Method and Apparatus for Utilizing Annotations to Facilitate Computer Retrieval of Database Material

This invention relates to systems for retrieval of computer stored database material and more particularly to a method and apparatus for annotating stored text, computer programs, computer graphics, audio, video and other material, either manually, semiautomatically or automatically and for utilizing the generated annotations to retrieve relevant material.

There exist at this time huge quantities of material stored in various computer databases and this quantity of information is growing each year at an astonishing rate. Further, not only is the quantity of material increasing, but the types of material which can be digitized for storage in a computer database is also increasing each year. For example, technology has made it possible not only to store material represented in alphanumeric form, such as text, computer programs, object classes, action specifications and related data, but it is now possible to store all types of graphic materials including pictures, photographs, drawings, charts, video images and the like as well as audio recordings including voice images and prints and all types of music.

However, having such large quantities of material stored is of little value unless it is possible to retrieve a desired item of the material or to determine if such item is available in the database. Unfortunately, while current techniques for accessing such diverse databases are moderately adequate for accessing certain types of textual databases, such techniques are virtually non-existent for accessing other types of textual databases and for accessing graphic, audio or other specialized databases. Consequently, there are large bodies of computer-stored material which are either difficult or virtually impossible to access and are, therefore, either not utilized at all or significantly under utilized.

In particular, the standard technique for accessing free-form texts is by means of a "key word" combination, generally with boolean

connections between the words or terms. An "inverted" file or key word index may be provided to reduce the time required for such search.

Such key word searching techniques have a number of drawbacks. First, such systems can only be used for text and are of no use for other types of material. Second, such systems are not adapted to accept a search request in the form which would normally be posed by a user. Instead, the user must be able to determine the terms and the boolean connections which will yield the desired information. Such a procedure requires a certain level of sophistication on the part of the user. Even with a sophisticated and experienced user, because of the vagaries of the English or other natural language utilized, and for other reasons, a search is frequently drawn too broadly so as to yield far more "hits" than the user wishes to review, or is drawn too narrowly so that relevant items are missed. In many instances, a search strategy is evolved by trial and error so that an acceptable number of relevant hits are obtained. While some front-end programs are available for assisting a user in developing a search strategy, the use of such front-end programs is also time-consuming and, depending on the skill of the user, may still not result, at least initially, in a proper search strategy being evolved.

A third problem with key word searches is that they in fact require that every word of the text be looked at during a search. Since textual databases such as those containing legal cases, scientific or medical papers, patents and the like may contain millions of words of text, a full key word search can often be a time-consuming and therefore an expensive procedure. This problem is generally dealt with through use of inverted file indexes which lead a user to an area containing a particular key word so that the entire body of text need not be searched. However, inverted file arrangements are not appropriate in all situations. The basic problem is the size of the inverted index, which can equal, or even exceed, the size of the main document file. Further, in order to maintain the usefulness of the inverted file, it is necessary that the large inverted file of key words be updated each time the main database is updated. For these and other reasons, inverted files are generally appropriate only if:

(1) the vocabulary of the text or database is homogeneous and standardized, as in medicine and law, and relatively free of spelling errors

(2) contiguous word and word proximity processing (where the location of the search words in the text is important) is not necessary and, hence, extensive location information is not required (i.e. extensive boolean connections between words are not critical in the key word search process);

(3) a limited class of text words (rather than the full text) are used for search purposes so that only these terms need appear in the inverted index;

(4) the database is not too large.

When the criteria indicated above are not met, an inverted file may be of only limited use in reducing the time required for a given search.

Thus, since the quantity of text is ever-increasing, it becomes progressively more difficult to rapidly locate text relevant to a particular question, assertion or the like; and it becomes virtually impossible to locate all text relevant to such an item. However, particularly in such areas as law and patents, failure to locate a relevant precedent or patent can have catastrophic results. Complete searching for graphics, such as a trademark logo for clearance purposes or a particular picture, for audio material or for other non-textual material, is virtually impossible. The result is that many functions which should or could be handled automatically by a computer are now being performed manually. For example, the "help-desk" function wherein users of for example a particular product or service may contact a particular E-mail address or telephone number to obtain information or assistance in dealing with various problems could be automated to provide textual, audio or graphic information in response to large classes of queries if stored material relevant to the queries could be easily located. However, the inability of computers to permit information to be retrieved in response to queries from large numbers of potential users, most of whom are unsophisticated in computer techniques, has resulted in the help-desk function normally being performed manually. Since queries frequently arrive in bunches, this can result in disgruntled users having to wait unacceptably long times to receive responses while some of the people performing the function may be under-utilized during other periods. It also means that a user has to rely on the expertise of the person providing the response where, particularly for sophisticated and fast changing technologies, the person manning the help-desk may not always be fully knowledgeable.

Therefore, a need exists for an improved method and apparatus for retrieving relevant material from large databases, and in particular for permitting such retrieval to be accomplished by a relatively unsophisticated user. Preferably, the user should merely be required to present a query in English (or other natural language) in a way which is normal for the user, with little or no preconditions attached.

It should also be possible to complete searches on all types of text, graphics, audio and other stored material and to complete the search expeditiously. To improve the response time on database searches, it is preferable that it not be necessary to refer to each word or other unit of the material during a retrieval operation and it is also desirable that, once relevant material is located, it be possible to locate other relevant material from the initially-located relevant material without necessarily requiring further searching. In particular, a match for a section of text should permit not only location of other relevant text, but also of relevant graphics, audio, etc. Such system should also facilitate the performance of an automated help-desk function wherein queries can be received in a variety of forms and received queries can be responded to automatically where data is available in the system for such response or manually (i.e. by a person) where relevant response data is not available from the system. Ideally, the system should be adaptive so that when a response is manually provided, such response is automatically available in the future in response to the same or similar query.

In accordance with the above, this invention provides a method and apparatus for facilitating computer retrieval of database material. The basic feature of the invention is that selected subdivisions of the material, such as sentences, paragraphs, sections, chapters, articles, columns, subroutines, pictures, video frames, graphs, charts, verses, object classes, action specifications or the like, are annotated, preferably with natural language questions, assertions, or noun phrases, or some combination or collection thereof. For the preferred form of the invention, the inputs are in a natural language, often the natural language in which the text appears, for example, English, Russian, French, German, etc. The annotations may however also initially be provided in a structured form. The structured form may utilize natural language words, which is the preferred form, or may use an artificial language.

The annotations may be generated manually, semiautomatically or automatically. Where the annotations are generated at least semiautomatically, similarities may be detected between a current subdivision to be annotated and previously annotated subdivisions; then previous annotations from subdivisions having predetermined detected similarities to the current subdivision may be utilized to annotate the current subdivision. The detected similarities may, for example, be key words, noun phrases, or the like. Annotations from a prior subdivision having the predetermined similarities may be modified either automatically or manually so as to be appropriate for the current subdivision. In a semiautomatic mode, the system may propose previous annotations and an operator may select one or more of the proposes annotations which are suitable for use with current subdivisions.

In another method for the semiautomatic generation of annotations, a database of annotation groups is formed from existing annotated subdivisions. One annotation for a current subdivision is then selected, either by an operator or by some automatic technique, and the database of annotations is then searched for a similar annotation. Other annotations in the annotation group containing a similar annotation to the selected one are then either used or proposed to annotate the current subdivision.

One or more pointers may be stored with each stored annotation so that, during a search, when an annotation is located, all subdivisions relevant to that annotation may be quickly located. Relevant subdivisions need not all be in the same form so that pointers for an annotation may point for example to both text and graphic subdivisions.

Where the annotations are originally presented in a natural language, they are preferably converted to a structured language form for storage so that, regardless of how the annotations are initially presented, they are always stored in a structured language form. To the extent pointers are not utilized as indicated above, other connections are established between the annotations and the textual subdivisions.

Once the annotations have been stored, relevant subdivisions may be located by merely searching the annotations rather than the full database. The queries to the system may be presented as standard natural language queries which the system then converts to the structured form for purposes of performing an annotation match. The ability of the system to convert a broad range of natural language queries into a form which may be matched with annotations stored in a structured form permits searches for relevant subdivisions to be easily performed by unsophisticated users without requiring the storage of excessive material.

If a search in response to a query does not yield a required subdivision, the query may be utilized as a generated annotation for such required subdivision, which subdivision may either be an existing subdivision in the database which is either unannotated or not annotated for this query, or may be a new subdivision which is added to the database in response to the query. In particular, one potential application is for an automated or semiautomated answer desk function. In performing this function, input queries which may for example be received over E-mail, scanned from written text or inputted to the system from voice recognition are utilized as the inputs to search annotations for relevant subdivisions. If a relevant subdivision is found, the material in the subdivision is provided to the user in suitable form (i.e. as an E-mail response, as hard copy, as an audio response, etc.). If the search does not locate a matching subdivision, the query is referred to a person for response and the response generated by the person is then stored in the system as a subdivision with the initial query, or a suitably modified form thereof, being stored as a generated annotation for this new subdivision.

The foregoing and other objects, features and advantages of the invention will be apparent from the following more particular description of preferred embodiments of the invention as illustrated in the accompanying drawings.

FIG. 1 is a block diagram of exemplary system in which the teachings of this invention might be utilized.

FIG. 2 is a general flow diagram of the annotation generating process for a preferred embodiment of the invention.

FIG. 3 is a flow diagram of a process for recovering relevant subdivisions utilizing the teachings of this invention.

FIG. 3A is a flow diagram for an optional "help disk" function which may be performed in conjunction with the process of FIG. 3.

FIGS. 4A and 4B are flow diagrams of a database generating process and a database utilizing process for supporting the automatic or semiautomatic generation of annotations.

FIG. 5 is a flow diagram of an alternative annotation support mechanism.

FIG. 6 is a flow diagram of a filtering process which might be utilized to reduce annotations either used or proposed for use in connection with an automatic or semiautomatic annotation generating mechanism.

FIG. 7 is a flow diagram of a learning mechanism which might be utilized in conjunction with the various automatic or semiautomatic annotation methods to reduce the number of inappropriate annotations used or proposed for use by use with such methods.

FIG. 8 is a simplified example of a classification tree which might be utilized in conjunction with the learning mechanism of FIG. 7.

As previously indicated, this invention relates to facilitating the searching of a large database of material, for example, all legal cases from a particular court or from courts in a particular geographic area, to find portions or subdivisions of the material which are relevant to a particular query. While, for discussion purposes, the database will generally be considered to be a textual database in the following discussion, this is not a limitation on the invention, and, as indicated earlier, databases from which material may be retrieved utilizing the teachings of this invention include ones containing graphics (i.e. pictures, graphs, charts, drawings, video images, etc.), audio (i.e. speech, music, sound effects, etc.), text (including computer programs in various codes or languages, object classes and action specifications) and any other type of material which may be computer storable. An action specification may, for example, be a series of commands for controlling a process, robot, entertainment device or the like. For purposes of this invention, a query may be a question (for example, "What is Neptune's color?"), an assertion (for example, "Neptune is blue"), or a noun phrase (for example, "the color of Neptune") or some combination thereof.

In order to facilitate the retrieval of relevant material from the database, at least selected subdivisions of the database are annotated. The annotations can be in the form of questions, assertions or noun phrases or some combination or collection thereof. As will be discussed in greater detail later, the annotations can be manually generated, semiautomatically generated or automatically generated. The subdivisions to be annotated may all be of the same type, for example paragraphs, or may vary depending on information content. Thus, where an entire article or paper is relevant to only a single query, only a single annotation might be utilized for the entire article or paper. However, in other situations, annotations might be appropriate for each paragraph, for selected paragraphs or even for sentences within a paragraph. Depending on the text, the subdivisions utilized for annotation purposes might include sentences, paragraphs, sections, columns, articles, papers, chapters and the like. For speech material, the same subdivisions might be utilized as for text and for this or other types of material, other subdivisions may be appropriate such as a verse, subroutine, picture, image, etc.

The annotations are stored in an annotation memory with suitable pointers or other connectors to the relevant subdivisions. When an annotation previously stored is found to be relevant to another subdivision, additional pointers may be added to the annotation so that when the annotation is matched during a search, both relevant subdivisions will be retrieved. There is no limit on the number of pointers which may be associated with a given annotation, although the memory utilized may impose some practical limitations. There is also no limit to the number of possible annotations for a given subdivision.

Annotations may be generated in a natural language or may initially be generated in a structured form. For purposes of this invention, a natural language will be a language such as English, French, German, Japanese or the like which is normally spoken in some part of the world. While for most of the examples, the natural language utilized is English, this is not a limitation on the invention. Further, while in the preferred embodiments the annotations are in the same language as the text, in some situations it may be desirable to have annotations in a different natural language from that of the text being annotated. This might be useful, for example, in retrieving documents in a foreign language that have not yet been translated into the language of the person looking for text. Thus, selective translation would be made only of relevant text. This would also be useful, for example, if the person looking for text is different from the person who will read the retrieved text or if the person looking for text is only minimally competent in the foreign language. A structured form is a more stylized form of the initial query which may, for example, be more useful for pattern matching purposes.

Each annotation expressed in natural language is preferably translated into structured form, which captures at least part of the underlying meaning of the annotation in a prescribed stylized form. Similarly, when a natural language query is entered, it is also translated into structured form, which captures at least part of the underlying meaning of the query in a prescribed stylized form. Thus, it is possible to match queries to annotations via their similar structured forms, even though the surface wordings in natural language are substantially different. This matching of structured forms allows a large number of differently expressed queries, with the same essential meaning, to match a smaller number of annotations.

Thus, for example, a small number of annotations, such as the following, might serve for a paragraph from an annual report:

Did the unification of Germany influence Siemens?

The unification of Germany stimulated Siemens' business.

Note that a single subdivision can have any number of annotations.

The first of these annotations might be translated into the following structured forms:

(unificationl influence Siemens)

(unificationl related-to Germany)

Subsequent to such annotation, the paragraph would be offered in response to the following queries:

How was Siemens influenced by the unification of Germany?

What effect did the unification of Germany have on Siemens?

Was business stimulated by the unification?

How does German unification affect Siemens?

How was Siemens' business affected by the unification of Germany?

What stimulated Siemens' business?

Did the unification of Germany have an impact on Siemens?

Tell me about the influence of the German unification.

Why did the unification of Germany stimulate Siemens' business?

Did the unification influence Siemens?

When a matching annotation is located, all subdivisions relative to the matched annotation may be retrieved. Searching may be done by relatively unsophisticated users since the queries may be presented as normal English questions or assertions. Such normal English questions or assertions are then converted to structured form for searching and match.

Referring to the figures for a more detailed understanding of the invention, FIG. 1 shows a system which might be utilized to practice the teachings of this invention. The system includes a computer 10 which may be any of a variety of standard data processors available on the market. The size of computer 10 may vary to some extent with the size of the database being accessed, on other functions which may be required of the processor and on the speed required to perform the various operations. While for purposes of the following discussion it is assumed that the same computer 10 is being utilized both to generate the annotations and to search the database, this is by no means a limitation on the invention, and different processors might be utilized to perform these functions.

Computer 10 has at least one large capacity memory 12 in which the database to be searched is stored. Memory 12 may be any of a variety of existing, large capacity memories such as various disk drives. Computer 10 may also have an additional memory device 14 in which annotations in structured form are stored. Depending on the size of the main database, memory 14 may also be of considerable size. Memories 12 and 14 may be separate storage devices or may be various sections of the same storage device.

A number of standard input devices 16 may be available for providing information to computer 10. These may include, but are by no means limited to, keyboard devices, mouse or roller-ball inputs, text/image scanners, modems, network inputs from other systems and the like. One option available with this system is to include a voice recognition module 18 which may be utilized to input annotations or to input queries into the system. Voice recognition module 18 may be any of a variety of such modules currently available.

Finally, computer 10 may have a number of standard output devices 22 such as a cathode ray tube or other visual display 20, a printer, a voice synthesizer, robotic manipulator, etc. A user may be able to interact with display 20 by use of one of the input devices 16 to, for example, select annotations from a group of proposed annotations or for other purposes.

FIG. 2 illustrates the general procedure utilized to annotate text in accordance with this invention. The first step in the operation, step 30, is to select a subdivision to be annotated. As previously indicated, this may be any logical subdivision of the material from a sentence, picture, etc. up through an entire article or book. Typically, textual subdivisions for purposes of annotation might be a paragraph or section of a document, with other similar subdivisions being selected for non-textual material. The decision on the subdivisions to be annotated would normally be made by a person doing the annotation who would read, view, listen to, etc. the material as appropriate and decide on logical groupings of material with respect to a particular annotation. A single annotation may for example relate to mixed material which includes both text and graphics. It is possible that a single piece of text might be relevant to two separate annotations. Thus, with three successive paragraphs, paragraphs 1 and 2 might be relevant to a first annotation while paragraphs 2 and 3 are relevant to a second annotation. Paragraph 2 would thus be part of the subdivision for both annotations.

While in the discussion above, it is assumed that step 30 is performed manually, it is, of course, possible that this step could also be performed automatically. Thus, for example, a decision could be made that an annotation would be provided for each paragraph, thus removing the human decision making process. However, such a simplistic mechanical approach might result in annotations being generated for paragraphs which do not provide useful information and in perhaps a single annotation being generated for a paragraph which might be relevant to two or more queries. While more sophisticated subdivision generation algorithms can be envisioned, human decision in this process appears preferable at this point.

Once step 30 has been completed, the operation proceeds to either step 32 or 34, depending upon whether annotations are initially generated in English (or other natural language) or in a structured form. Assuming annotations are initially generated in a natural language, for example, as a question, assertion or noun phrase (or combination/collection thereof), the operation would proceed to step 32 during which an annotation for the selected subdivision would be generated. This annotation could be generated by a person utilizing a keyboard or other input device 16 to generate the annotation. Since generating annotations is a labor intensive task, it might be faster to permit the annotator to orally input annotations using the voice recognition module 18, with the module converting the oral input to machine readable form which is inputted to computer 10.

It is also possible that step 32 may be performed either semiautomatically or automatically. Ways in which automatic or semiautomatic generation may be performed are discussed later. Generally, these techniques involve finding some relationship between the current subdivision and a prior subdivision and either utilizing the annotations for the prior subdivision, as suitably modified for the current subdivision (automatic mode) or displaying the modified annotations to the annotator as proposed annotations with the annotator making selections from the proposals (semiautomatic operation).

When an annotation has been presented to or generated by computer 10 during step 32, the operation proceeds to step 36 during which the annotation, which is still in natural language form, is translated into a structured form. As previously discussed, the structured form is an artificial language which may still use natural language words or may use symbols that do not necessarily correspond to natural language words. Examples of structured form might be, for example, a T-expression discussed in chapter 6 of ARTIFICAL INTELLIGENCE AT MIT--EXPANDING FRONTIERS, edited by Patrick H. Winston with Sarah A. Shellard (the MIT Press). Various pattern matching languages or relational database forms may also be utilized to perform this function. Examples of such language are provided in LISP, 3rd Edition, P. H. Winston and B. K. P. Horn, Addison-Wesley Publishing, 1989, Chapters 24, 28, 29 and 30 being of particular interest. A preferred procedure for the translation of natural language to structured form (i.e. a natural language interface) is described in Chapter 6 of Artificial Intelligence at MIT--Expanding Frontiers. Other natural language interface programs and procedures are also known in the art.

This translation procedure means that a large number of alternative natural language queries, once translated into structural forms, will match the structured forms produced by a small number of annotations, thus facilitating searching.

Step 34 indicates that the annotation may initially be presented by the annotator in structured form. While presenting annotations in this form is within the contemplation of the invention, and a skilled annotator might not find this difficult to do, particularly with voice input, it may be faster and result in a better annotated database if the annotator inputs are in natural language rather than structured form.

Either from step 34 if the input is in structured form or from step 36 if the initial input is in natural language, the operation proceeds to step 38 during which the annotation in structured form is stored in memory 14. During step 38, a single annotation in structured form would typically be stored. However, if two or more annotations in structured form are generated during step 36, these would be stored during step 38.

From step 38, the operation proceeds to step 40 during which a connection to a related subdivision is added to the annotation stored in memory 14. This connection might, for example, be a pointer to the memory location in memory 12 at which the related subdivision is stored.

While not specifically shown in FIG. 2, step 38 might also include the step of checking memory 14 to determine if an annotation in the structured form which is the same as the annotation being provided from step 34 or 36 is already stored in memory 14. If the annotation is already stored, then it would not normally be stored again. Instead, an additional pointer to a related subdivision would be added to the existing annotation during step 40. However, in some instances, it may be required to store an annotation more than once in memory 14 and to chain the two entries together.

When step 40 has been completed, the operation proceeds to step 42 to determine if there are more annotations for the particular subdivision. If there are more annotations for the particular subdivision, the operation returns to perform either step 32 or 34, as appropriate, and steps 32-40 are repeated until, during step 42, it is determined that all annotations for a particular subdivision have been entered. When this occurs, a "NO" output is obtained during step 42, causing the operation to proceed to step 44 to determine if there are more subdivisions to be annotated. If there are more subdivisions to be annotated, the operation returns to step 30 to select a new subdivision to be annotated. When during step 44 it is determined that there are no more subdivisions to be annotated, a "NO" output is obtained during step 44, causing computer 10 to exit from the annotation routine.

FIG. 3 is a flow diagram of the retrieval mechanism utilizing the teachings of this invention. During step 50, the first step in this operation, a query is presented to the system. This may be done using a standard input device 16 such as a keyboard or using voice recognition module 18. With the voice recognition module, input queries may be received via telephone through a modem. From step 50, the operation proceeds to step 52 to determine if the query is in natural language or structured form. Typically, the query would be by a relatively unsophisticated user and would be in natural language resulting in a "YES" output during step 52. This causes the output to proceed to step 54 during which the query is translated into structured form using the same technique that was used to translate an annotation in natural language into structured form during step 36 (FIG. 2). If, because of limitations on the translation procedure, the query cannot be translated into structured form, the user could be provided with an indication that his query could not be translated into structured form. The user might then be asked to restate the query, perhaps in a simpler form. Alternatively, other retrieval schemes, such as key word analysis, or statistical methods could be tried, either initially or after a second query attempt by the user.

From step 52 if a "NO" output is obtained during this step or from step 54, the operation proceeds to step 56 to search for an annotation that matches the structured form query. Step 56 may be performed by means of a number of known routines which automatically search for similar or identical expressions, such as, for example, the search routine described in Chapter 6 of ARTIFICIAL INTELLIGENCE AT MIT--EXPANDING FRONTIERS, mentioned earlier. During step 58, a determination is made as to whether a match has been found. If a match is not found, the retrieval routine of FIG. 3 would be exited.

At this point, the user could be provided with an indication that relevant information to the query could not be located in the database and/or other retrieval schemes, such as key word analysis or statistical methods, could be tried.

If a match is found during step 58, the operation proceeds to step 60 to use a pointer stored with the matched annotation to recover a related subdivision from main database memory 12. Then, during step 62, the retrieved subdivision is supplied to an output device 22 for presentation to the user, for example via display 20 or a speech synthesis system. Alternatively, the retrieved subdivision may be employed without direct presentation to the user, as, for example, when the retrieved subdivision is a command or series of commands for controlling actions, such as those of a robot or an entertainment device.

While step 62 is being performed, or thereafter, the operation proceeds to step 64 to determine if there are additional pointers stored with the matched-on annotation. If there are additional pointers, the operation returns to step 60 to recover the additional related subdivisions and to step 62 to present such subdivisions. Such related subdivisions may contain the same or different types of information. This sequence of operations continues until, during step 64, it is determined that there are no pointers to additional subdivisions for the matched annotation. This results in a "NO" output during step 64 which causes the system to exit from the retrieval mechanism until a new query is received.

As was discussed earlier, in order to achieve the benefits of this invention, it is necessary to annotate text. Particularly for an existing database which contains large amounts of information, this can be a time-consuming, labor-intensive task. While the ability to present the annotations in natural language form means that the annotator may be a semiskilled individual, the annotator still needs to be someone who is sufficiently familiar with the material being annotated so that he can determine the query or queries to which a database subdivision is relevant and generate appropriate annotations. The remainder of the discussion is directed primarily to ways of speeding up and simplifying the annotation task by permitting it to be accomplished either automatically or semiautomatically. One particular way this is accomplished is by utilizing previous annotations to facilitate future annotations as the annotated database material accumulates.

One technique for the automatic or semiautomatic annotation of subdivisions is illustrated in FIGS. 4A and 4B. As was discussed earlier, a number of different annotations might be utilized for a given subdivision. Such a set of annotations, which may be questions, assertions, noun phrases or the like are referred to as an annotation group. When there are multiple annotations for a given subdivision, these annotations are formed into an annotation group. These annotations may be stored together as a group or pointers or other means might be provided to chain the annotations of a group together. Chaining may frequently be required since a single annotation may be part of two or more different annotation groups. Such chaining may, for example, be accomplished by augmenting each annotation with a pointer or pointers to each annotation group that it is a part of. The operation just described is illustrated by box 70 in FIG. 4A.

When an annotation operation is being performed during step 32 (FIG. 2), the operations of FIG. 4B may be performed. The first of these steps, step 74, is for the annotator to select one annotation for the subdivision being annotated. The operation then proceeds to step 76 during which the annotations stored in memory 14 are searched to find an annotation which is the same or similar to the inputted annotation. If an annotation which is the same or similar has been used before, the pointers attached to that indexed annotation are used to find the annotation group that the matched annotation belongs to. The other annotations in those annotation groups may then be retrieved.

In the discussion so far, it has been assumed that annotations are stored in memory 14 basically in the order in which they are received. While this is one possible way in which the annotations may be stored, the searching of annotations may be facilitated if the annotations are indexed in some systematic way. For indexing techniques, see for example Chapter 14, Artificial Intelligence Programming, First Edition, Charniak et al, Erlbaum, 1980. With indexing, pointers to related annotations or annotations of a particular annotation group would almost surely be required.

There is also a question as to how "similar" is determined during step 76. Generally, this means similar in form, but with different objects. Annotations in similar form with different objects might be considered "similar" if the objects are drawn from the same class. Suppose, for example, that a text is annotated by the sentence "Scientists believe that Neptune has volcanoes." In structured form, this annotation might appear as (Scientists1 believe (Neptune has-as-part volcanoes2)). This is similar to (Scientists1 believe (Pluto has-as-part volcanoes2)) because both Neptune and Pluto belong to the "planet" class. While the structured form (Scientist1 believe (sun has-as-part volcanoes2)) is also similar to (Scientists1 believe (Neptune has-as-part volcanoes2)), it is less similar than (Scientists1 believe (Pluto has-as-part volcanoes2)) because the most specific class to which the objects belong is neither planet nor star, but rather the "heavenly body" class, which is a super class of both planet and star. As will be discussed later, these might initially be considered to be sufficiently similar for purposes of step 76, but this decision may later be changed. Alternatively, a search for annotation groups which are "similar" during step 76 might initially look only for annotations which belong to the same specific class, such as "planet". If an insufficient number of new annotations are found with that meaning of "similar", then the objects might be allowed to belong to more remote classes, such as the "heavenly body" class. Some trial and error may be required with a given database in order to achieve the right level of specificity in classes for purposes of "similar" during step 76.

From step 76, the operation proceeds to step 78 to suitably modify the annotations which were found during step 76. "Modify" means that the objects in the stored annotation group are replaced, preferably automatically, by the corresponding object in a new annotation for the new piece of text. Thus, the substitutions used when looking for "similar" annotations are reversed when "modifying" annotations. In other words, if the annotation (Scientists1 believe (Neptune has-as-part volcanoes2)) is located when a search is being made for information relating to the planet Pluto, the located annotation would automatically be changed to read (Scientists1 believe (Pluto has-as-part volcanoes2)) during step 78.

From step 78, the operation proceeds either to step 80 or step 82. During step 80, the other annotations in the annotation group as modified are stored to be used as annotations for the new subdivision. This is a more or less fully automatic mode of operation.

Alternatively, during step 82, the other annotations from the annotation group or groups, as modified, are proposed for use to annotate the current subdivision. Such proposal might, for example, be accomplished by presenting such annotations on display 20. The annotator could then, by means of a keyboard, mouse, rollerball, light-pen, voice recognition system or other suitable input means, select from the proposed annotations the ones which are appropriate for the current text subdivision (step 84). The accepted annotations would then be stored during step 38 (FIG. 2) and connections to related text added during step 40.

FIG. 5 illustrates another way in which annotations might be automatically or semiautomatically generated. Referring to FIG. 5, during step 90, for a subdivision to be annotated or further annotated, key words are identified. Step 90 may be accomplished in a number of ways. It may be accomplished automatically by, for example, identifying words in a text subdivision which match a list of specified key words. It may also be done manually by, for example, asking a human annotator to supply key words from a text or non-textual subdivision. Finally, it may be done semiautomatically by, for example, asking a human analyst to select from a list of words identified in text by a language analysis program. The possible key words could be presented to the annotator by a menu, or highlighted in mouse-sensitive text. Other techniques known in the art for identifying key words in text or other material might also be utilized.

When step 90 is complete, the operation proceeds to step 92 during which the database which has already been annotated and stored in memory 12 is searched for subdivisions with similar key words. "Similar" during this step can mean similar according to a variety of similarity measures, some of which are presented in an article entitled "Automatic Information Retrieval" by Gerard Salton in IEEE Computer, September 1980, pp 41-56. Systems for determining such similarities of key words are known in the art.

When a database subdivision having "similar" key words is located during step 92, the operation proceeds either to step 94 or 96. During step 94, the annotations for a subdivision that exhibits similar key words are suitably modified, for example, manually, and the modified annotations are then utilized for the new text subdivision.

Alternatively, during step 96, the same annotations are retrieved and modified in the same way; however, the modified annotations are then proposed for use to annotate the current subdivision in much the same way that annotations were proposed for use during step 82 (FIG. 4).

From step 96, the operation proceeds to step 98, during which selected ones of the proposed annotations are accepted by the annotator for use for the current subdivision. This step is basically the same as step 84 in FIG. 4B.

It is probable that, at least initially, the various automatic or semiautomatic annotating mechanisms discussed above, or other automatic or semiautomatic annotation techniques will provide or propose a number of inappropriate annotations. The filtering scheme of FIG. 6 and the learning scheme of FIG. 7 are intended to reduce and eliminate as many inappropriate proposed annotations as possible. The general idea is to restrict the use of an annotation group to annotations whose objects fall into particular classes. Class may be determined from a classification tree such as that shown in FIG. 8. This tree is obviously cursory and the tree formed for a particular database would be more extensive, although perhaps less inclusive. Definitions for such a tree could be, for example, taken from a dictionary and entered into the system manually, semiautomatically or automatically. Thus, a code could be stored with each annotation group specifying the classes of objects that limit the application of the annotation group. For example, an annotation group involving the planet Neptune might be restricted to the class "planet," or to the class "heavenly body," While it might be possible to restrict this annotation group to a higher item on the tree, for example, "physical object," this would almost surely result in many inappropriate annotations being utilized or proposed.

Referring to FIG. 6, during step 100, a determination is made if objects in a given annotation, obtained from the annotator in step belong to the allowed classes specified for the annotation group obtained in step 76 (i.e. the objects for the current annotation are included within the allowed classes for the annotation group). If the objects do belong to those allowed classes, then the annotation group passes the filtering step and the annotations in that annotation group, as suitably modified in step 78, are either accepted or proposed as annotations for the new subdivision (step 102). If the objects do not belong to those allowed classes, then the annotation group is rejected (step 104) and its annotations, therefore, are not used or proposed. A similar filtering technique could be utilized for other automatic or semiautomatic annotation techniques.

FIG. 7 illustrates how the filtering step may be improved based on decisions made by the annotator during, for example, step 84 or 98 where the annotator selects appropriate annotations from a proposed group. Thus, during step 110, the classes of objects in a proposed annotation that is ultimately rejected during step 84 or 98 is noted. During step 112, the classes of rejected annotations are utilized to make allowed classes more specific for the annotation group that produced the rejected annotation.

Thus, if the code in an annotation group limits the annotation group to the "heavenly body" class, and it is found that annotations involving the sun or other stars are rejected (step 104) when proposed in connection with that annotation group, then the code limiting that annotation group could be modified to limit the annotation group to the class "planet," which appears underneath the class "heavenly body" in the tree. This would reduce the likelihood of inappropriate future annotations being used or proposed.

As previously discussed, one possible application where queries might be presented as illustrated in FIG. 3 is an automated or semiautomated "help desk" function (see also FIG. 3A). In this function, individuals would contact the help desk to secure information or assistance relating to a specific product line, service, or other subject. For example, users of a particular computer program which is relatively complex might call the help desk to secure information or assistance to resolve problems in the use of the program, to interface the program with various hardware or other software, to upgrade or enhance the program or for other purposes. Queries could be presented over E-mail or through a network in electronic form, or over telephone lines through a modem as either data or voice queries which are transmitted through a modem to voice recognition module 18 or could be written inquiries which are scanned into the system for response with a printed reply being generated. Other techniques for receiving queries and providing responses known in the art could also be utilized.

When a query is received, which query is assumed to be in natural language, it would be translated into structured form during step 54 and then a search performed for matching annotations during step 56. If matching annotations are found, the corresponding subdivisions in the database would be presented to the querier in suitable form (i.e. transmitted over a network as E-mail, passed through a speech synthesizer to generate an oral response over a phone line, printed out for hard copy mailing to the querier, etc.). If the match is not found, rather than exiting during step 58, the query would be referred to a person manning the help desk for response (step 65). When the person generates the response (step 66), the response is added to the database (step 67) and the structured form of the query outputted during step 54 is added to the database of annotations with a pointer to the newly-generated database response (step 68). This is another way in which annotations could be automatically generated.

A method and apparatus is thus provided which permits large databases to be accessed by relatively unsophisticated users by the user posing a query to the system in normal natural language. The system can then access subdivisions relevant to the user's query and, if the original annotating has been done properly, all subdivisions relevant to the query will be presented.

While the system requires a relatively time-consuming initial annotating process, this process needs to be performed only once for a given database subdivision and, once a substantial database of annotations has been developed, it will be possible to perform the annotating process at least semiautomatically and possibly totally automatically. A powerful tool for enhancing the accessibility and usefulness of large databases is thus provided. The technique might also be used for "triggering," with new annotations being examined as they are made to determine if the corresponding database subdivision is relevant to a standing query.

While an exemplary system has been shown in FIG. 1 and exemplary annotating and retrieval techniques have been shown in the remaining figures, it is to be understood that this hardware and these techniques are shown for purposes of illustration only and that other hardware or other techniques could be utilized while still remaining within the spirit and scope of the invention. In particular, while various techniques have been described above for performing automatic or semiautomatic annotating, it is within the contemplation of the invention that other techniques for performing these functions might also be utilized.

Thus, while the invention has been particularly shown and described above with reference to preferred embodiments, the foregoing and other changes in form and detail may be made therein by one skilled in the art without departing from the spirit and scope of the invention.