Method and Apparatus for Computer Understanding and Manipulation of Minimally Formatted Text Documents

A portion of the disclosure of this patent document contains material which is subject to copyright protection. 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 otherwise reserves all copyright rights whatsoever.

This invention relates to the field of computer scanning and analysis of text documents. More specifically, it relates to the field of computer artificial intelligence and its use in converting English text documents into computer readable form and analyzing the computer readable text.

Information retrieval from text documents has been a major area of research in computer science for the past three decades. Although a number of techniques have been developed which enable computers to understand text, none of these techniques has been wholly satisfactory.

The simplest of these techniques is `keyword matching`. Using this technique, the computer scans through an ASCII text file looking for each and every occurrence of the desired `keyword`. By using a variety of logical operations such as AND and OR, the computer can be instructed to retrieve only the combinations of keywords which the operator believes will be most relevant. Unfortunately this technique is highly literal and even a well-crafted search can miss relevant documents. For example, if a paragraph contains the sentence:

California is a leading producer of oranges,

and the user wishes to know which fruits are grown in California, searching the paragraph for "California" and "oranges" will detect the sentence and provide the correct answer. However, if the sentence read:

California is a leading consumer of oranges,

the same search will still detect this sentence, in this case erroneously.

obviously this technique is less than wholly satisfactory. Another technique uses `Natural Language Understanding`. The use of this technique combines syntactic knowledge of English with semantic information about the topic in the text being analyzed. See, for example, James Allen, Natural Language Understanding, The Benjamin/Cummings Publishing Company, Inc., Menlo Park, Calif. 1987. These techniques are limited due to their dependence on the grammatical correctness of the text and their failure to make use of the text's physical layout.

Many documents and forms are comprised of numerous unrelated text phrases, which can be ungrammatical but which are nonetheless easily understood by humans. Examples of such documents include resumes, purchase order forms, bank statements, insurance form, and so on. Human understanding of these documents entails realizing how the spatial relationship between blocks of text contributes to the meaning of the document.

Resumes provide a classic example of documents which are very easily understood by humans through the use of spatial and textual analysis. To find the name of the person, for example, the human uses the fact that it normally appears near the top of the resume, often centered or set apart from the rest of the text. Current natural language understanding systems cannot use such spatial information in detecting names.

Another example of how such information is easily used by humans but not so easily used by computers is contained in the following excerpt from a resume:

In a computer this text fragment might actually be stored as:

Worked during this period of time for a large chip 1977--manufacturer de5igning chips and wrlting 1980 simulations for new designs. I participated in the design of several different memory chips.

Typographical errors during data entry result in text which natural language systems could not understand. Even if the typographical errors were corrected, it is evident that correct understanding of this stored text would be difficult as the range of years of employment has become separated and embedded in the job description. Another problem which would result in incorrect analysis by natural language processing systems is the fact that the first sentence is grammatically not a complete sentence.

It is an object of this invention to provide a method and apparatus for use with a computer which analyzes text documents using both sophisticated text pattern matching techniques which are insensitive to typographical errors and to the ungrammatical nature of text fragments, and spatial analysis techniques for analyzing the spatial structure of the text document, the method providing greatly improved computer textual analysis.

This object and others are fulfilled by the present invention wherein a document is first converted into a digital image form by means of an optical scanner. Once the document has been converted into a digital image, the digital image is analyzed to determine areas of contiguous text, contiguous being measured across both the document's vertical and horizontal axes. These areas of contiguous text are then further processed by a character recognition unit to generate an ASCII text form of the document. The contents of the ASCII form are analyzed using a tree-like "grammar" structure. The result of this analysis can be used to store the text in the computer's memory in a pre-selected format which the user believes will enhance data readability and useability.

These objects and other objects and advantages of the invention will appear more clearly from the following specification in connection with the accompanying drawings, in which:

FIG. 1 is a block diagram of the physical components of the present invention;

FIG. 2 is a flow chart of the process used by the present invention to analyze documents;

FIG. 3 is a flow chart of the method used to decompose the text into contiguous text blocks;

FIG. 4 is a flow chart detailing the process of selecting the proper, line spacing;

FIG. 5 is a sample text document;

FIGS. 6A and 6B show the output from step 4.1 in FIG. 4;

FIGS. 7A and 7B show the output from step 4.3 in FIG. 4;

FIG. 8 shows the output from step 4.4 in FIG. 4;

FIG. 9 is a flow chart of the process used to separate the text document into horizontal blocks;

FIG. 9A is a second sample text document;

FIG. 9B shows the horizontal blocks which make up the sample text document in FIG. 9A;

FIGS. 10A, 10B, and 10c show the output from step 9.3 of FIG. 9.

FIG. 11 is a flow chart of the process used to separate the vertical text blocks within the horizontal text blocks;

FIG. 11A shows a horizontal text block with multiple vertical text blocks;

FIG. 11B shows the same horizontal text as in FIG. 11A's horizontal block, broken up into individual vertical text blocks;

FIG. 12 shows the output after the initial blocking operation is completed;

FIG. 12A is a third sample text document;

FIG. 13 shows the output of the blocking method after the blocking/process is repeated;

FIG. 14 is a flow chart of the process used to classify, sort and convert the digital block into an ASCII block;

FIG. 15 shows the output of the blocking method after step 14.3 in FIG. 14;

FIG. 16 shows a text sample before and after the text blocking procedure;

FIG. 17 is a flow chart of the method used to sort the blocks;

FIGS. 18A and 18B are a printout of the ASCII text file after spatial analysis has been completed on the sample resume of FIG. 5;

FIG. 19 shows the components of the text analysis process and their relationship to one another;

FIGS. 20A, 20B and 20C form a sample grammar used to recognize dates in text documents;

FIG. 21 is a sample of the templates used to output the grammar;

FIGS. 22A, 22B and 22C are listings of the output from the extractor;

FIG. 23 is a flow chart of the text analysis process;

FIG. 24 is a block diagram of how the grammar is loaded into the host computer;

FIG. 25 illustrates how the text is initially scanned and analyzed;

FIG. 26 is a flow chart of the process used to match pattern objects of the grammar against the text;

FIG. 27 is a flow chart of the process used to match a specific pattern; and

FIG. 28 is a flow chart of the process used to fill in the templates used to output the data.

FIG. 1 illustrates the preferred embodiment of this invention. The apparatus comprises host computer 1, scanner 2, and character recognition unit 3. In this embodiment the document to be processed is digitized by the scanner and stored in memory 4, which may comprise either internal or external memory devices. Processing takes place on the host computer using character recognition unit 3.

The host computer 1 is a programmable digital computer having a bit map display and incorporating a keyboard and mouse for user input. A `Sun 3/60` workstation from Sun Microsystems, Inc. of Mountain View, CA has been used as the host computer. Character recognition unit 3 is an optical character recognition (OCR) processor. OCR processors translate digitized images into ASCII computer text files. Scanner 2 converts paper documents into digitized images for storage in computer main and secondary memory. The OCR and scanner may be contained in the same unit. For example, the `CDP-9000` from Calera Recognition Systems of Santa Clara, Calif. can be used.

The digitized document image is processed by host computer 1 using the present invention. The digitized image is represented as a matrix of picture elements (pixels), 300 pixels per inch. A pixel has a value 0 if it represents a white dot in the digitized image and 1 if it represents a black dot. The computer representation of an image of a single 8.5.times.11 inch page therefore requires approximately one megabyte of memory (packing eight pixels per byte).

FIG. 2 shows the operation of the present invention in broadest outline. There are three main stages of operation: spatial analysis of the text 2.1, composition of the document text 2.2, and text analysis 2.3. In the preferred embodiment these stages are each executed on the host computer.

In the present invention, many of the steps use one or more predefined constants. These constants are listed below in Table 1.

These constants were defined by a preliminary examination of numerous documents which might be processed by this invention. In some cases, values were initially chosen based on well-known printing conventions for documents. For example, many documents have a 1" margin on the left and right side of the paper. The initial values were further refined during testing. It should be understood that the exact value of these constants is not the essence of this invention. Other values could be used which might optimize the operation of this invention on a particular group of documents. The constants in Table 1 are therefore to be seen as examples only.

Spatial Analysis

The spatial analysis of the text takes the digitized image generated by the scanner and decomposes it into blocks of contiguous text. The decomposition process is illustrated in FIG. 3. In order to correctly identify blocks of related textual information the minimum spacing between lines (the line spacing) is first determined at step 3.1. In step 3.2 the image is partitioned into horizontal text blocks and, in step 3.3 the horizontal text blocks are partitioned into vertical blocks. This sequence of partitioning steps is performed twice (step 3.4).

The method used to determine the line spacing is shown in FIG. 4. First, and not shown in the figure, the image of the text document is down-sampled to 75 dots per inch (this reduces the amount of memory required during spatial analysis) and stored in row major order. Unless otherwise noted, constants used in the blocking process are chosen to be consistent with the 75 dots per inch density. Additionally, in this embodiment, pixels are operated upon as bytes. As a byte is 8 bits long and as each pixel is a bit, the down-sampled text is treated as being in vertical columns of 8 bits/pixels width.

The borders of the text document are ignored during blocking operations. A strip of a given width comprised of a preset number of pixel rows, the preset number being called TOP.sub.-- MARGIN, is ignored at the top of the text. Similarly, a row of BOTTOM.sub.-- MARGIN pixels in width is ignored at the bottom of the image, and columns of LEFT.sub.-- MARGIN*8 and RIGHT.sub.-- MARGIN*8 are ignored on the left and right side of the document. The size of these constants is listed in Table 1.

After the edges of the document have been ignored, the first step of the process, forming a histogram of the pixel values of the text's columns is performed (see step 4.1). In this step, the total number of pixel elements of value 1 in each vertical column of the document is summed. The column containing the most such elements is used for further processing. The results of performing this step upon the sample resume shown in FIG. 5 are given in FIGS. 6A and 6B. As stated, in creating the histogram of FIGS. 6A and 6B, the margins of the sample text document in FIG. 5 are ignored and the pixels of value 1 in each 8 bit column of the text are counted. Column 16 contains 936 such pixels and is determined in step 4.2 to have the highest total number of pixels (see line 0014 16 in FIG. 6A), thereby becoming the margin column. A band of MARGIN.sub.-- WIDTH columns, the leftmost column being the margin column, is now used to calculate the line spacing. As the MARGIN.sub.-- WIDTH variable equals 8 in this embodiment, a column of 64 bits in width is used for calculating the line spacing.

In step 4.3 of FIG. 4 the pixels of value 1 in each row of the band at margin column are summed to determine if there is text in the margin column. The results of this processing are shown in FIGS. 7A and 7B. Consecutive values equal to or greater than a preset threshold variable, called the MIN.sub.-- LINE THRESHOLD, indicate the presence of text.

In step 4.4 of FIG. 4 the distance from consecutive points where the number of pixels of value 1 went from below the MIN.sub.-- LINE.sub.-- THRESHOLD value to above that value is summed and a histogram of these values is formed. In other words, a count or measure of the distance from a first occurrence of text through the end of that first occurrence to the start of a second occurrence of text is taken. This count or measure is taken through the entire text document. These measures produce the histogram shown in FIG. 8, where column B is the distance from successive starting points of text and column C is the number of times that distance was found. The value for line spacing is selected in step 4.5 by choosing the most common value returned from step 4.4. This value must lie between the constants MIN.sub.-- LINE.sub.-- SPACE and a MAX.sub.-- LINE.sub.-- SPACE. These values are determined by the process discussed earlier. As shown in FIG. 8, the line space is set at 13 when the sample resume of FIG. 5 is used. It should be noted that the row histogram given in FIGS. 7A and 7B is not a complete histogram of the resume shown in FIG. 5. Consequently, all 13 occurrences of a line spacing of 13 are not shown.

FIG. 9 illustrates the method used to determine horizontal text blocks (see FIG. 3, step 3.2), i.e. those blocks of text which are vertically separated from one another by a space larger than a minimum spacing dependant on the line spacing. For example, the method of FIG. 9 will separate the document shown in FIG. 9A into the horizontal text blocks shown in FIG. 9B. Step 9.1 of FIG. 9 comprises summing the number of pixels of value 1 in each row of the image. In step 9.2 the values determined in step 9.1 are examined. Values above a certain minimum, denoted the MIN.sub.-- HORIZ.sub.-- LINE.sub.-- THRESHOLD, indicate the presence of text. Therefore, a transition from a row whose value is below the minimum to one whose value is above the minimum denotes the start of a text block. The block continues until a number of rows or a space without text larger than the HORIZ.sub.-- BLOCK.sub.-- FACTOR 5/8 is detected. This space is equal to 1.625 times the line space. The detection of a space of this width indicates the end of a text block. The first and last rows in the text document are also considered to be delimiters. The determination of the horizontal text blocks in the text sample shown in FIG. 5 is indicated by the output shown in FIG. 10. This is accomplished in step 9.3 of FIG. 9.

The process used to determine the vertical text blocks (see step 3.3 in FIG. 3) is illustrated in FIG. 11. This process divides each horizontal block, as determined in step 9.3 of FIG. 9, into separate vertical sub-blocks, where appropriate. As an example, the process of FIG. 11 takes the horizontal text block shown in FIG. 11A and resolves it into three horizontally separated blocks, as shown in FIG. 11B. In step 11.1 the number of pixels with value 1 are summed within each column (a column here being one byte in width) in the horizontal block. The sums indicate the presence or absence of text within a given vertical column. In step 11.2 the values of each column are compared to a preset value called the MIN.sub.-- COLUMN.sub.-- THRESHOLD. Columns which fall below this value are considered to have no text. In step 11.3, the vertical blocks are separated from each other, each block being considered to start from a first transition where the text in the vertical column exceeds the MIN.sub.-- COLUMN.sub.-- THRESHOLD to a point where there have been at least VERTICAL.sub.-- BLOCK.sub.-- FACTOR consecutive columns whose histogram value is below MIN.sub.-- COLUMN.sub.-- THRESHOLD. This ensures that the sides of each such group of columns are buffered by an appropriate amount of space. It is assumed that the first and last column delimit the start of the first vertical block and the end of the last vertical block.

At this point the image has been partitioned into a collection of text blocks. FIG. 12 shows the blocks which exist after one pass though the horizontal and vertical blocking procedures. It is possible that each of these blocks themselves has an internal block structure. A sample of such a document is shown in FIG. 12A. The original text is seen in FIG. 12A.1. As there is text on each line of this block, the horizontal blocking process only created one large horizontal block. After the first vertical blocking, the separation of blocks shown in FIG. 12A.2 is obtained. A problem still clearly exists in that spatially separated text in the left side block is still treated as a single block. The repetition of the blocking process catches this mistake and returns the properly blocked text shown in FIG. 12A.3. Therefore, the horizontal and vertical blocking process is repeated on each of the blocks shown in FIG. 12. This repetition is indicated by step 3.4 in FIG. 3. Once this is done the document image has been decomposed into blocks of contiguous regions of text. This is the output of the spatial analysis portion of this invention. For the document shown in FIG. 5, after the output shown in FIG. 12 passes through the blocking method a second time, the output shown in FIG. 13 is obtained. In both FIG. 12 and FIG. 13 the four variables related to each block are the X and Y location where the block begins on the document and the block's height and width. Note that the units used in FIG. 12 for the X location and width are in bytes whereas in FIG. 13 the units are in pixels.

Composition of The Document

After the spatial analysis of the document is completed, the document's composition is analyzed (step 2.2 in FIG. 2).

As shown in FIG. 14, the text blocks are classified, translated into an ASCII text file by the optical character reader, and sorted. First, blocks are obtained from memory (step 14.1). Each block is then classified in step 14.2. The classification of the text blocks is based upon the spatial location of each text block and results in the text block being classified as either a header or body block. What characterizes a header block is customized as to the particular type of document being analyzed. Generally, the classification will depend upon coordinate information derived from the blocking process. FIG. 13 shows how the blocks are stored as coordinates. In the example of the text document shown in FIG. 5, a header block is defined as either a block whose upper-left corner is within a preset distance DELTA from the left margin, or a block which is centered on the page and whose width is less than the result of the page width minus a preset variable whose value is CENTER.sub.-- WIDTH.

In step 14.3 in FIG. 14 the relationship of each block to the edge of the image is checked. If the block is within the preset distance DELTA from the right or left margin of the document, its edge is extended to the edge of the image. This contingency deals with the situation where the text in a text block actually extends into the margin area originally ignored when spatial analysis began. By extending margins into this space, the pixels of text that might have been lost are recovered. FIG. 15 lists the blocks after the left and right boundaries have been reblocked, i.e., extended to the image edges.

After being reblocked in step 14.3, each block, or more specifically, the region of the digitized image within the blocks' boundary is passed to the character recognition unit to translate the image into an ASCII text file. This occurs at step 14.4. Upon completion of this step each block has two components: 1) coordinates indicating where on the document page the block is physically located and 2) the text segment in the block, represented in the computer as an ASCII string.

In step 14.5, each individual vertical subblock is partitioned into vertically separated sub-blocks of text. This step is called text blocking. This text blocking step, which may appear to be duplicative of the previously described spatial analysis, is necessitated by the passing of the text blocks through the character recognition unit. It is possible that stray marks or other disfigurations of the document being analyzed result in pixels of value 1 in locations where no text actually exists. These pixels might result in the spatial analysis being unable to detect a space between blocks where such stray pixels are found. However, after these same pixels pass through the character recognition unit, they are not recognized as proper text and these stray pixels are ignored. Now, by performing a final text blocking step, the blocking errors caused by this type of problem can be eliminated. Also, the vertical blocking is now based on a minimum number of characters separating the blocks, not a certain number of pixels. Although the sensitivity of these measures is comparable, different errors are corrected by using the different measures. The separators between the sub-blocks must consist of at least a preset number of consecutive columns, each column one character wide, called WHITE.sub.-- WIDTH. The space between columns must contain only blanks. These sub-blocks are reformed into a single block by reading the sub-blocks sequentially from left to right. A sample of text shown before and after the text blocking is shown in FIG. 16. In step 14.6 the program loops back to step 14.1 and repeats until the classification and reblocking on all text blocks has occurred.

After all blocks have been classified as header or body blocks and, if necessary, text blocked, the texts of the blocks are reassembled by sorting the blocks at step 14.7. FIG. 17 shows the sort process in detail. The sort uses an ordering function which results in a list of text blocks such that the text contents of the blocks, when read sequentially, have the same visual order as a person would naturally want to see when scanning the original document. This ordering further results in blocks of related text being located in proper relational proximity to one another.

As shown in FIG. 17, the tops (measured by the Y position of the first row of text in the block) of two of the blocks (here A & B) are first compared to one another (step 17.1). If the tops of A and B are equal, the first columns of both blocks are compared (step 17.9). If the columns are equal, the process returns the value that blocks A and B are equal (step 17.10). If block A is to the left of block B, then the value that A is less than B is returned (step 17.11). Finally, if block A is to the right of B, then A is greater than B (step 17.12).

If the first rows are not equal, another series of tests begins. At step 17.2, if the bottom of block A is above the top of block B, then A is less than B (step 17.11). If the bottom of A is not above the top of B, the next test is if the bottom of B is above the top of A (step 17.3) if it is, then A is greater than B (step 17.12). If not, step 17.4 determines if A is to the left of B.

If A is left of B, then another test is made. If the first row of A is below the first row of B (step 17.5) and the bottom of A is below the bottom of B (step 17.6), then A is greater than B (step 17.12). The alternative result is that A is less than B (step 17.11).

If A is not to the left of B, then two other conditions are checked. If the first row of B is below the first row of A (step 17.7) and the bottom of A is above the bottom of B (step 17.8), then A is less than B. Otherwise A is greater than B (step 17.12).

In step 14.8 the contents of the text blocks are delivered to the text analysis process as a linear stream of characters using the ordering derived from step 14.7. Those blocks characterized as header blocks are padded with a ` ` character. Other indicators generated include line indicators and beginning and end of document marks. FIG. 18 shows a text sample after the spatial text analysis process has been completed.

Although this description discusses the processes of spatial analysis, composition of document, and text analysis separately, these processes are all codependant and the proper functioning of the entire invention depends upon their inter-relationship. Text analysis depends upon the document being correctly resolved into its component text blocks and the correct determination of the blocks' composition. The separate discussion of these processes simplifies understanding the invention, but it should not be concluded that the processes or modules are separate entities.

As shown in FIG. 2, once the text has been blocked in step 2.1 and translated into an ASCII text file in step 2.2, textual analysis, step 2.3, begins. The main components used to analyze the text and their relationship with one another are shown in FIG. 19. Blocked ASCII text strings 19.1 are the input for the textual analysis process. These strings can be denoted with special characters, such as a ` `, to indicate the type of string. The blocks so denoted carry special meaning because of their position on the documents. During textual analysis, predefined text patterns in strings 19.1 are located by the extraction process 19.2. A module called the extractor executes this pattern matching. The extractor finds all occurrences or instances of the predefined text patterns in the document.

In the preferred embodiment, the extractor module can be used on a `Sun 3/60` workstation and is realized by the C program attached as Appendix 1.

In addition to a properly blocked text, the extractor requires a knowledge base of relevant text patterns to perform text analysis properly. These patterns are described in a pattern definition language. A knowledge base written in this language is known as grammar 19.3. The extractor uses the grammar to extract text patterns of interest. A sample grammar, written in the pattern definition language is attached as Appendix 2.

The grammar used in the present invention is created prior to any specific text analysis and is unchanged by the text analysis. The grammar comprises pattern objects formed into synonyms and classes. The creation of the grammar is determined by the specific types of information whose referral or processing is desired.

The text patterns defined by the grammar are chosen by their relevance to the type of document being analyzed. In the case of resumes, patterns for job titles, educational degrees, job skills, etc. are defined in the grammar. In order to correctly identify such patterns, it is necessary to create patterns utilizing the ` ` character (inserted during the spatial analysis) which identifies header blocks. Header block patterns defined in a resume grammar include an education header, a work history header, an objective header, etc.

The pattern definition language or more specifically, the grammar, uses synonyms 19.4 and classes 19.5 to define key textual patterns for the extractor to locate and uses templates 19.6 to format and store the patterns located by the extractor. A synonym is a collection of patterns used for the same word. For example, "January", "Jan", "Jan.", "1", or "01", in the proper context, are all synonyms for January. A class is a collection of patterns, each pattern potentially having many synonyms, which collectively relate to the same type of object. For example, "Jan. 12, 1977", "Oct. 27, 1959", "12-01-57" are all in the class called `DATES`. The difference between classes and synonyms appears when they are matched as patterns.

The different patterns (see the previous example for January) which make up a synonym group are all treated as having the same value, regardless of which of the patterns occurs in the text. Classes, on the other hand, even when they contain only a single pattern, have a different value for each instance the pattern is detected in the text. For example, the class of "AGE" is comprised of only a single pattern, that being a single number, each pattern object within the preset range for the class "AGE" having a value equal to the actual number found in the text. When a pattern of a class is comprised of a plurality of elements, the value of the pattern is the concatenation of the values of the individual elements. It is impotant to note that a synonym, which is a pattern object, is formed from a plurality of patterns. Likewise, a class is a pattern object, the individual members which comprise the class being patterns.

Patterns consist of elements which can be numbers, words, or previously defined objects, separated by operators. The following operators are recognized by the extractor:

[I]=ignore;

[N]=not;

[n]=within n words of, where n is equal to 0 or a positive integer;

[W]=within;

[C]=containing;

[0]=optional;

[B]=before;

[F]=first; and

!:n spelling corrected by n letters, where n is 0 or a positive integer.

Thus "/CITY [4]/STATE [W]/FIRST.sub.-- TEN.sub.-- LINES" might be a pattern which is matched if a city is within four words of a state, and the city/state combination is within the first ten lines. If no operator is specified, the default, [1], is assumed.

FIGS. 20A, 20B, and 20C show the portion of a grammar used to define and recognize dates in a document. The synonym "Jan", defined on line 0006 of FIG. 20A, is a synonym for the month of January. There are four patterns which match the synonym "Jan". Each represents how "January" would appear in the text of a document. The second pattern "Jan [0].", contains two elements, "Jan" and ".", and one operator, "[0]". The "[0]" indicates that "Jan" may be optionally followed by a period. The class "MONTH", defined on line 0075, contains as patterns all of the synonyms for the twelve months.

Operators can be used to overcome typographical errors introduced during the character recognition phase. In the example above, if the period following "Jan" is missing after the character recognition step, the optional operator in the pattern for January ensures that this occurrence is still recognized as an instance of the month January.

Patterns make use of the formatting information left behind by the text blocking process by referring to lines, headers, and the beginning and end of document marks. This feature allows the authors of various grammars to incorporate spatial clues or information in pattern objects, thereby aiding computer understanding of documents in which text data location contributes to the document meaning.

For each pattern object an action can be specified. An action indicates what is to be done when instances of the pattern object are found in the text document. This allows text analysis of pattern objects as they are located in the document. For example, in FIG. 20B, line 0094, the class "YEAR" has action `Years` associated with it. The action is used to ensure that only numbers in a certain range are regarded as years. This allows the extractor to rule out 7893 as a year when it occurs in a document. Thus, actions perform such functions as range checking, counting occurrences of pattern objects, data output or other similar actions.

Templates are data structures that specify how the extractor should format the output. A template comprises a first pattern object which gives the template its name and a collection of slots. Each slot in turn has its own pattern object associated with it. A pattern object associated with a template slot must be a component (i.e. an element of, or an element of an element of, or an element of an element of an element of, etc.) of the first pattern object of the template. Each instance of the first pattern object of a template in the input document results in the filling in of the template's slots. The value used by the extractor to fill in each slot is the corresponding instance of the slot's pattern object. FIG. 21 illustrates a group of template definitions. For example, on line 0009 of FIG. 21, the template pattern object APPLICANT.sub.-- NAME causes all instances of the `Applicant` slot pattern object `APPLICANT` to be extracted and labeled with the slot name, `Applicant`. On line 0012, the template pattern object `APP.sub.-- ADDRESS` requires several pattern objects to fill in the template slots.

Once the grammar, with its synonyms 19.4, classes 19.5 and templates 19.6 is created, it may be stored in any known computer memory. When this invention is to be used, the grammar is loaded into the host computer (see step 23.1, FIG. 23). It should be understood that the loading of the grammar and the use of the extractor to find text patterns defined by the grammar do not in any way alter the grammar itself. Conceptually it may be easiest to view the extractor as creating a partial mirror image of the stored grammar, the practical image containing only those text patterns detected in the incoming ASCII text stream 19.1.

In other words, after the text blocking and composition processes occur, the resulting ASCII data streams 19.1 are scanned and those elements therein which match the synonyms 19.4 and classes 19.5 of the grammar 19.3 are used to fill out the grammar's templates 19.6. This filling out process does not alter the original grammar in any way. It does create a new text file comprised of only the information which was defined by the grammar and detected by the extractor. It is this file which is used to fill the templates 19.6 and to generate output. This output file can be altered by the user. The original grammar is untouched and available for processing other new documents.

The output (19.7 in FIG. 19) from the extractor can be used in many ways. For example, it can be saved in a computer file, displayed on an output device such as a cathode ray screen, printed, or placed into a database for future retrieval. Templates are defined depending on where and how the output will be used. An example of the output of the extractor is shown in FIGS. 22A, 22B, and 22C. In this example all class and synonym instances are printed. Lines 0149-0151 of FIG. 22C shows the output of template "YEARS", slot `Dates` from FIG. 21, line 0032, and the pattern object POS.sub.-- TITLE.

A high-level flow chart for the text analysis process is shown in FIG. 23. The grammar is loaded into the computer at step 23.1. In step 23.2 the document text is scanned, identifying key words from the grammar. The pattern objects specified by the grammar are matched with the words identified during the document scan in step 23.3. The templates are filled in using pattern objects found during the scan in step 23.4.

FIG. 24 illustrates the process of loading the grammar into the extractor (see step 23.1, FIG. 23). The primary effects of loading the grammar are to parse and analyze the grammar, and to build a hierarchical symbol table of patterns objects. The data structure for a pattern object in the symbol table is shown below in Table 2.

The grammar file is opened at step 24.1. In step 24.2, a line from the grammar file is read sequentially. Depending on the type of line, the appropriate action is taken in step 24.3. If a new class or synonym is read, it is added to the symbol table, at step 24.4 or 24.5, respectively. Each of these steps creates a new entry into the symbol table.

If a pattern is read, it is parsed at step 24.6. Patterns are made of numbers, words, or previously defined objects. In step 24.6, patterns are represented in the computer as lists of pairs, an element and its operator. The list of the pairs is stored in an array of pair list descriptors using the data structure shown in Table 3.

The actual pair is stored in the date structure shown in Table 4.

At step 24.7, if an action is read, this action is associated with the last class or synonym defined. Step 24.8 shows the entry of a new template definition. Each template is stored in a structure which is used to generate the output of the extractor. The data structure used for templates is shown in Table 5.

At step 24.9 slot definitions are parsed and inserted into the template structure.

At step 24.10, pattern object definitions store in other files can be added to the grammar. This allows pattern object definition files to be modularized. Step 24.11 tells the process to ignore comments placed in the grammar by the author. Any other line is unrecognized and hence an error. Step 24.12 indicates the handling of these errors.

In step 24.13, the process loops back to read the next input grammar line until the end of the grammar is reached.

The process of scanning a document, shown in step 23.2 of FIG. 23, is detailed in FIG. 25. The document is sequentially read into memory at step 25.1. During the reading, lines containing no alphabetic or numeric characters are ignored. Leading blank space on each line is removed and consecutive blanks are compressed to a single blank space. Once in memory, the document is scanned sequentially, one word at a time.

When a pattern defined by the grammar is found in the document text, it is recorded as an instance. The simplest instance is a single word or number. More complex instances arise from pattern objects whose patterns contain several elements separated by operators. Such instances are stored in the data structure shown in Table 6.

When the scanning process of FIG. 25 detects a number (step 25.2), a number instance is added into the set of matched patterns (step 25.3). At step 25.4, if a word is detected, the process checks to see if the word occurs as an element of or as a spelling corrected word of some pattern object in the grammar. If it does, an instance for that particular grammar element is added at step 25.5. A function "AddEntity" adds new instances by creating a new ENTITY structure. At step 25.6 a check is made to see if the end of document has been reached. If it has, the scanning phase is complete. If not the next word is processed.

After scanning the document for simple word and number instances, the more complex patterns of the document are matched and assembled. FIG. 26 illustrates this process. Each pattern object defined in the grammar is examined to see if any occurrences of it appear in the document. A first pattern object is loaded at step 26.1. Step 26.2 retrieves the pattern object phrases (patterns) in the current pattern object sequentially and one at a time. A search for instances of these pattern object phrases occurs at step 26.3. All instances are added by using a function "AddEntity". This process is repeated for each pattern object phrase in the current pattern object by looping at step 26.4. After each pattern has been examined any action associated with the pattern object is executed at step 26.5. This procedure is repeated for each pattern object in the grammar. See step 26.6.

Referring back to FIGS. 22A, 22B, and 22C, an example of how the process shown in FIG. 26 is performed can be seen.

FIG. 27 illustrates the major steps in the pattern matching process. The essence of this process is to build up instances of pattern objects by sequentially examining each element in the pattern. Each instance of an element is adjoined to previously constructed partial pattern instances, i.e. instances of a subpattern consisting of the elements of the pattern up to but not including the pattern element whose instances are currently being considered. Thus, for example, for the pattern

/MONTH /DAY /YEAR

the process first finds all instances of the first subpattern "/MONTH". Instances of the next subpattern, "/MONTH /DAY", are determined by first examining all instances of the first subpattern (months), and then creating a new instance if the month subpattern is followed by a day. Finally, instances of the last subpattern, "/MONTH /DAY /YEAR", are determined by examining all instances of the second subpattern (month day) and creating a new instance if the month-day subpattern is followed by a year.

At each step this process constructs all instances of the subpattern based upon what has previously been constructed. This is done for each pattern element and, as the last subpattern is the pattern itself, every instance of the pattern occurring in the document will have been found.

The steps used to create partial pattern instances of a given pattern are shown in FIG. 27. In step 27.1, a first pattern element is retrieved from the pattern. Instances of this element in the document are obtained from the extractor in step 27.2. As the pattern element is either a previously defined pattern or a simple word or number, all instances of it (if they exist) are already known. If none exist but the element operator requires instances, then the pattern has no instances (see step 27.6). If the operator associated with the element does not require instances of the element (for example when the operator is "optional"), a null instance is added at step 27.4. This process is repeated (see step 27.5) for each pattern element in the pattern object.

A first pattern element of the pattern is obtained at step 27.7. Instances of this element in the document are found at step 27.8. If the instance is null, the process of creating new partial pattern instances is skipped (step 27.9) and the process looks for the next element instance. Otherwise new partial pattern instances are created for all currently existing partial pattern instances, starting at step 27.8. The element operator and the instance returned in step 27.8 are checked to see if they are consistent with the partial pattern also retrieved at step 27.10 in step 27.11. If they are, the element instance is appended to the partial pattern instance to create a new pattern instance (see step 27.12). If not, the process begun at step 27.10 will be repeated for all previously existing partial pattern instances, if any (see step 27.13).

At Step 27.14, a check is made to see if there are more instances of the current pattern element. If there are, then the process returns to step 27.8 and repeats, this time with the next instance of the pattern element. If not, if there are more pattern elements in the pattern, the process returns to step 27.7 to work on the next pattern element (see step 27.15). If not, all partial pattern instances which have been completed are returned as valid matching patterns at step 27.16. This will be all instances of the pattern which occur in the document.

FIG. 28 illustrates the process used to fill templates. A first template from the grammar is retrieved at step 28.1. An instance of the template pattern is retrieved at step 28.2. The corresponding instance of the next slot of this template is retrieved by searching the children of the instance of the template pattern for the instance corresponding to the slot's pattern object. This occurs at step 28.3. This value is assigned to the slot in step 28.4. The program loops between step 28.3 and 28.5 until all template slots have been filled. The process loops between steps 28.2 and 28.6 until all template pattern instances have been operated upon. This process is repeated for each template in the grammar by looping between step 28.1 and step 28.7. Finally, at step 28.8 the results of the template filling process are supplied to memory or any other output form. This output can be used by other programs, displayed on a computer terminal, or stored into a database for future use.

In summary, the present invention accepts as input text documents which can be either grammatical or non-grammatical. The invention disassembles the document into separate blocks of contiguous text but preserves the information that the location of these blocks provides and maintains the blocks in ordered relationship to one another. The text blocks are then converted into ASCII text files. Using a grammar (a type of pattern definition language), the extractor scans each ASCII file and retrieves desired word patterns. The patterns are formatted by using a part of the grammar called templates. Finally, the output from the templates may be printed or stored. In this manner, a unique method and apparatus for analyzing non-grammatical text documents is realized.

In the foregoing specification, the invention has been described with reference to a specific exemplary embodiment thereof. It will, however, be evident that various modifications and changes may be made thereunto without departing from the broader spirit and scope of the invention as set forth in the appended claims. For example, the constants in Table 1 could be changed for optimum performance with different document types. The grammar could also be modified to retrieve more or different types of information. Many such changes or modifications are readily envisioned. The specification and drawings are, accordingly, to be regarded in an illustrative rather than in a restrictive sense. ##SPC1##