseball”, and if the second document was authored after the first document, then that would indicate that perhaps the person did not injure their back after all.

Second documents authored before the purported change in health may be examined to determine if the adverse health condition of the person existed before the alleged injury. The indicator terms, therefore, may be different depending upon whether the second document was authored before or after the first document. The indicator terms may also be different depending upon the nature of the alleged injury.

The third party may have a large number of similar application forms to automatically process in a short period of time. The third party, therefore, may develop a list 130 (FIG. 1) of indicator terms 134 and associated indicator weighting factors 135 such that if the indicator terms are found 138, 142 in the second documents, the semantic analysis system can then determine scores V′ 152 for the first documents based on the second documents. The first documents with scores above a threshold V′₀ may then be ranked by said scores and presented to a user 110 in the order of said ranking. Thus, the user can focus on the high scores first and investigate them further to determine if the associated application forms should be accepted.

Example 1: Automatically Adjusting Weighting Factors

Any method, including human judgment, may be used to develop an initial set of indicators and weighting factors. The weighting factors may be adjusted over time, however, as additional data becomes available. This is an example of machine learning.

FIG. 7 is a partial table 700 of results of the method 600 (FIG. 6) for automatically adjusting weighting factors after said method 600 had been applied to additional data that had become available. The table was presented on a screen 720 as part of a graphical user interface.

To execute the method 600, a set of one hundred workers' compensation insurance claims (i.e. first documents) from one hundred claimants 702 (i.e. persons in a set of persons) were adjudicated 712 to determine which claims were legitimate (i.e. accepted applications as indicated by a “0”) and which claims were not (i.e. applications not accepted as indicated by a “1”). The results were read in to the computer-based semantic analysis system 100. The system then accessed the social media sites of the persons submitting the insurance claims and identified second documents authored after the dates of the claims. The second documents had content about the persons. The system also read in a list 704 of six indicator terms from a permanent memory. The list of indicator terms comprised the terms “baseball”, “gym”, “hiking”, “running”, “tennis”, and “weekend”. Other indicator terms could have been included as well. The system then read in initial values of associated indicator weighting factors 722, the parameter A 718 and the constant D 716. The system then determined if the indicator terms were (“1”) or were not (“0”) found 705 in the second documents. The system then determined to total number N of indicator terms found in each 706 of the second documents. The system then calculated risk factors 708 for each person and associated probabilities 710 per the methods described above. The system then calculated the squares (e.g. 728, 726, 724) of the residuals between the adjudications and the probabilities. The system then determined the sum 714 of the squares of the residuals. The system then automatically adjusted the indicator weighting factors, the parameter A and the constant D used to calculate the aggregate weighting factors versus N so that the sum of squares of the residuals had a minimized value (i.e. a magnitude lower than its initial value). The final results of the minimization are shown in table 700.

Expressions of Methods as Formulas

In order to facilitate code writing, the methods described herein may be compactly expressed as formulas. Since formulas are only used to express methods in a compact form, they are not expressions of fundamental truths or laws of nature. For example, the method 300 of determining a semantic ranking may be expressed as:

$V^{\prime }=C_N\left(\sum _{i=1}^M{k}_i{\delta }_i+D\right)$ where

• • V′ is the score;
  • C_N is the aggregate weighting factor;
  • N is the number of indicator terms found in a second document;
  • M is the total number of indicator terms in a set of indicator terms;
  • i is an index number for each indicator term;
  • k_i is an indicator weighting factor for indicator i;
  • δ_i has a value of 1 when an indicator term is present in a second document and a value of 0 when an indicator term is not found in a second document; and
  • D is a constant.

Aggregate Weighting Factor Versus N

FIG. 8 is a graph 800 of suitable aggregate weighting factors C_N versus the number N of indicator terms found in second documents. The graph is presented on screen 720 as a graphical user interface. The method 400 of calculating C_N from the parameters A and N can be expressed as 802:

$C_N=\frac{1}{1+A\left(N-1\right)}$ where the variables are defined above. Any method that decreases C_N over a range of N, however, can be used. A C_N that decreases from 1 at N=1 to 0.8 or less at N=2 is suitable. Physically, a decreasing C_N with increasing N means that each additional indicator term found in a second document adds incrementally less to the overall score that what one would expect if the indicator weights were merely summed without the aggregate weighting factor.

Probability of Significant Discrepancy

The method steps 502 to 506 of determining a probability P′ of there being a significant discrepancy between the content of a second document and the content of a first document can be expressed as:

$P^{\prime }=\frac{{V^{\prime }}^2}{\left(1+{V^{\prime }}^2\right)}$ where the variables are defined above.

The method steps 502 to 506 are an example of an asymptotic transformation. As used herein, an asymptotic transformation is a transformation of an increasing variable to a variable with a finite upper limit, such as 1. Any asymptotic transformation may be used to convert a score into a probability. This includes an exponential transformation.

Example 2: False Positives and False Negatives

The methods described in Example 1 were repeated with a larger sample for 5,288 workers' compensation insurance claims. All claims were adjudicated to determine if they were legitimate (i.e. no significant discrepancy) or fraudulent (i.e. significant discrepancy). 2,330 of the claims were confirmed to have a significant discrepancy. The social media data (i.e. second documents) for the claimants (i.e. persons in a set of persons) was read in and compared to a list of 59 indicator terms. 2,126 of the claimants had one or more indicator terms in their social media data and had scores above a threshold. The social media data was all authored subsequent to the time of injury listed on the insurance claims. 1,770 of the claims that had scores above the threshold were adjudicated as having a significant discrepancy. 356 of the claims with scores above the threshold were adjudicated not having a significant discrepancy. Thus, the overall false positive rate of claims with scores above the threshold was about 356/2126=16.7%.

3,162 of the claims did not have an indicator term in the associated social media data (i.e. second documents) and hence had scores below the threshold. 630 of these claims were adjudicated as having a significant discrepancy. 2,532 of these claims were adjudicated as not having a significant discrepancy. Thus, the false negative rate of claims without an indicator term was about 630/3162=19.9%.

A desired probability may be selected depending upon the relative costs of false positives and false negatives so that the overall operating cost of the system is minimized.

Effect of Aggregate Weighting Factor

FIG. 9 shows a graph 900 of the effect of the aggregate weighting factor on the accuracy of the probability estimates of a significant discrepancy. The graph is displayed on screen 720 as part of a graphical user interface. The data points 902 show the percent of second documents with an adjudicated significant discrepancy versus the number N of indicator terms found in the associated second documents. About 65% of the claims with 1 indicator term were adjudicated to have a significant discrepancy. About 86% of the claims with 2 indicator terms were adjudicated to have a significant discrepancy.

The solid line 906 shows the calculated probability of a significant discrepancy using the methods described herein with an aggregate weighting factor that decreased with increasing N. The agreement between the calculated solid line and the adjudicated data points is good.

The dashed line 904 shows the calculated probability without an aggregate weighting factor. Without an aggregate weighting factor that decreased with increasing N, the calculated probability of a significant discrepancy is increasingly over estimated relative to the adjudicated data as N increases.

CONCLUSION

While the disclosure has been described with reference to one or more different exemplary embodiments, it will be understood by those skilled in the art that various changes may be made and equivalents may be substituted for elements thereof without departing from the scope of the disclosure. In addition, many modifications may be made to adapt to a particular situation without departing from the essential scope or teachings thereof. For example, as used herein, an “indicator term” may be a word, phrase, number, concept, static image, moving image, sound or any other content. A “document” can be any record in any medium, such as a video, social media posting, handwritten document on paper, or digital image. Therefore, it is intended that the disclosure not be limited to the particular embodiments disclosed as the best mode contemplated for carrying out this invention.

Claims

1. A semantic analysis system for ranking search results comprising: a) an input device for reading in data;b) a microprocessor for processing said data;c) an output device for outputting said processed data; andd) a permanent memory comprising computer readable instructions to physically cause said microprocessor to carry out the steps: i) receive from a user via said input device, a search query;ii) responsive to said search query, read in data identifying a first document authored by a person in a set of persons wherein: 1) said first document comprises textual content describing an injury suffered by said person on or before a first time;2) said first document was authored at said first time; and3) said first document is computer accessible by a link;iii) read in a second document wherein said second document was authored after said first time and wherein said second document comprises textual content about an activity performed by said person after said first time;iv) read in a set of indicator terms that have been statistically linked to a discrepancy between said textual content of said second document and said textual content of said first document wherein said discrepancy indicates that said person did not suffer said injury on or before said first time, and wherein said statistical linkage has been done without input from said first document, said second document, or said user;v) determine which of said indicator terms are present in said second document;vi) determine a score indicating how much of said discrepancy there is between said textual content of said second document and said textual content of said first document according to the steps: 1) read in an indicator weighting factor for each of said indicator terms present in said second document;2) determine a first sum of all of said indicator weighting factors for all of said indicator terms present in said second document;3) determine a second sum equal to said first sum plus a constant D;4) determine a count N of the number of said indicator terms present in said second document; and5) determine said score by multiplying said second sum by an aggregate weighting factor CN that has a value of 1 when N is 1 and decreases with increasing N over a range of N;vii) compare said score to a threshold; andviii) provide by said output device said link to said first document when said score is greater than or equal to said threshold.

2. The semantic analysis system of claim 1 wherein said steps further comprise: a) execute said steps ii to vii for each person in said set of persons to determine a score for each first document authored by said each of said persons, wherein each of said first documents is accessible by a link;b) rank said links to said first documents in order of highest score to lowest score; andc) output by said output device, said links to said first documents in said order of said rank.

3. The semantic analysis system of claim 1 wherein CN decreases from 1.0 to 0.8 or less over a range of N of 1 to 2.

4. The semantic analysis system of claim 1 wherein said steps further comprise: a) determine a first number by subtracting 1 from N;b) determine a second number by multiplying said first number by a parameter A and then adding 1; andc) determine said CN by taking the inverse of said second number.

5. The semantic analysis system of claim 4 wherein said parameter A is about 0.07.

6. The semantic analysis system of claim 1 wherein D is set to a value less than zero.

7. The semantic analysis system of claim 1 wherein said first document is an application form submitted to a third party for approval.

8. The semantic analysis system of claim 7 wherein said first document is one or more of: a) a job application;b) an application for citizenship;c) an insurance claim;d) a posting on a dating web site;e) an application for security clearance;f) a loan application; org) a patent application.

9. The semantic analysis system of claim 1 wherein said second document is one or more of: a) a social media posting authored by said person; orb) a blog post authored by another person who is not said person.

10. The semantic analysis system of claim 1 wherein said steps further comprise: a) determine a numerator equal to a square of said score;b) determine a denominator equal to 1 plus said square of said score;c) determine a probability of there being a significant discrepancy between said content of said second document and said content of said first document by taking a ratio of said numerator to said denominator;d) determine a plurality of said probabilities for a plurality of first documents authored by a plurality of said persons in said set of persons wherein: i) each of said plurality of said probabilities is based on a plurality of second documents each associated with each one of said plurality of said persons in said set of persons; andii) each of said second documents comprises content;e) determine an average of said plurality of said probabilities; andf) automatically adjust said threshold such that said average of said plurality of said probabilities has a desired value.

11. The semantic analysis system of claim 10 wherein said desired value of said average of said plurality of said probabilities is 0.8 or greater.

12. The semantic analysis system of claim 10 wherein said steps further comprise steps for statistically linking said indicator terms to said discrepancy between said textual content of said second document and said textual content of said first document, said steps for statistically linking comprising: a) read in for each of said first documents associated with said plurality of persons, an adjudication of whether or not there is a significant discrepancy between each of said contents of said second documents and each of said contents of said first documents;b) determine a plurality of residuals between each of said probabilities in said plurality of probabilities and each of said adjudications;c) determine a sum of squares of said residuals;d) automatically adjust said indicator weighting factors and said aggregate weighting factor such that said sum of squares of said residuals has a minimized value.

13. The semantic analysis system of claim 12 wherein said adjudications have a numerical value of 1 when there is a significant discrepancy and a numerical value of 0 when there is not a significant discrepancy.