JP2003515815A - Method and apparatus for storing and retrieving knowledge - Google Patents

Abstract

(57) [Summary] An apparatus for storing and retrieving knowledge includes a natural language parser, a knowledge research agent, and a directional (compass) search agent. Methods for storing and retrieving knowledge can include directional searches that combine query text with relationships to the search, and automatic knowledge model construction based on existing knowledge model content and extrinsic information searches.

Description

Detailed Description of the Invention

BACKGROUND 1. Field of the Invention The present invention relates to systems and methods for organizing, storing, and retrieving knowledge.

(2. Related Technology) At present, much human knowledge is configured into information structures such as a database, a set of text documents, and a set of expert system rules. However, the knowledge contained in such structures is difficult to efficiently identify, locate, and retrieve when needed.

One of the known and more efficient knowledge memory retrieval systems is Hyperkn
Hyperknowledge knowledge base developed by owledge Corporation and described by Trevor Davies in US Pat. No. 6,0023,296 issued Dec. 14, 1999. Hereinafter, this patent is referred to as Davies, which is incorporated herein by reference in its entirety. A knowledge base constructed according to the Davies principle,
Below is a linked knowledge model.
In this link knowledge model, four useful relationships between elements of the represented knowledge are included in links between the elements. These relationships are conveniently represented by the points of the compass as shown in Fig. 1. In this compass, the information of "West" of the selected concept is "Why" about the selected concept. , And the “east” information of the selected concept indicates that it details the “how” of the selected concept. Further, the “south” information of the selected concept details “what” about the selected concept, and the “north” information of the selected concept describes “what” about the selected concept.
"What?" By arranging information detailing "what", "what", "why", and "how" above, below, left, and right of the selected concept, the direction of north, south, west, and east Is conveniently represented as a visual representation of the link knowledge model.
As explained by Davies, a pair of opposing directions, namely north-
The relationship represented by South (what-what) and East-west (how-why) is
It is complementary.

SUMMARY OF THE INVENTION Embodiments of the present invention help users find relevant knowledge rather than raw data that is not relevant to the user. Knowledge is used herein to refer to information that is structured according to context and meaning, and information may be unstructured raw information or structured according to other criteria such as input format, vocabulary, origin, etc. Used to refer to information.

To find knowledge, embodiments of the present invention take information as a natural language and generate a link knowledge model from this information. Then navigate the link knowledge model according to the "how,""why,""what," and "what" relationships discussed above, using a user interface that makes it easy to connect information together. Knowledge can be extracted.

Embodiments of the present invention can augment a user's knowledge by augmenting the knowledge base available to the user in the form of a linked knowledge model. Voids in the link knowledge model and edges or boundaries of the link knowledge model can be detected automatically. Thus, embodiments of the present invention can detect such voids and edges and search for missing information on an ad hoc basis, on a regular basis, or by other criteria.

Some embodiments of the present invention automatically analyze the knowledge sought in queries received, for example, by email, and are based on the content of a dynamic link knowledge model accessible to the user. A user interface may be included that creates a response to these queries. Further, the content of the queries and other information received can be related to the information already contained in the link knowledge model to automatically augment the model. The link knowledge model can also be presented in various useful formats, depending on the type of knowledge being sought.

According to one aspect of the invention, there is a system for storing and retrieving knowledge.
This system has a text analyzer that has an input that receives information in natural language and that produces note information at the output, and an input that receives note information from the text analyzer and that generates a link knowledge model. It includes a parser stored in memory and a user interface that presents at least one complementary pair to the operator, through which the operator can navigate the link knowledge model to retrieve knowledge. The system may further comprise a knowledge exploration agent having an input for receiving the linked knowledge model and an output, the knowledge exploration agent generating at this output a query adapted to an external search engine. Hyper-knowledge agents can automatically augment the link knowledge model with new information found by executing queries against external search engines. In this system, the parser may further comprise a rules engine that executes a plurality of rules that define the structure of knowledge embedded in the information expressed in natural language. The link knowledge model may include associated noun / verb phrase pairs, and multiple rules describe how a natural language sentence can be associated with associated noun / verb phrase pairs in the link knowledge model. You can define what is relevant. A rule is at least one semantic rule, at least one syntactic rule,
Or it may include at least one context rule.

According to another aspect of the invention, there is a method of retrieving knowledge. The method presents the operator with a selector that the operator uses to indicate one of the complementary pairs and a query input that the operator uses to indicate the query text, and
Combining one of the indicated pairs of complementary relationships with the query text to form a query and retrieving knowledge with this query. In accordance with this aspect of the invention, the selector may further define four directions, characterized by pairs of complementary directions, each direction representing a relationship. The selector can also be presented in a graphic resembling a compass. The method further includes storing knowledge in a link knowledge model that includes noun phrase / verb phrase pairs and searching the link knowledge model for a first noun phrase / verb phrase pair defined by the query text. , And a second noun phrase / verb phrase pair having one of the indicated complementary relation pairs to the first noun phrase / verb phrase pair in the link knowledge model. . Finally, searching can further include parsing the natural language query text into searchable Boolean expressions.

According to another aspect of the invention, there is a method of storing knowledge. This method analyzes a natural language document into a link knowledge model that includes linked noun phrase / verb phrase pairs, and represents a link knowledge model part whose knowledge level is lower than other link knowledge model parts. Identifying and performing a search for external information for additional knowledge related to the identified link knowledge model portion. According to this method, the identified link knowledge model portion further comprises a second noun phrase / verb phrase that is related by at least one of the two complementary relationships defined between the noun phrase / verb phrase pair. The first noun phrase / verb phrase pair that lacks the pair may be included.

According to another aspect of the invention, each of the aforementioned aspects of the invention can be implemented as a software product that includes a machine-readable medium encoded with a series of software instructions, the software instructions being executed. Then, one of the methods discussed above is instructed to perform, or a general purpose computer operates as a special purpose system as defined above.

[0012]   In the drawings, like reference numbers indicate like elements.

DETAILED DESCRIPTION The invention will be better understood when the following detailed description of some embodiments and aspects thereof is read in connection with the drawings.

Embodiments of the present invention have four principle aspects, which are discussed below.
A complete architecture incorporating the principles of the present invention will then be discussed. These four modes are called 1) directional search ability, 2) knowledge extraction from natural language information, 3) automatic knowledge search, and 4) interface agent.

Directional Search Capability The directional search capability, eg, using a compass search agent (see 401 in FIG. 4), allows a user to specify a knowledge direction when performing a specified search. Therefore, the search provides a more focused answer to the user's query.
This capability is facilitated by using data structures and visual presentation by Davies.

When the user wants to search for knowledge, the user inputs a search phrase in a search box 201 as shown in FIG. 2, for example. The user then selects the desired knowledge direction by using a conventional pointing device such as a mouse to select one of the orientation points in the compass 202 that represents the four knowledge directions.

The operation is simple if the search is performed on a link knowledge model. A fuzzy match algorithm is used to search for processes or noun / verb phrases that match the search phrase. The link knowledge model is then navigated, i.e. the selected concept is changed in the direction of the request by one step and the resulting process or noun / verb phrase is returned. Therefore, searching for the phrase “why” in the phrase “hire an employee” does not mean general information about employee hire, how to approach employee hiring, or what employee to hire. , Will return knowledge explaining why an employee should be hired. If a “what” or “what” direction is requested, noun phrases in this direction are returned along with a selection list of verbs related to these noun phrases that are also extracted from the link knowledge model. If a general search is desired, the user can choose not to specify a direction,
When large link knowledge models are involved, it is likely that the information will be overloaded.

FIG. 3 shows another search example. In this case, the user has entered the search phrase 301 "maximum web market". The user asked the question "How to maximize my web market?" By clicking the "How" arrow 302 on the compass. The search results are narrowed down to "how to maximize the web market". To further refine the knowledge area to be explored, the user can click on the desired search direction in the compass 303 near the topic 304 of interest.

For a generalized source of information, eg, a collection of hypertext documents commonly known as the World Wide Web, stored on an international computer network commonly known as the Internet. When a search is performed, a connection to a conventional search engine is first established. A query is passed to the search engine, which query includes the request direction attached to the search phrase. For example, when searching for "employ employee" in the "why" direction, the query passed to the search engine is "why employ employee". By adding directions to the query, the page that answers the "why" question goes to the top of the search results. Once the relevant pages have been searched, analyze them, extract the knowledge contained in them,
It can be displayed or added to the link knowledge model selected or maintained by the user. Next, this automatic knowledge extraction process will be described in more detail.

Knowledge Extraction from Natural Language Information Eighty percent of the information stored by a typical business entity is estimated to be in the form of text-based documents, such as paper documents, emails, presentations, web pages. However, the knowledge represented by all this information is unstructured and stored inefficiently in the form of natural language. Embodiments of the present invention address this issue by automatically building a link knowledge model from natural language information using a knowledge exploration agent (see 402 in FIG. 4). Spoken words, written text documents, email messages, and web pages are transformed into linked knowledge models when connected to appropriate input devices and translators, which are easier to use with directional search. To be searched.

As shown in FIG. 5 and described in more detail below, the natural language parser (40 in FIG. 4).
3) analyzes the input information and builds the information as a link knowledge model. In summary, a text analyzer analyzes natural language sentences and annotates them with language and semantic tags. The context engine ensures that relationships across sentences are always known, allowing information from one sentence to be placed in the correct context with respect to information from another sentence. The parser engine then first notes the noun phrase / verb phrase pair that is generated as described below with respect to the text analyzer, and then converts the noted noun phrase / verb phrase pair into a link knowledge model corpus (KMC). The comparison produces noun phrase / verb phrase pairs and the links between them.
Optionally, the parser engine also relies on rules supplied and applied by the rules engine.

KMC consists of a number of linked knowledge models that are manually constructed from a range of texts and other knowledge sources. There will be many types of corpora each focusing on a particular area of knowledge, such as the knowledge contained in a typical business book. The structure of these corpus is described below. The user
If the input text matches the subject of a particular corpus, that corpus can be selected. Each noun phrase / verb phrase pair in the KMC is tagged with a carefully selected note selection so that the pattern between the text and the corresponding noun phrase / verb phrase pair can be detected. These notes can be considered to belong to different levels of knowledge, and these will be discussed next.

The lowest level note is a pure linguistic analysis of the text. These include parts of speech (for example, verb for using, noun for customer), and British Nat.
Included are word frequencies (highlighted when words or phrases occur more frequently than expected) as compared to standard English, possibly extracted from Ional Corpus. Optionally, a basic word form (eg, custom for customer)
ers, using for) and syntactic structure (commonly referred to as a sentence "parse tree"). All these notes can be automatically added to the KMC using existing tools such as Link Grammer and WordNet.

Intermediate level notes are at the semantic level. In this case, the text and noun phrase /
Each word or multi-word phrase in both a verb phrase pair is tagged with their semantic domain. This allows the meaning of a text or noun phrase / verb phrase pair to be compared with other text or noun phrase / verb phrase pairs, rather than just comparing the words themselves. This level of note also applies to Longman's Dic
It can be automatically generated in the KMC using a lookup resource such as a temporary of Contemporary English.

The highest level note is the noun / verb phrase pair itself. They depend on their structure (what, how, why, and what are linked to other noun / verb phrase pairs) and content (verb and noun phrases) by the KMC. Compose the most semantically rich data of all. By this and to ensure correctness, the noun / verb phrase pair and its associated natural language text are manually cross-referenced.

To facilitate pattern matching within KMC, noun phrase / verb phrase pairs must be expressed in the form of text markup that describes the structure of the link knowledge model. One of the possible implementations is Hyper Knowledge Markup Language (HKM
L, SGML (Standard General Markup Lan), which we refer to as Hyperknowledge Mark-up Language
However, other representations can be adopted. H
The KML structure contains any number of tags, some of which are defined below. Each tag is used to describe a unique part of the link knowledge model. Most tags should be closed with a closing piece (usually indicated by a forward slash (/) inside the tag). Some tags are described below, but other tags can be added to define other parts of the link knowledge model. Also, other names or syntactic constructs known in the art can be used.

<ENVIRONMENT> Defines the scope of the entire environment. </ ENVIRO
Close with NMENT> tag. -<MODEL> Define each model in the environment. Close with the </ MODEL> tag. <KNOWDES> Defines the section of the model noun / verb phrase pair.
Close with the </ KNOWDES> tag. <KNOWDE> Define a single noun / verb phrase pair in the model. </ K
Close with NOWDE> tag.・ Attribute KEY = number ・ Attribute CONNECTOR = cnAND / cnOR / cnTHEN ・ Attribute SUPERIOR = number (instance noun phrase / verb phrase pair, type noun phrase / verb phrase pair) ・ <OBJECTIVE> special noun phrase / Define a verb phrase pair. </ OBJ
Close with ECTIVE> tag. Enclose the <KNOWDE> tag when the noun phrase / verb phrase pair is the object. <VERBPHRASE> Defines a verb phrase of a noun phrase / verb phrase pair. Close with the optional </ VERBPHRASE> tag. Attribute TEXT = string <NOUNPHRASE> Defines the noun phrase of a noun phrase / verb phrase pair. Close with the optional </ NOUNPHRASE> tag. Attributes TEXT = Strings <RELATES> Defines the relevant section of the model. </ RELA
Close with TIONS> tag. -<RELATION> Defines a relational link. Optional </ RELAT
Close with ION> tag. • Attribute RELKEY = Numeric • Attribute NEXTREL = Numeric (defines the following relationship) <HOW> Defines the "how" noun phrase / verb phrase pair of the relationship. Close with the optional </ HOW> tag. Attribute HOWKEY = number. <WHY> Defines a "why" noun / verb phrase pair for the relationship. Optional <
/ WHY> Close with tag.・ Attribute WHYKEY = number

One environment file can contain one or more link knowledge models. Hyperknowledge environment files are written in HKML, although other environments can be built using other markup languages. Each hyper-knowledge environment file (denoted by <ENVIRONMENT> tags) contains an unlimited number of link knowledge models (denoted by <MODEL> tags). Each model contains two sections. That is, a section of a pair of noun phrases / verb phrases indicated by a <KNOWDES> tag and a section of a pair of related noun phrases / verb phrases indicated by a <RELATION> tag.

The noun phrase / verb phrase pairs section includes an unlimited number of noun phrase / verb phrase pairs,
Each is indicated by a <KNOWDE> tag. The <OBJECTIVE> tag can be used to enclose a <KNOWDE> tag to indicate that the noun / verb phrase pair is the object of the model. Each noun phrase / verb phrase pair has several attributes, namely its unique identifier (key), its connector type, and its owner if it is the "what" of a noun phrase / verb phrase pair. And a noun phrase / verb phrase pair. Between the start and end <KNOWDE> tags, the tag <VERBPHRASE
> And <NOUNPHRASE> must be included. The two tags define the two text parts that make up the noun / verb phrase pair.

The Relations section contains an unlimited number of relations (each indicated by a <RELATION> tag). Each <RELATION> tag defines a noun / verb phrase pair-to-pair relationship. Between the opening and closing <RELASION> tags, <HOW
> And <WHY> tags must be included. The two tags define the actual two "how" and "why" noun / verb phrase pairs that are included in this relationship. Each relationship comprises several attributes, including its unique identifier, relKey, and the next relationship in the chain, nextRel.

Within the <KNOWDE> and <RELATION> tags, unique identifiers or keys are used to match individual noun phrase / verb phrase pairs with those contained in the relationship. This helps to describe the unique ties between noun / verb phrase pairs so that the structure is clear.

The rule engine contains syntactic and semantic rules, both of which are manually developed and automatically generated via a data-oriented analysis of the corpus.

The final output of the parser engine is a link knowledge model containing noun phrase / verb phrase pairs linked by "how-why-what-what" links. It then employs various interfaces to output this structure to the user. These will be described in more detail later.

Conventional parsers for obtaining machine understanding of natural language are very complex due to the range of problems to be solved. However, the Davies model structure is inherently based on natural language noun phrase / verb phrase pairs. Semantic or actual knowledge is held in combination with the phrase pairs and the "how-why-what-what" relationship between the various phrase pairs. Thus, the task of mapping from natural language to knowledge that can be stored, searched and retrieved using machines is much easier to handle.

The natural language parser (403 in FIG. 4) detailed in FIG. 5 may be on a remote or local server computer, or on a client computer, such as a personal computer used by a user. . Placing the parser on a server on the Internet is particularly advantageous when used to extract knowledge from the World Wide Web. Placing a natural language parser (403 in FIG. 4) on the user's client computer is particularly advantageous when used to extract knowledge from the user's personal documents, reports, presentations, emails, etc. is there.

Automatic Knowledge Survey When a link knowledge model of any size is constructed, the structure of the link knowledge model itself reveals the scope of knowledge contained in it. The process at the west end of the link knowledge model (see the compass in Figure 1), the noun phrase / verb phrase pair, defines the understanding of "why" and the boundaries of knowledge, and the east end of the link knowledge model (the compass in Fig. 1). The process in (see figure) demarcates the understanding of "how". Similarly, "how"
Or a process with few "why" ties may represent a void in knowledge. Also, a process that lacks a "what" or "what" connection may represent a void in knowledge. Such voids or edges can be filled with a knowledge research agent (see 402 in FIG. 4) as follows.

First, we identify the processes at the ends of the model, or processes whose number of connections indicates a void in the knowledge. A directional search is then used to find the missing information, eg on the Internet or on the user's unanalyzed document. Knowledge is then extracted from the information thus found using the parser described above.

After being initiated, an automated survey can be essentially left idle and ongoing to continuously expand machine understanding of a topic.

Interface Agent The Interface Agent (see 404 in FIG. 4) facilitates the implementation of the processes described above by facilitating the import of natural language information in various formats. Agent is Hyper in Woburn, Mass.
It is implemented based on the application programming interface (API) in the product of knowledge Corporation (manufacturing the link knowledge model by Davies). The hyper-knowledge products used include standard automation, DCOM, and CORBA interfaces. Other standard and custom interfaces can also be implemented and used in other products.

Next, the operation of the electronic mail interface agent will be considered in detail.

The interface agent detects the arrival of new email. When the email arrives, the interface agent invokes the natural language extraction technique described above to extract knowledge from the email message. A model built from a single email may be useful to help understand particularly long or complex messages.

The processes extracted from the email and formed into a model of the email message can then be searched in a local or other linked knowledge model accessible to the user. References to the processes involved and other information can then be integrated into the email model. Similarly, the model of email can be integrated with existing link knowledge models, thus extending the knowledge and understanding contained therein. One of the advantages of extending the email model as mentioned is that the user will have more complete information when finally reading the email. One of the advantages of extending the local link knowledge model as mentioned is that the users of this local link knowledge model will have more complete information when examining the local link knowledge model.

The knowledge extracted on the fly from emails as described above can be used in a variety of creative ways. The extracted knowledge can be used to dynamically generate organizational structures within the email reader client software. For example, the key subject of an email can be used to create a folder in which the email can be stored or linked. The extracted knowledge can also be used to automatically generate informative replies to email questions. When the extracted knowledge points to a specific part of the local link knowledge model and the agent finds the ties to the desired level of confidence, it generates an automatic reply containing the relevant parts of the local link knowledge model. can do.

Consider the following simple example. An email arrives just saying "How do you delete the records?" The natural language parser uses "delete" as a verb,
Recognize "record" as a noun and create a process consisting of the noun / verb pair "delete record". The word "how" then triggers a link knowledge model search in the "how" direction. The returned link knowledge model portion will then be sent as a reply to the original email.

[0045]   Next, a more complex example is provided.

[0046]   In this example, the user received the following email.

“To John, we need to meet urgently to understand the problems we face when we start to tackle the global economy. We must make the proposed e-commerce strategy a success! I'm looking forward to your thoughts. David "

In order for the user to (a) summarize the content, (b) mark up keypoints with hypertext links, and (c) know more using the links generated in step (b) First, the natural language parser 403 decides to read the text.

First, the email text is annotated with syntax and meaning. This information is used to extract candidate noun phrase / verb phrase pairs from the email text. These include:

[0050]   ・ Global economy / work   ・ Proposed e-commerce strategy / success

The noun phrase / verb phrase pair matcher then compares the text, syntax, and semantic information of the candidate noun phrase / verb phrase pair with the annotated noun phrase / verb phrase pair in the KMC. Searches for a noun phrase / verb phrase pair that is close to a match in KMC. Consider the following extraction from KMC.

“Your site should deal with the global economy. The point is that your site will achieve this by considering foreign languages. The arrival of the global economy is the success of e-commerce. Offer more opportunities to experience. ”

[0053]

[Table 1]

The above represents only a small section of the KMC, showing only the text and manually generated noun / verb phrase pairs that represent the text.

It can be seen that the candidate noun phrase / verb phrase pair “work on global economy” matches in meaning and syntax well with the KMC noun phrase / verb phrase pair “address the global economy”. Thus, the model builder 405 stores the "work on global economy" noun / verb phrase pairs along with links to the "address the global economy" noun / verb phrase pairs in the KMC. This link is a note of a candidate noun phrase / verb phrase pair and KM
Tagged with a percentage confidence number based on the match level between the notes of the C noun phrase / verb phrase pair. In addition to this, all notes are stored with a noun / verb phrase pair, including a link back to the original email text. This association allows a hyperlink to be displayed in the email text.

Next, the KMC noun phrase / verb phrase pair “to deal with the global economy” was examined for “how” and “why” and other candidate noun phrases / verbs extracted from the email text. Compared to a pair of phrases. A good match is also found between the candidate noun phrase / verb phrase pair “successful proposed e-commerce strategy” and the KMC noun phrase / verb phrase pair “experience successful e-commerce”. In this case as well, the model builder 405
Remember this tie and percentage match. Also, this candidate noun phrase /
The verb phrase pair is combined by the model builder 405 as the “why” of the “work on a global economy” candidate noun phrase / verb phrase pair.

Suppose no more significant matches are found for KMC. in this case,
The original email text was marked with two hyperlinks representing the two noun phrase / verb phrase pairs generated by Model Builder 405. When the user moves the compass pointing device over these hyperlinks, the user can see that "work on global economy" has "why" in the document (indicated by the red arrow at the western end). ). When clicked on, the user jumps to the text "Successful proposed e-commerce strategy". But importantly, the user
It can also be seen that the KMC has suggestions on "how" this can be done (indicated by the blue outlined arrow at the eastern end). If you follow this link, you'll see a list of noun / verb pairs of "how" from KMC.
Consider foreign language "is included. This system not only added structure to the document, but also formed links to related knowledge.

Of course, other colors, geographic shapes, etc. could be used as indicators in this context.

Natural Language Parser Next, the natural language parser 403 will be described in more detail with reference to FIG.
Natural language parser 403 comprises several sub-components. All of the various types of inputs 500 are converted by a text analyzer 502 into text 501 ready for analysis. After analysis, noun phrase / verb phrase pairs can be extracted from the text as candidates for inclusion in the final model. The text analyzer tags the text with the same notes used in KMC. These will be described in detail below. These notes are generated using the Language Structure Tool 503, which includes Link Grammar, WordNet, and Longman's Dictionary of Contemporary.
Includes, but is not limited to, English. The text analyzer then optionally summarizes the supplied text using techniques such as, but not limited to, Shannon's Information Theory.

The context engine 504 is used to ensure that cross-sentence relationships are always known and that information from one sentence is placed in the correct context with respect to information from another sentence. . Resources such as the Anaphora Resolution Tool assist in this process. Unfortunately, sentences are not always self-contained knowledge units that the parser engine 505 can act on individually. For example, consider the sentence "Click it to start the music." The referent of the pronoun "that" cannot be determined without examining the previous sentence. Assuming the previous sentence is "you will see that there is a button labeled Play", then the context reveals that "it" refers to "Button labeled Play". The context engine 504 builds fine-grained noun associations across multiple sentences, thereby allowing context reference resolution across sentences.

Finally, the text analyzer 502 uses the previously constructed note to construct candidate noun / verb phrase pairs from the text. For example, part-of-speech analysis reveals the presence and location of verbs and nouns, which can be used to generate noun phrase / verb phrase pairs. At this point, the results of the summary analysis (if applied) and the word frequency notes can be used to assign a value of "importance" to the noun / verb phrase pair. When a word or phrase appears more often than expected, the noun phrase / verb phrase pair it produces becomes more important.

Parser engine 505 uses text with notes and linked candidate noun phrases /
The verb phrase pair 506 is received from the text analyzer 502. The rules engine 501 and hyper-knowledge corpus 508 are then utilized to generate the final linked noun phrase / verb phrase pair structure. Before describing the components of the parser engine 505, the rule engine 507 will be described.

The rules engine 507 includes a list of rules that facilitates faster analysis of natural language text so that corresponding noun phrase / verb phrase pairs are output faster and more accurately. Lead. Such rules can be found manually and
C can be used to develop, test, and refine these invented rules. The possible rules can be quickly applied to the KMC and the number of cases where the rule works and where there are exceptions is found. If the number of exceptions is small, the rule can be accepted and further refined to handle these exceptions.

In parallel with this, the natural language parser 403 can also initiate automatic extraction of patterns and rules contained within the KMC by applying data-oriented analysis techniques such as Hidden Markov Models.

Parser Engine The first task of parser engine 505 is to pick candidate noun phrase / verb phrase pairs in order of importance and find a matching noun phrase / verb phrase pair in the KMC. A match is found by matcher 508 comparing the notes of the candidate noun / verb phrase pair with the noun / verb phrase pairs and their notes within the KMC. KMC
The presence of the matching pattern of linguistic data and semantic data indicates that there is a matching noun phrase / verb phrase pair. The noun phrase / verb phrase pair that best matches, ie, matches most notes, is selected and passed to Model Builder 509.

Then, the noun phrase / verb phrase pair in the KMC linked to this noun phrase / verb phrase pair is semantically analyzed to find candidate noun phrase / verb phrase pairs for a match. To be searched. If found, replace the words in the KMC noun phrase / verb phrase pair with the words in the supplied text, keeping the same semantic context, and
Generate pairs of verb phrases. This process is repeated until most of the supplied text is covered and linked to noun / verb phrase pairs.

Optionally, rule engine 507 is employed to assist in finding matching noun phrase / verb phrase pairs in the KMC, or to propose new noun phrase / verb phrase pairs and ties. It For example, consider the sentence "He put food in the refrigerator because he wanted to freeze it." In this case, the two candidate noun phrase / verb phrase pairs identified by the text analyzer are "food in refrigerator".
And "Frozen food". The presence of the connective word "do" indicates that these noun phrase / verb phrase pairs are likely to be connected by a "why" tie. This can be further confirmed by the fact that there are semantically matching examples in KMC.

The model builder 509 accepts noun phrase / verb phrase pairs from the noun phrase / verb phrase pair matcher 508 and builds a model of linked noun phrase / verb phrase pairs. Some noun phrase / verb phrase pairs match the text supplied to the parser engine, maintaining the links between these items. Other noun / verb phrase pairs exist only in the KMC, with some accuracy and relevance of the noun / verb phrase pair. These noun phrase / verb phrase pairs are stored as links from the noun phrase / verb phrase pairs in the constructed model to the noun phrase / verb phrase pairs in the KMC. In this case, which noun phrase / verb phrase pair represents the structure of the supplied natural language text, and which noun phrase / verb phrase pair is not included in the text but is already known in the KMC. You can clearly see whether to extend.

Browsing Output Structure FIG. 6 illustrates a sample output of parsed natural language text. Text 600 with identified noun phrase / verb phrase pairs is displayed with an underlined hyperlink 601 for each noun phrase / verb phrase pair. When the cursor hovers over the hyperlink, the compass 602 is presented to the user. Use this compass
Navigate the text according to what / how / why / what structure generated by the model builder. This structure will be described further below. In FIG.
By clicking on the blue arrow 603 pointing east, the hyperlink 601 “Design a killer application” can be navigated in the “how” direction. The destination is another piece of text in the document.

The hyperlink 601 is also not mentioned in the supplied text.
It also corresponds to a noun phrase / verb phrase pair in a KMC that has a "why" noun phrase / verb phrase pair. This is indicated by the red outline arrow 604 pointing west. Following this link, the user can jump to a page listing the "design killer application" noun phrase / verb phrase pairs of the "why" noun phrase / verb phrase pair for further knowledge navigation.

By clicking on the center 605 of the compass, the user is taken to a map view of the text as shown in FIG. This view displays a "site map" of natural language text, allowing the user to quickly and efficiently absorb and navigate the text.

Synthesis Next, the complete system shown in FIG. 4 will be discussed. User is homepage 4
Enter the system via 06. Home page 406 can be implemented as a display using proprietary software, or otherwise as a hypertext page on the World Wide Web. The home page presents the user with the option of browsing the link knowledge model 407 directly or searching for the model 407 using directional search as described above. Directly browsing the link knowledge model 407 is described in Davies.

When a directional search is invoked by a user, the parameters of the user's query are passed to the compass search agent 401, which operates as described above in the section on directional search capabilities. Thus, the compass search agent 401 connects to the local link knowledge model 406 and, if desired, other search engines 407 to implement the desired level of functionality.

The knowledge research agent 402 can also connect to the compass search agent 401 and operate, for example, in the background. According to the result of the search performed by the compass search agent 401, the knowledge research agent 40
2 can continuously increase the available knowledge in the link knowledge model. In addition, the knowledge exploration agent 402 may be connected to receive input from the natural language parser 403, which includes new queries and information, including queries issued and results returned by the compass search agent 401. Receive input from source.

Interface agents 404 to other products and builders 405 may also be provided to accommodate various input types. Interface agents 404 for other products have been described above with respect to surveying responses to e-mail queries, for example. The builder 405 is a conventional product for manually constructing the link knowledge model 407 and is described in Davies and has been described above.

Underlying Data Structure The underlying data structure is similar to that described in Davies. Generally, three types of associations are supported.

Level 1 knowledge: Relationship between things (noun phrases). However, there is no knowledge direction. For example,
Customer-to-order relationship.・ Level 2 knowledge Relationships between things (noun phrases). However, there is a knowledge direction (why and how). For example, customer-why? -Cash. • Level 3 knowledge The relationship between processes that defines why-how. For example, bill a customer-why? -Collect cash. This is the current level at which HyperKnowledge remembers knowledge.

The output from the natural language parser can be stored by various mechanisms. One suitable mechanism is a relational database, but alternative storage mechanisms can be employed. The output is a structure similar to KMC and represents the link between the output link knowledge model and the original natural language parser input text.

However, this structure also includes an association between the candidate noun phrase / verb phrase pairs generated by the model builder and the noun phrase / verb phrase pairs in the KMC. A percentage of match accuracy is stored for each association.

Knowledge Model Corpus Structure The KMC structure is implemented as a large-scale relational database. The relational database consists of a number of relational tables, each relational table holding information about an individual note in the corpus. However, the KMC structure is not limited to the structure described below, as more relationship tables can be added to accommodate new note types.

The highest level relationship table lists all noun phrase / verb phrase pairs. Each noun phrase / verb phrase pair has a unique key. This key identifies individual noun phrase / verb phrase pairs and allows cross references to entries in other relationship tables. The numbers in each field are foreign keys in other relational tables.

[0082]

[Table 2]

Each noun phrase / verb phrase pair is divided into a noun phrase and a verb phrase. As a result, 2 more
One relation table is generated, and each relation table stores the structure of each phrase. By subdividing these phrases, two more relationship tables hold individual verb phrase words and noun phrase words. Individual words in these tables are linked to other relational tables, including those that store parts of speech, semantic markup, base forms, word frequencies, syntactic structures, and more types. For each entry in the individual word table, the position of that word in the original corpus text is also stored.

The original corpus text is listed word by word as a single relational table. Each word in this table is also linked to another relational table that stores the various note types mentioned in the previous paragraph.

Other relationship tables include those that store connections between noun phrase / verb phrase pairs and those that store connections between words. Each noun phrase / verb phrase pair stored in the main relationship table has a link to a relationship table that stores a "how" connection, and a link to a relationship table that stores a "why" connection. Also has. Each of these relationship tables stores uniquely identified ties and respective "how" and "why" noun / verb phrase pairs. A relational table that stores the syntactic structure of the text is also defined. It stores the types of links between individual words, subject links or object links. An example of a connection table of noun phrase / verb phrase pairs is shown below for the "how" connection.

[0086]

[Table 3]

In the above table, it can be seen that relation 4 defines the tie between the two noun / verb phrase pairs listed in the first table above.

Conjunctions are also stored in a single relationship table. A collocation is defined as the repetition frequency of a word pattern, that is, the frequency with which two words appear next to each other a certain number of times in the corpus.

The present invention has been described above with some specific embodiments. However, many modifications that are considered to be within the scope of this invention will now be apparent to those of ordinary skill in the art. Therefore, the scope of the present invention should be limited only by the appended claims.

[Brief description of drawings]

[Figure 1]   FIG. 5 is a schematic diagram of a compass diagram used to show connections between information, including knowledge.

FIG. 2 is a screenshot of a user interface for exploring a knowledge base.

FIG. 3 is a screen shot of a user interface presenting search results in a navigable format.

[Figure 4]   1 is a block diagram of a system in which an embodiment of the invention is incorporated.

[Figure 5]   FIG. 3 is a block diagram of a natural language parser implementing aspects of the present invention.

FIG. 6 is a portion of a screenshot displaying a user interface for showing relationships between information in input text including knowledge.

[Figure 7]   FIG. 5 is a diagram of knowledge relating information in a structure representing knowledge.

─────────────────────────────────────────────────── ─── Continuation of front page (51) Int.Cl. ⁷ Identification code FI theme code (reference) G06N 5/04 550 G06N 5/04 550Q 570 570C (81) Designated country EP (AT, BE, CH, CY, DE, DK, ES, FI, FR, GB, GR, IE, IT, LU, MC, NL, PT, SE, TR), OA (BF, BJ, CF, CG, CI, CM, GA, GN, GW, ML, MR, NE, SN, TD, TG), AP (GH, GM, KE, LS, MW, MZ, SD, SL, SZ, TZ, UG, ZW), EA (AM, AZ, BY) , KG, KZ, MD, RU, TJ, TM), AE, AG, AL, AM, AT, AU, AZ, BA, BB, BG, BR, BY, B , CA, CH, CN, CR, CU, CZ, DE, DK, DM, DZ, EE, ES, FI, GB, GD, GE, GH, GM, HR, HU, ID, IL, IN, IS, J P, KE, KG, KP, KR, KZ, LC, LK, LR, LS, LT, LU, LV, MA, MD, MG, MK, MN, MW, MX, MZ, NO, NZ, PL, PT, RO, RU, SD, SE, SG, SI, SK, SL, TJ, TM, TR, TT, TZ, UA, UG, UZ, VN, YU, ZA, ZW

Claims (22)

[Claims] 1. A system for storing and retrieving knowledge, comprising a text analyzer having an input for receiving information in natural language and producing noteworthy information at an output; A parser having an input for receiving information, generating a link knowledge model of said information and storing it in memory; and a user interface for presenting at least one complementary pair to an operator, said operator having said link knowledge model A user interface that can be used to navigate through and retrieve knowledge. 2. The knowledge exploration agent of claim 1, further comprising a knowledge exploration agent having an input for receiving the link knowledge model and an output for generating a query adapted by the knowledge exploration agent to an external search engine. system. 3. The system of claim 2, wherein the hyper-knowledge agent automatically augments the link knowledge model with new information found by executing the query to the external search engine. . 4. The system of claim 1, wherein the parser further comprises a rules engine that executes a plurality of rules defining a structure of knowledge embedded in information expressed in natural language. . 5. The link knowledge model comprises linked noun phrase / verb phrase pairs, and the plurality of rules describes how a natural language sentence is linked noun phrase / linkage model in the link knowledge model. The system of claim 4, wherein the system defines if it is related to a verb phrase pair. 6. The system of claim 5, wherein the rules include at least one semantic rule. 7. The system of claim 5, wherein the rules include at least one syntactic rule. 8. The system of claim 5, wherein the rules include at least one context rule. 9. A method for retrieving knowledge, the operator comprising: a selector used by an operator to indicate one of a pair of complementary relationships; and a query input used by the operator to indicate query text. Presenting, and combining said indicated one of said complementary pairs with said query text to form a query and retrieving knowledge by said query. 10. The method of claim 9, wherein the selector defines four directions characterized by pairs of complementary directions, each direction representing a relationship. 11. The method of claim 10, wherein the selector is presented in a compass-like graphic. 12. Storing knowledge in a link knowledge model that includes noun phrase / verb phrase pairs; searching the link knowledge model for a first noun phrase / verb phrase pair defined by the query text. And searching the link knowledge model for a second noun phrase / verb phrase pair having the indicated one of the complementary pairs for the first noun phrase / verb phrase pair. 10. The method of claim 9, further comprising: 13. The method of claim 12, wherein searching further comprises parsing the natural language query text into searchable Boolean expressions. 14. A method of storing knowledge, wherein a natural language document is parsed into a linked knowledge model comprising linked noun phrase / verb phrase pairs, the knowledge level represented by the linked knowledge model. Identifying a portion of the link knowledge model that is low compared to other portions, and performing a search of external information for additional knowledge related to the identified portion of the link knowledge model. A method characterized by the following. 15. The identified portion of the link knowledge model is a noun phrase /
Including a first noun phrase / verb phrase pair that lacks a second noun phrase / verb phrase pair related by at least one of two complementary relationships defined between the verb phrase pairs. 15. The method of claim 14 characterized. 16. A software product comprising a machine-readable medium encoded with a series of software instructions, the software instructions, when executed, a selector used by an operator to indicate one of complementary pairs. Presenting to the operator query input used by the operator to indicate query text, and combining the indicated one of the pair of complementary relationships with the query text to form a query, A software product characterized by instructing it to perform a method including retrieving knowledge by query. 17. The software product of claim 16, wherein the selector defines four directions characterized by pairs of complementary directions, each direction representing a relationship. 18. The software product of claim 16, wherein the selector is presented in a graphic resembling a compass. 19. Storing knowledge in a link knowledge model that includes noun phrase / verb phrase pairs; searching the link knowledge model for a first noun phrase / verb phrase pair defined by the query text. And searching the link knowledge model for a second noun phrase / verb phrase pair having the indicated one of the complementary pairs for the first noun phrase / verb phrase pair. The software product of claim 16, further comprising: 20. The software product of claim 16, wherein searching further comprises parsing the natural language query text into searchable Boolean expressions. 21. A software product including a machine-readable medium encoded with a series of software instructions, the software instructions, when executed, a natural language document, with associated noun phrase / verb phrase pairs. Analyzing to a link knowledge model including; identifying a portion of the link knowledge model whose represented knowledge level is lower than other portions of the link knowledge model; and identifying the identified portion of the link knowledge model. A software product characterized by instructing a method to perform a search of external information for additional knowledge related to a part. 22. The identified portion of the link knowledge model is a noun phrase /
Including a first noun phrase / verb phrase pair that lacks a second noun phrase / verb phrase pair related by at least one of two complementary relationships defined between the verb phrase pairs. 22. The software product of claim 21.