Method of indexing and retrieval of electronically-stored documents
A document indexing and retrieval system and method which assigns weights to the key words and assigns a relative value to pairs of key words (i.e. defines a relative relation on K.times.K) based on their frequency of occurrence and co-occurrence in the document data base. In response to a query both the weights and this relative relation are used to suggest additional and/or alternative key words which are very likely to find relevant documents. Documents are then ranked by number of hits adjusted for the weights of hit words and their relative values.
1. Field of the Invention
This invention relates generally to document storage and retrieval systems and more particularly to a method of indexing documents so that they can be retrieved in response to a query in order of their relevance to the query. It also permits, general query to be easily modified based on the content of the documents so that the new query will retrieve documents that are relevant to the original query.
2. Description of the Prior Art
Document retrieval based on indexing of the documents in a document data base is well known. Typically the documents are indexed by creating an index file which records the documents that each word is in. Then when the user inputs a query, the documents that contain one or more words of the query can be quickly identified. However, if the query consists of general words that are not terms of art, the query may produce unsatisfactory retrieval results by either producing few documents that are of interest to the user or producing many documents that are not interesting to the user or both.
SUMMARY OF THE INVENTIONA principal object of the present invention is to provide an improved method of indexing and retrieving documents which:
(A) allows a user to easily modify his query based on the content of the documents so that the new query will retrieve documents that are of interest to the user;
(B) accurately ranks the documents in order of relevance to the query; and
(C) allows the user to peruse the documents extremely quickly.
Another object of the present invention is to use the Soft Boolean Connector concept to adjust the number of hits (i.e., the number of query words that a document is credited with for ranking purposes) by giving less than a full hit to a word that often co-occurs with other query words. Another object of the present invention is to use the Soft Boolean Connector concept to adjust the number of hits (i.e. the number of query words that a word is credited with by virtue of its being related to those query words) for a possible suggested word by giving less than a full hit to a word that often co-occurs with the other query words.
These objects, as well as other objects which will become apparent from the discussion that follows, are achieved according to the present invention by the following steps (note: in the following the words "term" and "keyword" stand for both a single word and a phrase consisting of a group of words, e.g., "patent application".):
1. Indexing the documents by creating index files of which documents contain each term, how many times the term appears in the document, and how many documents each term appears in.
2. Assigning as many weights to each term as there are documents that contain that term, where the weight of a term in a document depends on the number of times the term appears in the document, the number of documents that the term appears in, and the total number of terms in the document.
3. Constructing for each term a ranked list of companions of said term which list contains the terms (companions) that appear in the same documents as said term in order of the sum of the weights of the companions over all documents that contain both the term and the companion. Associated with each companion is the companion percentage which is the sum used to rank the companions.
4. Using the companion lists to construct relative lists for each term which relative lists usually contain only those companions which also have said term as a companion. Associated with each relative is the relative percentage which is a weighted average of the companion's percentage as a companion of the term and the term's companion percentage as a companion of the companion. The relative percentages are used to rank the relatives.
5. Assigning a "polysemantic" weight to each term, which polysemantic weight depends on the number of documents that the term is in, the number of relatives that the term has, and the relative strength of the first few relatives to the other relatives.
6. Presenting to the user, in response to a query, a list of "SWAPS" (Synthetic Word Association Pattern Search) terms that are the best relatives to the entire group of terms contained in the query and allowing the user to add one or more of the presented terms to the query.
7. Ranking the documents according to how many query terms are contained in the document, their polysemantic weights and their weights in the documents.
The present invention facilitates the rapid searching of a document data base for documents that are of interest to the user. By using the suggested SWAPS terms the user can modify his query so as to retrieve those documents, if they exist in the data base, which are of interest. Since the SWAPS terms that are presented are in many of the documents that the original query terms are in, adding them to the query is guaranteed to retrieve those documents and others containing the SWAPS terms. By using the SWAPS feature repeatedly the user can in effect roam around the data base without actually retrieving and reading documents. Only after the query has been modified to include all the interesting SWAPS terms, does the user need to actually retrieve the documents. The user can start with a poor query and modify it using SWAPS so that it becomes a good query. The user need not waste time formulating a good query that will not retrieve any relevant documents because there happen to be no such documents in the data base. The SWAPS terms that are suggested will always retrieve documents that contain them i.e. documents that are likely to be relevant.
The ranking of the documents also facilitates rapid searching because the user can be confident that the highest ranked documents will be the documents that are most relevant to the query and that all documents which have any relevance will be retrieved and ranked.
The foregoing and other objects, features and advantages of the present invention will become apparent from the following, more particular description of the preferred embodiments of the invention, as illustrated in the accompanying drawings.
BRIEF DESCRIPTION OF THE DRAWINGSFIG. 1 is a block diagram of a computer system embodying the present invention;
FIG. 2 is a view of the display screen showing an entered query and the result of parsing it;
FIG. 3 is a view of the display screen showing suggested SWAPS terms for the query of FIG. 2;
FIG. 4 is a view of the display screen showing the modified query;
FIG. 5 is a view of the display screen showing suggested SWAPS terms for the modified query of FIG. 4;
FIG. 6 is a view of the display screen showing a second modification of the query based on choosing SWAPS terms from FIG. 5;
FIG. 7 is a view of the display screen as a result of ranking the documents for the query of FIG. 6;
FIG. 8 is an operational flow diagram for indexing a set of documents;
FIG. 9 is a procedure tree for the QSEARCH program used for searching an indexed set of documents using the SWAPS and RANKING features;
FIGS. 10A to 10J are description of the program modules in FIG. 8;
FIG. 11 is a description of the program modules in FIG. 9; and
FIGS. 12A to 12C are description of the ABSTRACT program module.
DESCRIPTION OF THE PREFERRED EMBODIMENTSThis invention will now be described as embodied in a computer system of the type shown in FIG. 1. This embodiment utilizes the following computer hardware and software:
(1) IBM compatible personal computer with at least 4 MB of RAM, a large capacity hard drive, a display screen, and a keyboard.
(2) MS-DOS compatible operating system and LIM 3.2 compatible expanded memory manager.
(3) A vocabulary file of terms (words and phrases)
(4) A series of programs that index the documents by constructing various files that hold information about which terms are in which documents, which documents contain which terms, the weights of the terms, and which terms are relatives of other terms by virtue of occurring in the same documents and how strongly are they are related.
(5) A user program that accepts a query, suggests modifications to the query, and ranks the documents based on the modified query using the weights and relative strengths of the terms of the query.
The Vocabulary file is structured as a list of headwords each with a short synonym list. All of the synonyms of a given headword are assigned the same code.
The full list of indexing programs can be found in FIG. 10. Here we will describe the most important of these programs: AIM, AIMPASS2, FREQCOMP, RELATIVE, and POLYSEMY.
The first indexing program is AIM.BAS: Automatic Indexing Module. It creates DocKeys, DocIndex, and IDF. DocKeys holds all of the Keywords and Keyword-Counts for all documents. IDF holds the document frequency, i.e., the number of documents a keyword appears in.
As the words in the documents are checked against the vocabulary to see if they are keywords, the case (upper or lower) is possibly changed and they are stripped of prefixes to see if the different case or stem is a keyword according to the following algorithms: (UC=upper-case and LC=lower-case)
IF UC word is at the beginning of a sentence AND we don't have it in our vocabulary as a LC word THEN look for it the Vocabulary as an UC word
IF UC word in middle of sentence AND we don't have it UC THEN look for it if it doesn't have a typical proper name ending
In USER Program Only: IF word NOT found THEN find both the stem AND find the Good prefix (In the following "find" means that the stem and/or prefix is said to be in the document if the prefix is of the right type and the stem has the indicated length and is a keyword.)
IF GOOD prefix THEN
Find GOOD prefix if stem>3 characters long
IF word is found THEN find if stem >8 characters long
IF word is NOT found THEN find if stem >5 characters long
IF POOR prefix THEN
If word is found THEN DON'T find stem
If word is NOT found THEN find if stem >5 characters long
List of Poor Prefixes:
hi, co, de, en, ex, im, in, un, re, con, eco, dis, epi, mal, mid, mis, non, off, out, pre, pan, sub, uni, demi, down, fore, hemi, high, meta, over para, peri, post, self, semi, after, inter, quasi, trans, under
List of Good Prefixes:
air, bio, sea, sky, top, aero, anti, auto, back, head, home, homo, hemo, mega, mini, mono, rear, poly, self, tele, viro, chemo, ferro, homeo, hyper, infra intra, macro, micro, multi, hydro, radio, super, supra, ultra, contra, hetero, thermo, techno, nucleo, counter, electro, magneto
The next indexing program is AIMPASS2.BAS. It creates Key and Weight files. The nth Rec of Key.Ndx contains NumKeysinDoc(n) followed by up to 127 Key codes which have Weight greater than or equal to the Adaptive Threshold Value. The Adaptive Threshold Value is the average Weight value of the 80th Keyword in each document (0 if there are less than 80 Keywords in a document). The nth Rec of Weight. Ndx contains up to 127 (or as many Keywords are above the Adaptive Threshold Value) Document Weights computed with the following weight formula: ##EQU1##
FREQCOMP.BAS implements the Inverted Index access method along with the weighted values to calculate the frequent companions for each of the words used in the document collection.
For each word ("A") in the controlled vocabulary dictionary, the WEIGHT (see above formula) values for each co-occurring word in the document (a co-occurring word to A is one that appears as a Keyword in the same document that A appears as a Keyword) are summed, along with the WEIGHT values for A in that document, respectively in all documents in which they co-occur. The sum values for each co-occurring word are converted to a percentage, scaled to the sum value for A (i.e., percentage=sum for word's WEIGHT values divided by the sum for A's WEIGHT values). Note that the percentages for the co-occurring words can be higher than 100% if they are heavily weighted in the same records in which A appears. The co-occurring words are then sorted in descending order (from highest percentage value to lowest) and the top 127 are written to a file (see below for structure). If there are 127 co-occurring words or fewer, then all of the co-occurring words will be written in descending sorted order..
______________________________________
Definitions:
{ } = co-occurring
weight = WEIGHT value
Example:
Doc# 1 Doc# 2 Doc# 3 Doc# 4 Doc# 5
##STR1##
##STR2##
##STR3##
##STR4##
##STR5##
A { } B C D E . . .
B { } A C D E . . .
5.5 = sum of weight of all A's
6.4 = sum of weights of B's co-occurring with A
Sample:
A 5.5 { } B 6.4/116% . . .
B 6.4 { } A 4.0/63% . . .
Resulting File:
Main Word Co-Occurring Words . . . (sorted)
______________________________________
A B 116% . . .
B A 63% . . .
.
.
______________________________________
After the frequent companions have been found RELATIVE.BAS is run to define the relatives of each Keyword (A) according to the following algorithm:
are there any FreqComps for A? If so, then for each FreqComp of A (F):
look for F in A's FreqComp List and get its value
look for the word itself (A) in word A's FreqComp List and get its value
a apply formula of (Lower.times.6+Higher)/7, where Lower is the lower of the two values obtained in the above two steps and Higher is the higher of the two values.
a sort in the resulting list of words and values in decreasing order, by value
a save the first 63 (or as many as are found) of this list as the relatives for keyword A
For each word (called "A") in the dictionary which has Frequent Companions (not all do, because some words in the dictionary are not used at all in a database), take each Frequent Companion of A (called "F") and its Frequent Companion Percentage Value [FCPVal] in A's Frequent Companion List [FCList](called "F-VAL") and look for the FCPVal of A in F's FCList (called "A-VAL"). NOTE: If A is not found in F's FCList, then A-VAL is zero (0). The RELATIVE value for F is calculated by multiplying the smaller of F-VAL and A-VAL by 6, adding the larger of F-VAL and A-VAL, and then dividing that sum by 7. If both A and F are in each other's FC lists, the resulting Relative value will be added to both words' Relative lists. If F is in A's FC List, but A is not in F's then F-VAL will be divided by seven and added only to A's Relative list.
After all the RELATIVE values are calculated for each Frequent Companion (F) in A's FCList, they are sorted in descending order and the top 63 of these words are written to A's Relative List. If there are fewer than 63 Relatives, then all of the Relatives will be written to A's Relative List, in descending order of RELATIVE value. ##EQU2##
Here the SmallerPercent Value is the smaller of the A-VAL and the F-VAL and the LargerPercent Value is the larger of the A-VAL and the F-VAL.
______________________________________
Sample:
##STR6##
Resulting File
Main Word Relatives . . . (sorted)
______________________________________
A B 70 . . .
B A 70 . . .
.
.
______________________________________
After the relatives have been found each of the keywords is given a single polysemantic weight that does not change from document to document by the program POLYSEMY.BAS which uses the following formula: ##EQU3##
Here Avg.sub.n is the average of the relative percentages of the first n relatives of the keyword, TotRelVal is the sum of relative percentages over all relative lists that the keyword is in, and DocFreq is the number of documents that the keyword is in (having a WEIGHT above the adaptive threshold).
Once the indexing programs have been run, the ABSTRACT program is run to create highlights of the full text that will be presented to the user before or in place of the full text itself. First the documents are broken into sentences using a Sentence Ends Algorithm. Then the sentences are assigned weights (values) as a whole and the top ranked sentences are chosen to be part of the highlight. Finally a Sanitize algorithm is used to "X" out (eliminate) proper names from in the highlights. See FIG. 12 for specific details on the algorithms used in the ABSTRACT program.
Once the indexing and optionally the ABSTRACT programs have been run, the QSEARCH program can be used to search for documents. This is done by entering a query in natural language. The user program will parse the query to find all the keywords it contains using algorithms similar to those in the AIM program.
After the query is parsed the user is shown the keywords that are contained in the query in order of their polysemantic weight and is given the opportunity to add and delete words in the query and to have the program suggest SWAPS terms based on the query. These SWAPS terms are generated by generating for each keyword in the vocabulary a summed-relpoly-percentage which is the sum, over all terms that are in the query, of the relpoly percentages of that keyword, where the relpoly percentage is the product of the relative percentage and the polysemantic weight. Then the summed RelPoly percentages are adjusted using a concept called Soft Boolean Connectors to come up with a final SWAPS value for each keyword. The keywords are then ranked by SWAPS value and the highest ranked are presented to the user as suggested SWAPS terms to be added to the query.
The Soft Boolean Connectors concept involves penalizing pairs of terms that co-occur often (i.e., in many documents) when calculating the adjustment to be applied to the summed relpoly percentages.
______________________________________
First, Multiply the last group of SWAPS words by
Boost Factor (=2)
Then add relative values of relatives of
main word after each 15 multiplied by the
PolyValue of the Word
(The previous value will be called "Temp Value")
Create table for everyv pair combination of query words, e.g.,
for words A, B, & C, there are three pairs.
AB
AC
BC
______________________________________
For each pair of query words, ("A" & "B"), the Relative Value used in the formula below is B's Relative Value in A's Relative List, or, if B doesn't appear in A's Relative List, then the value is taken from A's Relative Value in B's Relative List (this is possible because the Relative Value between any two words is mutual), i.e., if B is found in A's Relative list, take just that value. You don't need to look at B's list to find A's value there because, if it is there, it would have the same value as B has in A's list. Only if B is not in A's Relative list check for A in B's list. Enter the Relative Penalty value resulting from the following formula into the table for each combination (pair):
______________________________________
##STR7##
Example:
if Relative values are
AB = 70
AC = 75
BC = 65
and A appears in 5 documents and B appears in 4 and
the avg. doc freq is 4, then using the
following formula for AB
##STR8##
the table value for AB is 1.77
(look below for maximum, making this 1.0 instead)
for a hit of 3 words - 3 relatives of main word or
3 words in a document
A B C
Adjust Value = 2.5.sup.(# of hits - sum of penalties)
Note: The "# of hits" value is:
##STR9##
Word/Document Value = Temp Value .times. Adjust Value
______________________________________
MAXIMUM PENALTY TABLE (SWAPS)
query words
Max.
______________________________________
(for each pair)
2 0.3
3 1.0
4 & up 0.9
(for sum of pairs)
2 0.3
3 1.4
4 1.8
5 2.3
6 & up 2.8
______________________________________
After the user has modified the query by choosing SWAPS terms, he can have the program suggest new SWAPS terms based on the new query. In this case the program boosts the relative percentages of the last chosen set of SWAPS terms before calculating summed relpoly percentages. This allows the user to navigate in the data base by modifying his query so that it will find documents containing the SWAPS terms.
For example, FIG. 2 shows the options the user will be presented with after entering the query "when can a contract be enforced". If the user chooses the menu option "Related Terms" he will be shown a list of SWAPS terms as shown in FIG. 3. This first set of SWAPS terms that are presented to the user includes the term "statutes". The user may choose one or more of these suggested SWAPS terms to add to the query. In FIG. 4 we see that the user has chosen to add the term "statutes" to the query. At this point the user can again ask the system to suggest SWAPS words. This time the previously added SWAPS term "statutes" will be given extra weight in determining which new terms are suggested to the user. In FIG. 5 we see the resulting suggested SWAPS terms generated from the four query terms "agreement", "statutes", "enforcement", and "can", with "statutes" given more weight than the other three terms. Notice that the SWAPS words are ranked somewhat differently than in FIG. 3 and in particular a new SWAPS term "statute of limitations" is suggested. By adding the term "statutes" to the query and then asking again for suggested SWAPS terms the user has "moved" the query to "an area of the database" that contains documents dealing with "statute of limitations", which is a term of art that makes the original query more focused and is likely to find documents that are relevant to the intent of the original query. Here the fact that both terms "statutes" and "statute of limitations" contain the same word is fortuitous. It is the meaning of the term "statutes" which makes it a close relative of "statute of limitations" by virtue of the fact that these two terms co-occur in many of the same documents.
Once the user is satisfied with his query he asks the program to retrieve documents that are relevant to the query. In FIG. 6 he would choose the View Documents option. The system will then use its index files to assign a value to each document and then rank the documents. The documents are ranked by generating for each document a summed-weightpoly-value which is the sum, over all terms that are in the query, of the weightpoly values of that keyword, where the weightpoly value is the product of the weight of the keyword in that document and its polysemantic weight. Then the summed-weightpoly values are adjusted using the Soft Boolean Connectors concept to come up with a final value for each document. The documents are then ranked by value and presented to the user in order of rank.
The Soft Boolean Connectors concept involves penalizing pairs of terms that co-occur often (i.e. in many documents) when calculating the adjustment to be applied to the summed relpoly percentages. First, multiply original query words by
Boost Factor (=2)
Then add WEIGHT values of key words in a document after each is multiplied by the PolyValue of the word.
______________________________________
(The previous value will be called "Temp Value")
Create table for every pair combination of query words (A B C)
AB
AC
BC
______________________________________
For each pair of query words, ("A" & "B"), the Relative Value used in the formula below is B's Relative Value in A's Relative List, or, if B doesn't appear in A's Relative List, then the value is taken from A's Relative Value in B's Relative List (this is possible because the Relative Value between any two words is mutual), i.e., if B is found in A's Relative list, take just that value. You don't need to look at B's list to find A's value there because, if it is there, it would have the same value as B has in A's list. Only if B is not in A's Relative list check for A in B's list. Enter the Relative Penalty value resulting from the following formula into the table for each combination.
______________________________________
Example:
if Relative values are
AB = 70
AC = 75
BC = 65
and A appears in 5 documents and B appears in 4
and the avg doc freq is 4
then using the following formula for AB
##STR10##
the table value for AB is 2.65
(look below for maximum, making this 1.0 instead)
for a hit of 3 words - 3 words in a document
A B C
Adjust Value = 1.8.sup.(# of hits - sum of penalties)
Note: The "# of hits" is the same as above for the SWAPS.
Word/Document Value = Temp Value .times. Adjust Value
______________________________________
MAXIMUM PENALTY TABLE (RANKING)
query words
Max.
______________________________________
(for each pair)
2 0.5
3 1.3
4 1.2
5 & up 1.1
(for sum of pairs)
2 0.5
3 1.6
4 1.9
5 2.3
6 & up 2.8
______________________________________
To facilitate very rapid perusal of the ranked documents, the document values (used in the ranking) are presented as a bar graph as shown in FIG. 7. Also the documents are presented in 3 forms. The first form consists of a ranked array of the highest ranked terms in the document that requires only about 1/3 of the display screen (FIG. 7). The second form consists of a program generated "highlight" of the document which consists of very short portions of the document of less than a dozen words that contain the highest ranked terms. This highlight scrolls in about 2/3 of the screen and is shown along with the array of highest ranked terms. The third form consists of the full text of the document which can be scrolled. The user can use arrow keys to move rapidly from one document to the next.
Appendix 1 contains the full BASIC program source code that implements the preferred embodiment described above. This code must be compiled using the Microsoft 7.1 BASIC compiler to produce object modules which must then be linked along with libraries containing object code for assembler routines from the Crescent Software QuickPak Professional Advanced Programming Library for BASIC Compilers Version 4.12 to produce an executable file.
There has thus been shown and described a novel document indexing and retrieval system which fulfills all the objects and advantages sought therefor. Many changes, modifications, variations and other uses and applications of the subject invention will, however, become apparent to those skilled in the art after considering this specification and the accompanying drawings which disclose the preferred embodiments therefor. All such changes, modifications, variations and other uses and applications which do not depart from the spirit and scope of the invention are deemed to be covered by the invention which is limited only by the claims which follow. ##SPC1##
Claims
1. A method of indexing.[.and retrieving.]. documents, said method using a digital computer system having a central processing unit, a memory,.[.a display screen, a keyboard,.]. and a large capacity file system, said method comprising the steps of:
- (a) storing in said memory a vocabulary of terms, each term consisting of one or more words, and for each term an associated term-code;
- (b) storing on said file system a collection of documents each with an associated unique document-number;
- (c) creating index files which contain for each said term-code in (a)
- (i) the set of document-numbers in (b) such that the corresponding documents contain the corresponding term; and
- (ii) for each said document-identifying-number in (i) the frequency-in-document of the corresponding term which is the number of times that said term appears in the corresponding document;
- (d) creating a weight-in-document file which contains for each document-number in (c)(i) the weight-in-document of the corresponding term which is calculated using the frequency-in-document in (c) (ii), the number of document-numbers in (c) (i), and the total number of terms in (a) which are in the corresponding document (counted multiple times);
- (e) creating a frequent-companion file which contains for each occurring term-code in (a) a ranked set of pairs of numbers where each pair consists of a first element term-code and a second element companion-percentage, where the companion-percentage is calculated by summing the weight-in-document values of said first element term-code over documents that contain both the term corresponding to said first element term-code and the term corresponding to said occurring term-code and then dividing by the sum over all documents of the weight-in-document of said occurring term-code;
- (f) creating a relative file which contains for each occurring term-code in (a) a ranked set of pairs of numbers where each pair consists of a first element relative term-code and a second element relative-percentage, where the relative-percentage is calculated by taking a weighted average of the companion-percentage of said first element term-code calculated in step (e) and the companion-percentage of said occurring term-code that was calculated in step (e) when said first element term-code was the occurring term-code and said occurring term-code was the first element term-code.[.;
- (g) creating a polysemantic file which contains for each occurring term-code in (a), a polysemantic weight which is calculated using the number of sets of pairs in the relative file created in step (f) that said occurring term-code appears in, the number of documents-numbers for which the weight-in-document of said occurring term-code calculated in step (d) is greater than some threshold value, and the averages for several values of N of the first N relative-percentages of said occurring term-code calculated and ranked in step (f);
- (h) accepting a query consisting of a sequence of words entered by a user using said keyboard and creating a parsed-query table of term-codes which consist of the term-codes in said vocabulary that are associated with the terms that are contained in said query;
- (i) creating a temporary swap table of pairs of first element term-codes and corresponding second element summed-relative-percentages consisting of those relative term-codes created in step (f) where said corresponding second element summed-relative-percentages are the sum, over all said occurring term-codes that are in said parsed-query table, of the relative percentages of said first element term-codes;
- (j) creating a modified swap table by modifying said second element summed-relative-percentages created in step (i) by multiplying them by a function of the polysemantic weight of the corresponding first element term-codes;
- (k) sorting said modified swap table by said modified summed-relative-percentages in descending order;
- (l) displaying on said display the terms corresponding to the term-codes of said modified swap table;
- (m) accepting user keypresses or other actions which identify one or more of the terms displayed in step (l) and adding the corresponding term-codes to the parsed-query-table;
- (n) repeating steps (i) through (m) as many times as the user indicates by his input;
- (o) accepting an input from the user indicating a command to retrieve documents;
- (p) creating a temporary rank table of pairs of first element document-numbers and corresponding second element summed-document-weight.times.poly values which pairs comprise those document-numbers for which any of the term-codes that are in said parsed-query table have weight-in-document above a threshold value, and summed-document-weight.times.poly values which are the sums, over all term-codes in said parsed-query table, of a function of me polysemantic weight of the term-code and the weight-in-document of the term-code;
- (r) creating a sorted rank table by sorting said temporary rank table by the value of the second elements of the pairs in descending order;
- (s) displaying on the display screen some portion of the document corresponding to the first document number in the sorted rank table and some indication of the corresponding summed-document-weight.times.poly value;
- (t) displaying other documents corresponding to other document-numbers in the sorted rank table in response to inputs from the user.]..
2. A method as in claim.[.1 wherein additional steps (j)(l) and (p)(l) are carried out after steps (j) and (p) respectively to implement the soft boolean connector algorithm which consists of.]..Iadd.8 and further comprising.Iaddend.the following steps:
- .[.(A).]. creating a table of relative penalties for each pair of said term-codes in said parsed-query table where said relative penalty is a function of the relative percentage corresponding to the two term-codes of said pair, the number of documents that each of the term-codes of the pair are contained in with a document-weight above a threshold, and the average over all terms of the number of documents that the term is contained in with a document-weight above said threshold;
- .[.(B) modifying said relative penalties by taking the minimum of the relative penalty and some maximum value which depends on the number of terms in the parsed-query table;.].
- .[.(C).]. summing said modified relative penalties to produce a sum of relative penalties;
- .[.(D) modifying said sum of relative penalties by taking the minimum of said sum and some maximum sum value which depends on the number of terms in the parsed-query table to produce a modified sum of penalties;.].
- .[.(E) summing some function of the polysemantic weights of the term-codes in the parsed-query table that are either relatives of a potential SWAPS term (jl) or are contained in a document (pl) to produce a number of hits value;.].
- .[.(F).]. Calculating some function of the number of hits value and the modified sum of penalties value to produce a power value;
- .[.(G).]. Raising a number approximately equal to 2 to the power value to produce an adjust value;
- .[.(H).]. Multiplying either the.[.modified.]. summed relative percentages.[.calculated in step j).]. or the summed document weight.[..times.poly values calculated in step (p).]. by the adjust value.
3. A method as in claim 1 where the formula for calculating the weight-in-document in step (d) is: ##EQU4##.[.
4. A method as in claim 1 where the formula for calculating the polysemantic weight in step (g) is:.]..[.5. A method as in claim 1 where
the function in step (j) is the identity function..].6. A method as in claim.[.1.]..Iadd.8.Iaddend.where the function in step.[.(p).]..Iadd.ii.Iaddend.is the identity function..Iadd.7. A method of processing a search request from a device having a keyboard and a display screen for searching a group of documents indexed according to the method of claim 1, said method of processing comprising:
- (A) accepting a query consisting of a sequence of words entered by a user using said keyboard and creating a parsed-query table of term-codes which consist of the term-codes in said vocabulary that are associated with the terms that are contained in said query;
- (B) creating a swap table of pairs of first element term-codes and corresponding second element summed-relative-percentages consisting of those relative term-codes created in step (f) where said corresponding second element summed-relative-percentages are the sum, over all said occurring term-codes that are in said parsed-query table, of the relative percentages of said first element term-codes;
- (C) sorting said swap table by said summed-relative-percentages in descending order;
- (D) displaying on said display the terms corresponding to the term-codes of said swap table;
- (E) accepting user keypresses or other actions which identify one or more of the terms displayed in step (D) and adding the corresponding term-codes to the parsed-query-table; and
- (F) repeating steps (B) through (E) as many times as the user indicates by
his input..Iaddend..Iadd.8. A method for retrieving documents identified pursuant to a query processed by the method of claim 7, said method for retrieving comprising:
- (i) accepting an input from the user indicating a command to retrieve documents;
- (ii) creating a temporary rank table of pairs of first element document-numbers and corresponding second element summed-document-weight values which pairs comprise those document-numbers for which any of the term-codes that are in said parsed-query table have weight-in-document above a threshold value, and summed-document-weight values which are the sums, over all term-codes in said parsed-query table, of a function of the term-code and the weight-in-document of the term-code;
- (iii) creating a sorted rank table by sorting said temporary rank table by the value of the second elements of the pairs in descending order;
- (iv) displaying on the display screen some portion of the document corresponding to the first document number in the sorted rank table and some indication of the corresponding summed-document-weight value;
- (v) displaying other documents corresponding to other document-numbers in the sorted rank table in response to inputs from the user..Iaddend.
| 4495566 | January 22, 1985 | Dickinson et al. |
| 4674066 | June 16, 1987 | Kucera |
| 5062074 | October 29, 1991 | Kleinberger |
| 5168565 | December 1, 1992 | Morita |
| 5263159 | November 16, 1993 | Mitsui |
| 5301109 | April 5, 1994 | Landaver et al. |
| 5307266 | April 26, 1994 | Hayashi et al. |
| 5406480 | April 11, 1995 | Kanno |
| 5873056 | February 16, 1999 | Liddy et al. |
- G. Salton, "Automatic Text Processing," Addison-Wesley, Reading, MA, 1989, pp. 229-371.
Type: Grant
Filed: Apr 4, 1997
Date of Patent: Jun 6, 2000
Inventors: Karl-Erbo G. Kageneck (Clinton Corners, NY), Ted Young (New York, NY)
Primary Examiner: Paul V. Kulik
Attorney: Whitman Breed Abbott & Morgan LLP
Application Number: 8/832,891
International Classification: G06F 1730;
