JP5682031B2 - Method and system for generating text - Google Patents

Description

  The present invention relates generally to methods and systems for generating text, and more particularly to, but not limited to, methods and systems for generating syntactically correct text for reports.

  The exponential increase in computer capacity, including processing speed and memory capacity since the mid-20th century, has dramatically increased the usefulness of using computers in every part of society, right in our daily lives. One of the major uses of computers is the generation and storage of increasingly large amounts of data. However, raw data by itself has limited value. In most cases, the true value can only be acquired when the raw data is interpreted by someone and the necessary understanding and insight is provided. This interpretation process is a value-added process that converts “data” into “knowledge” and often converts it into “judgement”. This knowledge or judgment is often expressed in text reports.

  Computerized processes help to extract, contrast, and store both numerical and textual data, but the ability to effectively interpret this data by humans or computers is a large amount of data and associated complexity. May be limited by gender.

  For humans, the ability to make decisions to accurately interpret the body of data in a timely manner is that the data has been preprocessed and sufficiently reduced so that significant features are evident. Necessary.

  In the case of rule-based expert systems, there is a further related requirement that each rule be as general as possible to avoid the proliferation of rules required to take into account all the peculiarities of large and complex data sets. Exists. Since more general rules are built using higher level abstractions from the data set, variations in the details of the underlying data do not necessarily invalidate those rules. These higher level abstractions are just significant features used by human experts who build rule-based expert systems.

  That is, just like a human expert, an expert system reduces complex data to a form that allows inferences based on a smaller set of significant features rather than a larger set of raw data values. Need to be done.

  Therefore, finding a way to reduce the data complexity of the data to be interpreted by preprocessing the data into a smaller, less complex set of significant values that can be presented to a human or computer for subsequent interpretation. Is an issue.

  There are two main factors that contribute to data complexity.

  The first factor is the exact number of data item values that may need to be interpreted, i.e., there are a large number of elements that need to be analyzed in a given system.

  For example, to create a patient test report for a referring physician, a laboratory pathologist must interpret the results of hundreds of protein biomarkers used in a diagnostic instrument that analyzed a patient's blood sample. There is.

  The second factor that spurs complexity is the size of the individual data values themselves and possibly the unstructured ("free form") format. A single numeric value or enumerated value (ie text code) by itself has a clear relationship between this “atomic” value and the corresponding data item, eg a troponin value of 3.4 mmol / L. So it may be relatively easy to interpret.

  However, large free-form text may contain ambiguous expressions, typographical errors, abbreviations, two or more data values, or one of many different possible interpretations of the same data value, which makes the interpretation much more Make it difficult.

  For example, to create a patient test report for a contracted physician, the laboratory pathologist must interpret the machine-generated test results in the context of the patient's long textual history provided by the contracted physician. It may not be. The medical history is complex because it is a large unstructured data item, and relatively small variations in the text can completely alter the resulting interpretation. For example, the abbreviations “DM” (known diabetes mellitus), “FH DM” (family history of diabetes mellitus), “? DM” (suspected diabetes), “not DM” (not diabetes) ) All change the pathologist's interpretation of a given set of glucose test results. Synonyms in clinical memos ("DM", "Diabetic" (diabetic), "Diab" (diabetes), "Diabetes noted")), typographical errors ("Diabetes Melitis"), word order changes ("? Note also that any “DM”, “DM?”) Must be understood by the pathologist when interpreting.

  The medical history may also include the phrases “on Zocor” (during Zocor administration) or “on lipid lower treatment” (during administration of lipid-lowering drug), both phrases (phrases) taken by the patient taking some heart medication Represents a second concept that tells the pathologist whether or not This kind of phrase also affects the pathologist's interpretation of the test results and the reports produced for the referring physician.

  Looking at the specific example “DM, on Zocor”, there is no clear association between the “medical history” data item and the atomic value. Rather, the history as a complex data item implicitly includes two simpler atomic data items, such as “diabetes (yes)” and “therapeutic drug administration (yes)”.

  Another example of this second type of complexity due to the size of data item values and the lack of structure is that while the first laboratory performs "internally" some of the patient tests , When sending blood samples to a second laboratory for some more specialized tests. The second laboratory returns test results in a text report. From the perspective of the first laboratory pathologist, the report received from the second laboratory is a complex data item. In addition to this report, the pathologist must also interpret the results of the first laboratory in order to produce a final report for the referring physician.

  Another example of a clinical field with complex data is the allergy field. In this area, to identify one or more relevant allergens, potential allergens need to be tested in blood samples and then matched to symptoms in a free-form patient history that may be lengthy. Other examples include contagious diseases (identification of pathogenic bacteria) and multi-organ diseases (eg, identification of neurological, endocrinological, and oncological distant causes).

  Similar difficulties in interpreting complex data include fraud detection (eg, reissue of tickets, driver's licenses and passports, credit card purchases, and e-commerce), logistics audits, inventory management (eg, It also occurs in non-medical areas such as serial number management (for counterfeit detection or for product recall purposes) or IT support services.

  In the case of airline fraud detection, a number of events are recorded, including unstructured or semi-structured data regarding ticket sales and passenger flights, and then the correct price for the specified ticket is recorded. To identify whether it has been applied, it needs to be matched against pricing fare tables and other criteria for ticket reissue. This is a cumbersome task. This is because the information contained in the fare table and ticket is unstructured or only semi-structured, and is each set actually in accordance with the conditions expressed in the fare table? This is because it is individually interpreted by human experts to determine whether or not.

  In order to allow efficient and accurate interpretation by human experts, the complex data in the fare table (in this example) needs to be reduced to a set of conditions applicable to a particular ticket. The relevant characteristics (departure city and destination city, boarding date, boarding class, price) of the ticket also need to be extracted. Once the fare table and ticket data are pre-processed to these significant features, a human expert can make a decision as to whether there is a fraudulent or incorrect ticketing event.

  Real estate assessment work is another area where complex data interpretation is required. In this area, the interpretation required is an assessment consisting of the amount and the underlying explanation. Interpreted data consists of a variety of complex and disparate data, including house and land size, house orientation, nearby zip code and recent assessment, or other comparable real estate. Free-form text memos that describe various properties of a property (eg, a view blocked by an adjacent high-rise apartment) may contain important factors that affect the assessment and therefore require interpretation.

  Another example of a non-medical field that requires complex data interpretation is the field of IT support services. Imagine an online transaction processing system where companies provide regular value-added output to contract customers, such as news feeds or other reports.

  The reliability of a company's online transaction processing system is essential to the performance of this service. To achieve a very high level of reliability, the system must be continuously monitored for all factors that can affect the reliability of the system.

  These factors include alarms and warnings generated by the operating system and alarms and warnings generated by the transaction processing application itself, in addition to resource usage such as transaction rate, user activity, memory, disk, and CPU. The standard way to record these factors is to continuously log all of this information to a central facility, such as a log file, where the information is collected by the company's IT support staff. It can be analyzed regularly. The goal is for IT support staff to respond to any serious alerts or anxious trends recorded in the log file before the online transaction system fails.

  Since log entries are often generated by various operating system or application system components belonging to different manufacturer products, they are not formatted by the universal coding system and are essentially free text. In the case of a large online transaction processing system, the log file can be very large, for example several tens of megabytes per day, which is beyond the range that IT support staff can inspect manually. In addition, certain classes of alerts may require immediate action, in which case the alert and the corresponding repair action determination may need to be quickly confirmed.

  To allow efficient and accurate interpretation by human experts, as in the previous example, the complex data in the log file must be It needs to be preprocessed to a set of significant features (eg, alarms and trend conditions) that are the basis for making decisions.

  Computer-based expert systems attempt to mimic human interpretation processes. For example, RippleDown is a computer-based instructor taught by field experts how to do a highly specific interpretation on a case-by-case basis, as described in US Pat. No. 6,553,361. It is an expert system (decision engine).

  Like human experts, rule-based expert systems need to ensure that data is presented to the system with respect to those features so that the system can infer from the relevant significant features. If you infer from complex raw data (for example, data in the fare table and the ticket itself), not only can you handle the number of specific rules that are needed, but once you build the rules, In any case, any new variation encountered in the fare schedule or ticket will fail to interpret.

  In high transaction environments, expert systems can play an important role in enhancing human expertise and providing rapid interpretation of raw data. For example, a pathology laboratory may need to provide interpretation reports for tens of thousands of patients per day, far exceeding the human capacity of a small number of pathologists that can be employed at the laboratory. is there.

  However, the ability of expert systems to interpret data is limited by the same factors that limit human professionals, namely data complexity. Complex data is preprocessed in a format that allows rules to be built using higher concepts than those used by human experts, avoiding the proliferation of rules and report definitions that would otherwise occur There is a need to.

  In the following, two more detailed and specific examples of data complexity issues are presented.

  The first more specific example is in the field of medical pathology. In this field, complex investigations usually performed by specialists such as pathologists often require a large number of tests. Interpretation of test results is often difficult and requires the skill of an expert or expert system. An expert or expert system generates text that is contained within a report that contains useful analysis and interpretation of the test results, sometimes in a highly summarized form, and the report is specialized for interpreting the raw test results themselves. Sent to a referring physician (eg family doctor) who may not have knowledge. Until now, knowledge bases (knowledge bases) of expert systems have been established in the field where examinations are relatively independent of each other. For example, a knowledge base for thyroid reports primarily considers thyroid function test results (ie, TSH, FT3, and FT4). Other patient demographic data, such as age and gender, generally need to be considered, as well as findings recorded in clinical memos from physical examination or dictated medical history. In addition to referring to these individual tests and their values, reports generated using these knowledge bases also provide diagnosis and often provide treatment and follow-up test recommendations. In general, in these areas, there are less than 20 tests to consider, as well as patient demographic data such as age and gender, and findings in clinical notes provided by the practitioner. Because test results may interact, they may be related to some extent (eg, if one test is abnormal, another test is likely to be abnormal), and the tests to consider If the number of test interactions is small, the rules in the knowledge base can refer to the individual test results themselves and still maintain their generality. That is, the inspection result need not be reduced to a smaller set of significant features by any pre-processing step prior to interpretation.

  Specific rules composed of text comments given under certain conditions can be written by considering each individual test result or by considering a relatively small number of significant combinations of test results . For example, in the case of a group of thyroid tests, if the TSH test results have increased, a comment “consistent with primary hypothyroidism” may be generated.

Traditional clinical fields, such as the thyroid example above, have only a small number of Attributes. However, for newer clinical areas where studies that could be hundreds or even thousands are expected, it will not be possible to apply specific rules to each type of study. For example, a practitioner may require many food allergy tests, such as peanuts, soy, milk, wheat, and eggs. If soy and milk return very high positive values (eg, 24.3 and 30.1, respectively) and other tests are negative, the pathologist returns a “very high result” Detected against milk (30.1) and soy (24.3) "
Would like to include comments like

The rules that allow interpretation of test data to give this comment are:
10 ≦ Milk ≦ 50, indicating very high result 10 ≦ Soy ≦ 50, indicating another very high result Milk> Soy, indicating that the milk value should be before soybean in the report Peanuts = 0
Wheat = 0, egg = 0
Based on.

In this simple example where at most five allergens are tested, the number of combinations of the above comments is 2 ⁵ = 32 (ignoring the order of importance). Different rules need to exist for each combination of inspection results.

  Even with this simple comment, it is clearly not practical to individually define each of the 32 possible combinations of this comment and the corresponding rule. And the real world example is much more complex than this.

  In the case of an allergy knowledge base, there are literally hundreds of possible tests that can be performed in the study, each measuring the same chemical (IgE), and the value of each test indicates the patient's response to a particular allergen. When there are hundreds of tests in a survey, it is impossible for an expert to define the interactions that can occur between test results and provide the large number of comment variations that an accurate report requires. Before an interpretive knowledge base can be defined, the data complexity in this area must be significantly reduced.

However, the computational challenge of generating reports that take into account highly complex data is beyond the capabilities of conventional expert systems. For example, if there are 400 tests, there are 2 ⁴⁰⁰ possible combinations of test results, even if each test has only two types of output, such as “positive” or “negative”, and each combination is Need unique report text conclusions that are pre-generated and stored on the computer system. It does not even take into account other possible inputs such as clinical memos, which can occur between test data and the situation is greatly complicated. Conventional approaches that attempt to interpret complex data are not feasible if there are hundreds of observations or more.

  In a clinical environment, the various cases and corresponding reports can be enormous even if the number of tests is not so large, especially if patient history information and clinical notes are also considered.

  A second more specific example is an airline where the ticket may be issued directly by the airline or indirectly through a travel agent, ticket dealer, or online travel website. It is a ticket issuing application. In case of needing to reissue a ticket (for example, due to a travel plan change or exchange of a lost or damaged ticket) in light of the fare table (documents of the period and conditions affecting the ticket) and Details of the original transaction need to be verified against the original transaction details (eg, amount paid, number of tickets purchased, transaction currency, passenger name, purchase date and location). What is particularly difficult is that the air fare table is a complex text data item. The air fare table does not follow any clear format, but nevertheless contains certain important information. This information is often expressed as a number of key terms (key terms) such as “cancel”, “before travel”, “lost ticket”, etc., and also as monetary values and dates. Within a single fare schedule and between multiple fare schedules, each key term may appear in various forms. For example, “free of charge” (free), “foc”, and “no penalty” (all without penalty) all mean the same thing.

  Each fare table not only includes key terms, but also specifies specific information such as penalty for cancellation before boarding, penalty for lost tickets, and so on. Each of these key concepts (Key Concepts) is expressed in a variety of different ways using key terms.

  Therefore, in the above example, it is necessary to analyze a block of free text containing relevant information expressed in various ways, and then analyze the information from the free text along with other data to reach a conclusion. is there. Similar problems arise in the context of medical diagnosis (clinical notes may contain important information expressed in free text and must be interpreted in conjunction with pathological examinations and demographic data).

Difficulties in interpreting blocks of free text include:
(A) Difficulties in extracting these significant features from a block of free text so that rules can be constructed using one or more significant features.
(B) The knowledge base is difficult to handle a small number of variations of free text blocks. If the text data in a free text block is not exactly the same as the text for which the rules were built, those rules may not be generic enough to still apply to new free text blocks.
(C) The knowledge base has difficulty handling different representations of significant features themselves within a single free text block or between multiple free text blocks.
(D) The need to build rules based on a free text block that contains multiple key terms and possibly some higher-level key concepts. A “key concept” is a significant feature embedded within free text that is used by an expert or expert system when interpreting. The key concept is a higher-level unique code that represents a series of key terms (a single key term sequence). Several variations of key term sequences may be mapped to a single key concept.

In summary, traditional computer application expert systems used to mimic the human interpretation process when interpreting data, when used to interpret complex data, have many constraints including: Receive.
(A) If a very large number of data values need to be considered to reach a conclusion or express a judgment (eg final diagnosis), the rules that manipulate the interpretation process become overly complex and unwieldy, It is difficult to interpret very large data values.
(B) It is difficult to handle large unstructured data item values, and as a result, such complex data cannot be interpreted. Shrinking complex data items into canonical forms that allow simpler, atomic data items and values to be extracted and used in rules and conclusions is a cumbersome process and maintains a knowledge base This will cause long-term difficulties.

  Thus, conventional expert systems are constrained in interpreting increasingly large amounts of complex data and converting such data into knowledge or judgment (knowledge or judgment expressed in text reports). A large number of complex, including numeric and textual data obtained from different sources and presented in a variety of formats, including free-form text or alternatively structured text, such as in a “synoptic” report There is a need for a computer application method and system for generating text (such as text reports) capable of interpreting data.

  It is an object of the present invention to provide a method and system that resolves the above-mentioned limitations of conventional expert systems in interpreting complex data and converting such data into knowledge or judgment expressed in text reports. That is.

According to a first aspect of the present invention, a method for generating information from a plurality of data items is provided. This method
(A) populating an aggregate data item using at least one of a plurality of data items;
(B) generating information using aggregated data items, and
An aggregate data item is a form of a derived attribute (derived attribute),
A derived attribute uses an expression so that one or more superordinate concepts can be extracted from the data items so that a condensed amount of more relevant data can be considered when generating the information. A data item constructed from the plurality of data items,
The method of generating information is performed by a decision support system,
The generated information is stored in the following groups i. Text information,
ii. Belongs to one or more of the machine instructions.

According to the second aspect of the present invention, a system for generating information from a plurality of data items is provided. This system
(A) a preprocessor,
i. Populate an aggregate data item with at least one of the plurality of data items; and ii. A preprocessor that constructs one or more other derived attributes from multiple data items;
(B) an information generator that generates information using a derived attribute that uses the aggregated data item and other derived attributes;
An information generator forms at least part of a decision support system;
The aggregated data item is a form of a derived attribute;
A derived attribute uses an expression so that one or more superordinate concepts can be extracted from the data items so that a condensed amount of more relevant data can be considered when generating the information. A data item constructed from the plurality of data items,
The generated information is stored in the following groups i. Text information ii. Belongs to one or more of the machine instructions.

  In one embodiment, the method includes a previous step of populating each of a plurality of elements of a predetermined structure with a corresponding one of the data values. This structure associates each of the plurality of elements with an aggregate data item. The method may populate aggregated data items using at least one of the plurality of data items according to this structure.

  In one embodiment, the method includes processing the structure to determine one or more characteristics of the aggregate data item.

  In one embodiment, the structure correlates a plurality of aggregate data items, and the method includes processing the structure to determine one or more characteristics of each of the plurality of aggregate data items. Including.

  In one embodiment, the information includes text information. The information may be clinical decision support information. In addition, the information includes machine instructions.

  In one embodiment, generating information includes generating information using a knowledge base or decision support system.

  In one embodiment, populating the aggregated data item includes receiving a plurality of data items. In one embodiment, the text information may be syntactically and / or grammatically correct. The text information may be human readable. In one embodiment, the text information may form at least part of the report. The report may be associated with one or more test results. In one embodiment, each data item can correspond to one of the test results. The tests may include any of allergy tests, leukemia tests, pathology tests, blood tests, and any other type of medical test. The data item may correspond to any other information, such as gender, age, demographic information, clinical symptoms.

  In one embodiment, the aggregate data item includes a plurality of data items associated with each other. In one embodiment, populating the aggregated data item includes populating the aggregated data item by applying a rule to at least one of the plurality of data items or another aggregated data item. This rule may be a rule specific to the field. In addition, this rule may be a rule specific to the case. This rule may be for populating aggregated data items using one or more of the plurality of data items. The rule may be for populating aggregate data items using one or more of the data items above a threshold. The rule may be for populating aggregated data items using one or more of the data items below a threshold.

  In one embodiment, the method includes (a) populating one or more aggregated data items using at least one of the plurality of data items; and (b) the one or more aggregated data items. Populating one or more further aggregated data items using data items from the one or more aggregated data items by applying one or more rules; and (c) one or more further Generating information using the aggregated data items.

  In one embodiment, each of the plurality of data items is associated with an identifier and a value. Each of the plurality of data items may include its identifier and value. This identifier may be associated with the name or label of the data item. In one embodiment, generating the information includes including in the information the name or label of the data item that populates the aggregated data item. The step of generating information may include the step of including in the information a value associated with the data item that populates the aggregated data item. The step of generating information may include the step of determining the order of names or labels in the information.

  In one embodiment, generating the information includes determining one or more characteristics of the aggregate data item. The step of determining characteristics includes determining the number of data items that make up the aggregated data item, determining whether the aggregated data item is empty, and whether the aggregated data item includes a particular data item. Including one or more of determining, determining whether the aggregated data item does not include a particular data item, and determining whether multiple aggregated data items share the data item Also good. One of the characteristics may be a value.

  In one embodiment, generating the information includes including the determined characteristics of the aggregate data item in the information.

  In one embodiment, populating the aggregate data item may include populating the aggregate data item with one or more other aggregate data items. The aggregate data item and each of the one or more aggregate data items may be associated with one aggregate identifier. Each aggregation identifier may be associated with an aggregation name. The step of generating information may include a step of including an aggregate name. The step of including the aggregate name in the information may include a step of determining the order of the aggregate name in the information.

  In one embodiment, the step of populating the aggregated data item includes performing an operation on two other aggregated data items. The step of performing the operation may include one or more of a difference, a sum, and a product of two other aggregate data items. In one embodiment, populating the aggregated data item includes determining which data items that make up another aggregated data item have a value within a specific range or a value belonging to a specific discrete set. .

  In one embodiment, generating the information includes applying one or more rules to the aggregate data item. These one or more rules may form at least part of a ripple down rule knowledge system.

  In one embodiment, generating the information includes including in the information the identifier (s) of the data item (s) that make up the derived attribute.

  In one embodiment, generating the information includes including in the information the values of the data items that make up the data items.

  In one embodiment, generating information includes determining the order of data item identifiers within the information.

In one embodiment, the method further comprises building a conceptual representation of the information;
Conceptual expression is
(A) a plurality of data items;
(B) one or more derived attributes;
(C) a conclusion given by evaluating a rule of the decision support system based on one or more analyzes of one or more conclusions evaluated in a previous iteration of one or more rules, Construction is done in successive reevaluations of these rules.

  In one embodiment, the method further includes building a conclusion, where the building of the conclusion is performed in an iteration of the rule evaluation, and each successive rule evaluation utilizes the conclusion built in the previous rule evaluation. .

  In one embodiment, a system for generating information from a plurality of data items, the system comprising: an aggregate data item populator that populates the aggregate data item using at least one of the plurality of data items; and aggregate data And an information generator for generating information using the item.

  In one embodiment, the information generator is a text information generator that generates text information. Alternatively, the information generator is a machine instruction generator that generates machine instructions.

  In one embodiment, the system comprises a knowledge base or decision support system. In one embodiment, the system comprises a data item receiver that receives a plurality of data items. The information generator may be configured to generate syntactically and / or grammatically correct text information. The information generator may be configured to generate human readable text information. The information generator may be configured to generate encoded text information. The encoded text information may be machine instructions.

In one embodiment, a system for generating information from a plurality of data items, the system comprising:
(A) a preprocessor,
i. Populate an aggregate data item with at least one of the plurality of data items; and ii. A preprocessor that constructs one or more other derived attributes from multiple data items;
(B) a derived attribute that uses the aggregated data item and an information generator that generates information using other derived attributes;
An information generator forms at least part of a decision support system;
The aggregate data item is a form of derived attribute,
A derived attribute uses an expression so that one or more superordinate concepts can be extracted from the data items so that a condensed amount of more relevant data can be considered when generating the information. A data item constructed from the plurality of data items,
The generated information is stored in the following groups i. Text information ii. Belongs to one or more of the machine instructions.

  In one embodiment, the information generator may be configured to generate text information that forms at least part of the report. In one embodiment, the aggregate data populator may be configured to populate the aggregate data item by applying a rule to at least one of the plurality of data items.

  In one embodiment, the information generator is configured to include in the information a name or label associated with the data item that populates the aggregate data item. The information generator may be configured to include in the information values associated with the data items that populate the aggregated data items. The information generator may be configured to determine the order of names or labels within the text.

  In one embodiment, the system further comprises a builder that builds a conceptual representation of information that includes an interpretation portion, wherein the interpretation portion represents an operation on an aggregate data item that includes a plurality of data items.

  In one embodiment, the information generator is configured to determine characteristics of the aggregate data item. The information generator is capable of determining the number of data items that make up the aggregate data item, determining whether the aggregate data item is empty, and determining whether the aggregate data item includes a particular data item. You may be comprised so that one or more of them may be included.

  In one embodiment, the information generator is configured to include the determined characteristic of the aggregate data item in the information.

  In one embodiment, the aggregate data populator is configured to populate the aggregate data item with one or more other aggregate data items. The aggregate data populator may be configured to include in the text the aggregate data item name associated with the aggregate data item. The aggregate data populator may be configured to determine the order of aggregate data item names in the text.

  In one embodiment, the aggregate data item populator is configured to operate on two other aggregate data items.

  In one embodiment, the aggregate data item populator is configured to determine which data items that make up another aggregate data item have a value within a particular range. In one embodiment, the information generator is configured to apply one or more rules to the aggregate data item.

  In one embodiment, the information generator is configured to take into account the local characteristics for which the inspection was performed.

  In one embodiment, a method for generating information from a plurality of data items is provided, the method comprising one or more rules using one or more aggregated data items, each including one or more of the data items. And evaluating the result, and generating information according to the result.

  In one embodiment, the information includes text information. Alternatively, the information includes machine instructions.

  In one embodiment, generating information includes generating information using a knowledge base or decision support system.

  In one embodiment, evaluating the results of one or more rules includes using the characteristics of the aggregate data item as a basis for one or more of these rules.

  In one embodiment, a system for generating information from a plurality of data items is provided. The system includes an evaluator that evaluates the result of one or more rules using one or more aggregated data items each including one or more of a plurality of data items, and information generation that generates information according to the results. With a vessel.

  In one embodiment, the information generator is a text information generator that generates text information. Alternatively, the information generator is a machine instruction generator that generates machine instructions.

  In one embodiment, the system comprises a knowledge base or decision support system.

  In one embodiment, a method for generating information is provided, the method receiving a conceptual representation of information that includes an interpretation portion, wherein the interpretation portion represents an operation on an aggregate data item that includes a plurality of data items. And generating information from the interpretation part.

  In one embodiment, the information is text information. Alternatively, the information is a machine instruction.

  In one embodiment, generating information includes including in the information one or more names or labels associated with each of the data items. The step of generating information may include a step of including a group name of a plurality of data items in the information. Including one or more names or labels may include integrating information with a literal portion of a conceptual representation of the text.

  In one embodiment, the conceptual representation is pseudo text.

  In one embodiment, generating the information may include generating syntactically and / or grammatically correct text information.

  In one embodiment, a system for generating information is provided, wherein the system is a receiver that receives a conceptual representation of information that includes an interpretation portion, wherein the interpretation portion operates on an aggregate data item that includes a plurality of data items. A receiver, and an information generator for generating information from the interpretation part.

  In one embodiment, the information is text information, in which case the information generator is a text information generator. Alternatively, the information is a machine instruction, in which case the information generator is a machine instruction generator.

  In one embodiment, the generator is configured to include in the information one or more names or labels associated with each of the data items. The generator may be configured to include a group name of a plurality of data items in the information. The generator may be configured to integrate information with the literal part of the conceptual representation of the text.

  In one embodiment, the generator may be configured to generate syntactically and / or grammatically correct text information.

  In one embodiment, the present invention provides a computer program comprising instructions for controlling a computer to perform the method according to the first aspect of the present invention.

  In one embodiment, the present invention provides a computer readable medium providing a computer program according to the seventh aspect of the present invention.

  In one embodiment, the present invention provides a computer program comprising instructions for controlling a computer to implement a method according to the third aspect of the present invention.

  In one embodiment, the present invention provides a computer readable medium providing a computer program according to the ninth aspect of the present invention.

  In one embodiment, the present invention provides a computer program comprising instructions for controlling a computer to implement a method according to the fifth aspect of the present invention.

  In one embodiment, the present invention provides a computer readable medium providing a computer program according to the eleventh aspect of the present invention. The term “server” herein is intended to encompass any combination of hardware and software that, in part of the client-server architecture, performs services for connected clients.

  The client and server can be a single piece of hardware or separate software running on multiple connected hardware.

  Thus, in a preferred embodiment, the present invention limits the limitations of conventional expert systems by providing a means that can interpret large amounts of complex data, including numeric and textual data obtained from different sources. A computer application method and system for generating text (such as a text report) that overcomes at least some of the above. In one embodiment, the present invention further provides a means for interpreting data presented in various formats including free-form text.

  To achieve a better understanding of the nature of the present invention, embodiments of a method and system for generating text information are described below by way of example only with reference to the accompanying drawings and examples.

  Here, Example 1 is an example of a method and system for generating text according to a preferred embodiment in the form of a leukemia report knowledge base.

  Example 2 is another example of a method and system for generating text according to a preferred embodiment that takes the form of an allergy report knowledge base.

  Example 3 is an example of a method and system for generating text according to another embodiment that takes the form of an airline ticketing audit system.

  Example 4 is an example of a method and system for generating text according to another embodiment taking the form of a log file monitoring system.

It is a user interface window showing an example of a “normal form” using a text block and a text normalization attribute “NormCat”. The example shown relates to ticketing. It is a window of a user interface showing an example of a regular expression defining a list of key terms and each key term in “NormCat”. The example shown relates to ticketing. FIG. 5 is a user interface window showing two examples of comments with variables that extract currency and value from normalized text. The example shown relates to ticketing. 1 is a block diagram of one embodiment of a system for generating text or text information, such as a text report. FIG. 3 is a flow diagram of one embodiment of a method for generating text or text information, such as a text report. FIG. 6 is a block diagram of another embodiment of a system for generating text or text information, such as a text report. FIG. 6 is a flow diagram of another embodiment of a method for generating text or text information, such as a text report. FIG. 6 is a flow diagram of yet another embodiment of a method for generating text or text information, such as a text report. FIG. 6 is a flow diagram of a third embodiment of a method for generating text or text information such as a text report. FIG. 4 is a user interface window illustrating an example of a text condenser attribute (TCA) that defines a key term of an embodiment. FIG. The example shown relates to ticketing. FIG. 11 is a user interface window illustrating an example of a text summary attribute (TCA) according to the embodiment of FIG. 10 in which a key concept is defined along with a key term. The example shown relates to ticketing. User interface window for text summarizer attribute (TCA) in FIG. 10 that gives sample cases values for themselves (“TCA”) and values for key concepts “CxBt”, “CxAt”, “RiOb1” and “RiRtc” It is. The example shown relates to ticketing. FIG. 11 is a user interface window of the TCA in FIG. 10 showing an example of a matching form that defines evaluation of a key concept. The user is prompted to provide an example of raw text for each matching format. The example shown relates to ticketing. FIG. 11 is a window of the user interface of the TCA of FIG. 10 showing an example of a matching format for a derived match. Because keywords have been added, the matching format no longer matches those examples. The example shown relates to ticketing. In the TCA user interface window of FIG. 10 showing how the matching format needs to be changed so that the matching format matches the normalized version of those examples when new keywords are added. is there. The example shown relates to ticketing. FIG. 11 is a window of the TCA user interface of FIG. 10 showing how the key concept can be used directly on the variables in the comments. The example shown relates to ticketing. FIG. 11 is a TCA user interface window of FIG. 10 showing how a keyword name change from “BT” to “BeforeTravel” automatically updates the collation format. The example shown relates to ticketing. Figure 5 is a user interface window of one embodiment of a TCA that extracts dates and Boolean values. FIG. 19 is a user interface window of the embodiment of FIG. 18 showing exemplary Boolean and date values for the key concept. FIG. The example shown relates to ticketing. FIG. 4 is an exemplary window of a user interface of an embodiment of a TCA according to the present invention in which a derived attribute and a portion of the raw text it represents are provided to a user as a tooltip. FIG. 6 is a schematic diagram of an embodiment of a hierarchical relationship between data items and aggregated data items.

  Table 1 is a dictionary of terms defined according to the present invention. The terms defined in Table 1 are shown using capital letters throughout this document. Where capital letters are not used in a term, they should be interpreted in their ordinary meaning unless otherwise indicated.

  In a preferred embodiment, the present invention provides at least one of the limitations of conventional expert systems by providing a means capable of interpreting large amounts of complex data including numerical and textual data obtained from different sources. The present invention provides a computer application method and system for generating information (such as a text report) that eliminates a part. In one embodiment, the present invention further provides means for interpreting data presented in various forms, including free text analysis means that allow interpretation of data presented in free form text.

  FIG. 4 is a block diagram of one embodiment of a system for generating text from a plurality of data items, indicated generally by the number 1. The system 1 may comprise any system capable of processing information, and in this embodiment has a computer program that resides on the computer readable medium 2 and includes instructions for controlling the central processor 4 of the system. It can be described as a computer system 1. These instructions are for implementing a method 500 for generating text information, such as information in a human readable text report, from a plurality of data items. A flow diagram of the method 500 is shown in FIG.

  Alternatively, the generated information is not text information presented as a report, but one or more machine instructions, and the components of the system 1 are changed accordingly. It should be understood that the term “textual information” should be construed in a broader sense from now on to encompass this alternative embodiment, where appropriate.

  Referring to FIG. 4, a computer readable medium 2 includes a non-volatile memory 2 in the form of a hard drive disk 2 connected to a processor 4 by a suitable bus 6 such as SCSI. In some embodiments, the non-volatile memory 2 includes, for example, a flash memory, CD, DVD, or USB flash memory unit.

2. One or more data items 8 are received via a data receiver 10 that is part of the communication interface with the other system or user from which the data items originate. The receiver
(A) Data items (b) Derived attributes (c) One or more of the conclusions can be received.

  Each data item 8 represents input data to be processed, such as the results of one or more examinations of a survey, or any other simple or complex data that requires processing.

  Such processing can be executed by a builder that constructs a conceptual representation of information including an interpretation part, where the interpretation part represents an operation on an aggregate data item including a plurality of data items. To do. In some embodiments, the source of these data items is an information system 37 that is external to one.

The generated text information 26 is via a data transmitter 11 (referred to as a transmitter) that is part of a communication interface with other systems or recipients (such as one or more users) that require text information. Sent. In some embodiments, the text information destination is an information system 37 external to one. The transmitter sends the generated information
(A) Machine (b) Send to one or more of the recipients.

  In some embodiments as shown in FIG. 6, the system 3 is an embedded system. The components of FIG. 6 that are similar to the components of system 1 of FIG. 4 are similarly numbered. The embedded system 3 in this embodiment forms part of a device that performs a test such as a medical test. It should be understood that any suitable architecture can be used, such as a terminal / mainframe, client / server, cloud computing, etc., as well as the illustrated architecture.

  In the embodiment shown in FIGS. 4 and 6, the computer readable medium (eg hard drive) 2 is for defining aggregate data items, other derived attributes, and rules for generating text information. Holds computer instructions.

Generally speaking, a “derived attribute” is a data item that does not exist in the original data presented to the knowledge base for interpretation, but is constructed from this data using some expression. This formula is based on the following considerations: (a) the identifier of the data item or other derived attribute (b) the value of the data item or derived attribute (c) multiple instances of the data item or derived attribute presented below i . Time-based sequence ii. Based on any other vector format.

  The aggregate data item is an example of a derived attribute. The original (or “primary”) data is presented as a mapping from data items to value pairs. When historical data is considered, the original data is presented as a map from a data item to a time-based sequence of values for that data item. A data item in the original data is called a “primary attribute”.

  Derived attributes are a higher level concept that can be used more naturally and generically than primary attributes in rules and reports. For example, the primary attribute may be the name of the referring doctor. A more useful derived attribute may be a derived attribute “specialist” having the value “true” if the name of the referring physician matches a name on the specialist physician list. Another example is when the primary attributes are patient height and patient weight. A useful derived attribute may be an attribute “BMI” consisting of a numerical value evaluated as a ratio of weight to square of height.

  With reference to FIGS. 4 and 6, the system 1 is configured to process the data items 8 by performing the method of generating text shown in FIGS. In addition, the data item 8 may be related to an arbitrary specialized field such as real estate assessment. Data items 8 for real estate “inspection” or evaluation include, for example, house and land size, house orientation, nearby postal code and recent assessment, or other comparable real estate. Other examples of specialties include one or more of fraud detection, bone mineral concentration reports, medical alerts, genomic reports, molecular reports, and allergy reports. The systems 1, 3 and method 500 described herein may be configured to pre-process such data items 8.

  In the exemplary embodiment of FIG. 4, the system 1 has a data receiver 10 that receives a data item 8 that is then stored on a hard drive (or other computer readable medium) 2. Or may not be stored. In embodiments where testing is performed at a location remote from the system 1, for example, at a remote site 12, the system 1 may be configured to be connected to a network 14 to which the remote site 12 is also connected. The network 14 may be a wide area network such as the Internet or the cloud, but the remote site 12 may be very close, for example, a room next to the system 1, in which case the network 14 is local It can be an area network or a wireless network such as WiFi or WLAN. In addition, in the example shown in FIG. 6 in which the system 3 is a part of the inspection device 5, the data receiver 10 may operate as an interface between the processor 4 and the data source 22. An example of a data source 22 is the system 3 sample inspection device, which performs a physical, chemical, biological inspection or other analysis on the sample.

  The processor 4 (FIGS. 4 and 6) populates the aggregated data item 24 using at least one of a plurality of data items 8 stored on the hard drive 2 (or other computer readable medium). Programmed as an aggregate data item populator. In one embodiment, the aggregated data item 24 is a data structure of some kind of memory 20 for processing by the processor 4 (eg, file, list, array, tree, record, table used in a database, flat file, or Any form of appropriate data structure, such as an indexing system). Data item 8 can also be stored in memory 20. The memory 20 in this embodiment may be a CPU register, on-die 8 RAM cache, external cache, DRAM, and / or paging system, virtual memory or swap space on the hard drive (or other computer readable medium) 2, or any other Includes types of memory. However, embodiments may have more or fewer memory types where appropriate.

  The processor 4 is programmed to be an information generator that uses the aggregated data items 24 to generate information 26 (eg, as a text report or as one or more machine instructions). The information generator 4 is configured to store the information 26 so generated in the memory 20. In this embodiment, text information 26 represents syntactically and / or grammatically correct human readable text.

  The output of the system 1, 3 is textual information, preferably in human readable form, for example text printed on a monitor or screen 28 (eg text report), printed on a paper report 33 by the printer 30. One of the text or email or other type of electronic message 34 sent by the data transmitter 11 over the network 14 to the user's workstation 32 (eg, a physician or surgeon's computer) or another information system 37 More than one. The text information generated by the processor 4 may be text information such as some other decision support result derived from the data item 8. In one embodiment, the SMS gateway (or other SMS relay mechanism) 34 is an electronic message, such as an SMS or email, that includes text information 26 in human readable form (ie, syntactically and / or grammatically correct text). Is transmitted by the system 1 to a receiver 36 such as an electronic device. Device 36 may be a mobile phone, smartphone, PDA, or other handheld electronic device, any other computing device with processing capabilities. In one embodiment, the system 1 is configured to send an instruction to send an SMS to the handheld mobile device 36. This is advantageous when the test results are abnormal and require immediate follow-up, or when the test results are sought quickly (eg, when inspecting an air ticket).

  Referring to FIG. 5, one embodiment of a method 500 for generating text, such as information in a text report, from a plurality of data items is shown. The processor 4 (FIGS. 4 and 6) operating as an aggregate data populator is programmed to populate the aggregate data item 24. Referring to FIG. 7, another embodiment of a method for generating text is shown. This method includes a sub-step of populating an aggregate data item (labeled 24 in FIGS. 4 and 6) by applying one or more rules to at least one of the multiple aggregate data items. Including.

  The rules may form at least part of a rule-based knowledge / expert system or decision engine. One example of a suitable rule knowledge system is a proprietary system known as RippleDown disclosed in the specification of applicant's US Pat. No. 6,553,361, incorporated herein by reference. As described in that US specification, a collection of rules is a knowledge base built by experts. The rules may be specific to the field. For example, the rules may be specific to the field of allergy testing or to the field of leukemia testing. However, in some other examples, the rules are case specific rules, i.e. rules specific to a set of associated test results / data items 8. In this case, the system 1 is a knowledge base or a decision support system.

4 and 6, in one case, the data item 8 has an associated name or label portion and a value portion, for example,
Milk, 25
Soybean, 30
Peanuts, 0
become that way.

Each data item 8 is associated with an identifier (here milk, soy or peanut) and a value (here 25, 30 or 0). In these embodiments, each of the data items 8 includes an identifier and a value. The identifier is, in this example, the name or label of a data item (eg, “milk”) that can be used to generate textual information 26 (eg, see step 504 in FIG. 5). The system uses this identifier to specify the specified (a) language (if the generated information is human readable), or (b) instruction format (if the generated information is a machine instruction)
Contains rules that can be translated and / or expressed. This allows the information to be generated to be generated in a language / format appropriate to the situation or as otherwise determined by rules.

The aggregated data item 24 with the name or label “very high food allergen” includes, for example, if the rule milk> 25, the milk is included in the “very high food allergen” AND
If soybean> 25, include soybean in "very high food allergen" AND
If peanuts> 25, peanuts can be populated from data item 8 above by including them in a “very high food allergen” (see, for example, step 502 in FIG. 5).

In addition to this, the aggregated data item 24 (FIGS. 4 and 6) having the name or label “very high food allergen” includes, for example, the following data preprocessing operation (see, for example, step 702 in FIG. 7).
“Very high food allergens” can be populated from data item 8 above by applying food allergens in the range (25,100).

Generating information from multiple data items
(D) a preprocessor (preprocessing device),
i. Populate an aggregate data item with at least one of a plurality of data items; and ii. A preprocessor comprising one or more other derived attributes from these multiple data items;
(E) a derived attribute that uses this aggregated data item and an information generator that generates information using other derived attributes;
An information generator forms at least part of a decision support system;
The aggregate data item is a form of derived attribute,
A derived attribute is an expression that allows one or more superordinate concepts to be extracted from the plurality of data items so that the condensed amount of more relevant data can be considered when generating the information. A data item composed of the plurality of data items using
The information so generated is stored in the following group iii. Text information iv. Belongs to one or more of the machine instructions.

  The processor 4 (FIGS. 4 and 6) is also programmed as an evaluator that evaluates the result of one or more rules, as illustrated above, using one or more aggregated data items (eg, 24). Is done. In the above example, the text information generator 4 generates the text information for the report 33 according to the rule conclusion, for example.

Thus, the processor 4 (FIGS. 4 and 6)
(A) an aggregated data item populator that populates one or more aggregated data items 24 using personal data items 8;
(B) an evaluator that evaluates the conclusion of one or more rules applied to the aggregated data item 24; and (c) using the aggregated data item 24 (eg, as a text report or one or more machines It can function as one or more of the text information generators that generate the text information 26 (as instructions).

  It will be appreciated that the processor 4 (FIGS. 4 and 6) may examine each data item 8 before including it in the aggregated data item 24. An exemplary embodiment of a system for generating text (as shown in FIGS. 4 and 6) may include two or more processors 4 that perform the outlined functions in parallel or serially. You will understand that.

In the example outlined above, one conceptual representation of an aggregate data item (labeled 24 in FIGS. 4 and 6) having the name “very high food allergen” is
Milk, 25
Soybean, 30
It becomes.

The text information generator 4 (FIGS. 4 and 6) may be configured to include in the text information 26 a name or label associated with the data item 8 that populates the aggregated data item 26. For example, the processor 4
It may be desired to form textual information that a very high result was found for a “very high food allergen” (eg, at step 504 of the method shown in FIG. 5).

Continuing the same example, the processor 4 functioning as the text information generator 4 is:
Text information can be generated that represents the text that very high results were found for soy and milk.

The text information generator (processor) 4 has a higher value for soy than milk, so the best way to present this text is to order names or labels in the text so that soy is on top. It is determined that it is attached. Further, the generator 4 determines that “and” should be placed between “soy” and “milk” because there are only two items in this aggregated data item 24. If there is a third item in the aggregate data item 24, such as honey with a value of 26, then one of the grammatically correct texts to be generated is
Generator 4 includes machine instructions that allow generator 4 to determine that very high results have been found for soy, honey and milk.

The text information generator 4 is configured to include in the text information 26 values associated with the data items 8 that populate the aggregated data items 24, if necessary. For example, the above text would instead
Very high results may have been found for soy (30), honey (26) and milk (25).

  The above is an example of one commonly required ordering, but there may be other orders in different situations.

  Another such example is the use of lab pathologists in diagnostic instruments that have analyzed patient blood samples, such as hundreds of protein biomarkers, to generate patient test reports for contracted physicians. This is where the results may need to be interpreted. To allow such interpretation, the system that generates the text organizes the biomarker results into multiple subgroups, each of which is considered a higher level marker that has some diagnostic significance. Can do. For example, one group of biomarkers may be tested for the presence or absence of a specific BCC type of leukemia, and another group may be tested for the presence or absence of a specific type of AML of leukemia.

  Thus, the system that generates the text derives a single result from all of the biomarker results in each subgroup, for example, a single value representing the combined result of the markers of the BCC group, and the AML group. Data complexity is reduced by deriving a single value that represents the result of combining the markers. If this happens, the results of the patient's blood sample will be suitable for interpretation by a laboratory pathologist, who only needs to consider a much smaller but higher marker.

  In addition to simplifying the interpretation process, the reports generated by the text generation system and provided to the referring physician by the pathologist are simplified by using the result values corresponding to groups of markers rather than individual marker values. The A report written from the point of view of a group of markers is more concise and less subject to variation due to changes in the value of individual marker values themselves.

  The advantage of grouping markers is that, unlike when all rules need to refer to individual marker values, the expert system will be able to follow human expert interpretation processes and group value inferences. The ability to build expert systems with far fewer rules. Similarly, various reports can be written in terms of groups of markers and their group values rather than specific marker values, so those reports need to be defined by human experts. A small number of report types can still be generated by the expert system.

  Many data item values may arise from the need to view a data item historically (on a time basis).

  For example, a pathologist who monitors myocardial enzyme outcome values, such as troponin, interprets the current results against all previous results over the past few weeks to assess whether to alert the emergency response team May be necessary. The amount of data and complexity represents this rate of change of the time series and is therefore reduced by providing a new top (high level) result that summarizes the key characteristics of the entire time series relative to the current value. In that case, the pathologist can interpret the significance of the current results in the light of this higher trend result.

  In some embodiments, the generated text information does not generate human-readable text information (ie, syntactically and / or grammatically correct text), but instead one or more machine instructions. Generate text that takes the form In this case, the system includes a machine instruction generator. Machine instructions can control the workflow. For example, if the test result indicates that no allergen has been detected, the machine instructions can cause the system to automatically send the report without checking the report by a human evaluator. In addition, the machine instructions may cause additional tests on the retained samples to be performed before the report is generated, or may be instructed to perform the additional tests.

In another embodiment, the systems 1, 3 (FIGS. 4 and 6) can include a receiver 36 that receives a text report or other output. Referring to FIG. 8, another embodiment of a method for generating text allows a user 38 to enter a conceptual representation of text via a keyboard or other input device connected to the CPU. 802). The conceptual representation is stored in the non-volatile memory 2 by the system. “Conceptual representation” is an expression of a rule condition from the viewpoint of a derived attribute (derived data item) including an original data item or an aggregated data item. Using the above example, the conceptual expression entered by the operator is
It takes the form of “pseudo-text” that very high results have been found for “very high food allergens”.

  The pseudo-text in this example is a compact and informal description of the conclusion / decision based on an analysis of the individual test results being matched. The pseudo-text represents the high-level description of the text desired by the operator, but importantly omits the details that the systems 1, 3 are intended to calculate. Pseudotext is a natural language description of the details of computer processing. Pseudotext is easier for humans to create and read than a more technical description of the desired text that can be achieved using a programming or scripting language.

The conceptual expression includes an interpretation part. In this case, the interpretation is “very high food allergen”
It is.

The interpretation part represents an operation on an aggregate data item having the name “very high food allergen”. Referring to FIG. 8 (step 802), in one embodiment of a method 800 for generating text, in systems 1 and 3, user 38 inputs a conceptual representation of the text as pseudo-text that includes an interpretation portion. Upon receipt of data item 8, text information generator 4 generates text information 26 from its interpretation as described elsewhere in this document (see step 804 in FIG. 8). The text information generator 4 is configured to include in the text information 26 one or more names or labels associated with each of the data items. The text information generator 4 may be further configured to include a collective name (identifier) for a plurality of data items in the text information 26. The text information generator 4 may be further configured to integrate the text information 26 with a literal part (a verbatim part) of the conceptual representation of the text. In this example,
Very high results have been found for soy, honey and milk.

In the embodiment shown in FIGS. 4 and 6, the text information generator 4 is configured to determine the characteristics of the aggregated data item 24. For example, the text information generator 4
(A) Determining the number of data items that make up an aggregate data item (b) Determining whether the aggregate data item is empty (c) Determining whether the aggregate data item contains a specific data item You may comprise so that one or more of things to do may be included.

These are examples of operations performed on aggregated data items in embodiments of methods and systems for generating text. For example, the text information 26 is
Generated from pseudo-text such as (very high number of food allergens) found very high results for one food allergen (step 804 in FIG. 8),
Very high results have been found for three food allergens.

  Thus, the text information generator 4 is configured to include in the text information 26 information about the determined characteristics of the aggregate data item. “Number of” (“number of”) is a type of operation performed on the aggregated data item “very high food allergen”.

  The aggregate data populator 4 (FIGS. 4 and 6) may be configured to populate the aggregate data item 24 with one or more other aggregate data items. The former aggregated data item may include a plurality of related data items, for example, all foods that the patient has been found to exhibit high allergies. Thus, the aggregate data item “food” is also an aggregate data item (eg, peanuts, tree nuts. Tree nuts are also data items such as almonds, Brazil nuts, walnuts, hazelnuts, macadamia, pistachios, pecans, and cashews). May be populated with data items (e.g., nuts).

  The aggregate data populator 4 may be configured to include the aggregate data item name (identifier) associated with the aggregate data item in the text. The aggregate data populator 4 may be configured to determine the order of the names of the aggregate data items 24 in the text. In one embodiment, the aggregate data item populator 4 is configured to perform operations on two or more other aggregate data items 24. For example, one aggregated data item 24 may be a very high outcome food allergen and the other aggregated data item 24 may be a food allergen of interest. In that case, the populator 4 may generate a new aggregated data item 24, for example a very high-resulting food allergen by taking the intersection of the two aggregated data items. Other possible operators include differences, unions, and products. In another embodiment, the aggregate data item populator 4 is configured to determine which data items that make up another aggregate data item have a value within a certain range.

  In one embodiment, generating text information 26 includes including in text information 26 information about the determined characteristics of aggregate data item 24. For example, if the determined characteristic is the maximum value of the items that make up the aggregated data item, the text information is “the highest pollen allergen has the result <highest pollen allergen value> mmol / L <highest Pollen allergen>, where <highest pollen allergen> is a characteristic of the pollen allergen aggregate data item defined as the pollen allergen having the highest value, <highest The value of the pollen allergen> is the value itself.

In some embodiments, the text information generator 4 is configured to apply one or more rules to the aggregated data item 24 to control program flow (see, eg, step 702 in FIG. 7). . An example of a logical check associated with such a rule is
If Number of moderate food> 1 AND if Number of symptoms> 1 AND Number of very high food + Number of food = 0
It is.

  The workflow action associated with such a rule may be to queue test results and reports to the reviewer, rather than automatically disclosing reports to the pathologist to the referring physician.

  It can be seen that the aggregated data item can be treated as a data item for generating the text information 26 when used in the evaluation of the Boolean conditions constituting the rule. Populating the aggregated data item 24 may include populating the aggregated data item 24 with one or more other aggregated data items, each of the other aggregated data items having a name or label It may have a related aggregate identifier that takes the form Populating an aggregated data item (eg, step 502 in FIG. 5) can be done by combining two or more other aggregated data items (eg, sum or product operations), or which data comprises another aggregated data item. This can be accomplished by applying more general conditions, such as determining whether an item has a value within a certain range (eg, a pollen item within the range [20-50]).

  In that case, the aggregation identifier (name or label) can be used in the text information 26 just as the data item name is used in the text information 26. Again, the text information generator 4 may determine the order of the aggregate names in the text information 26.

Some embodiments of the system and method include a new or improved data preprocessing method that reduces data complexity prior to interpretation by a knowledge base (knowledge base), the method comprising:
(A) grouping personal data items into one or more subsets (subsets) of data (each subset group is called an aggregate data item);
(B) a statistical value (eg, maximum value, minimum value, group size, median value, average value, mode value, or any other statistical value), or other numeric value for each aggregated data item, a Boolean value, Or a process for calculating a text value (hereinafter referred to as an “aggregation” value),
(C) further executing a specified operation (eg, sum or product) performed on the collection of aggregated data items to generate other aggregated data items (eg, each of which is a collection of specific cancer markers) The union of the aggregated data items “BCLL diagnosis”, “AML diagnosis”, “BCLL support”, “AML support” may represent another aggregate data item “leukemia” consisting of all leukemia cancer markers),
(D) generating one or more data items and values from data items having values consisting of free-form text, and / or
(E) generating one or more data items and values from data items associated with a series of values.

Thereby, one aspect of the data preprocessing method considers a collection of personal data items and their values, and derivations each having a value by grouping, filtering, mapping, correlation, or other processing. Reduce the complexity of this data by generating attributes (including aggregated data items).

  Another aspect of the data pre-processing method is by examining the complex free-form text values of the data items and generating other simpler data items, each having a value, through string pattern matching and filtering processes. Reduce complexity.

  Another aspect of the data pre-processing method is to examine data items associated with a set of values and filter, trend analysis, or other analysis to provide other simpler data, each having a value. Reduce complexity by generating items. This method considers a single derived data item and its value, instead of having to consider each individual data value or a complex data item value that is a free-form text or sequence in a set of original data items. Also make it possible. Here, the “derived” data item and the value thereof indicate the data item and the value configured by the preprocessing, and include the “aggregation” data item. This significantly reduces the amount and complexity of the data values that need to be interpreted, and thus the rules and decision points needed to reach a judgment or conclusion (represented later in the generated text report) Is also significantly reduced. Aggregated data items and their values can also be used as knowledge base output, greatly reducing the complexity of the resulting report text.

In other embodiments, the system and method for generating text further includes means for interpreting data presented in various formats. Free text analysis means that allows interpretation of free text data items is also included in the data interpretation means. The free text analysis means executes a method of preprocessing free text data items, and the method is to interpret a “regular expression” in the text data as follows: Mapping to one or more of a series of keywords that allow to consider a remarkably simple "canonical" representation of the data item instead of requiring it;
(B) assigning complex text data items to a number of simpler “atomic” data items, wherein the value of each atomic data item is
i. Boolean value (eg true or false, yes or no),
ii. Finite enumeration ("a", "b", "c"), or iii. Process, which is one of the numerical values.

  By providing a new or improved method for preprocessing complex data items as described herein, the preferred embodiment overcomes at least some of the limitations of conventional expert systems, Allows the interpretation of large amounts of complex data (including numerical and textual data obtained from different sources and presented in various formats including free-form text). Preferred embodiments convert complex data into knowledge or judgment (including conclusions, results, or other findings based on interpreted data). Knowledge or judgment is expressed in the text report as text information (including machine instructions).

  Data pre-processing methods reduce data complexity to a manageable degree by filtering, grouping, mapping, and other operations. For example, if there are hundreds of protein biomarker test values to be interpreted, the filtering operation may exclude certain results that are not associated with a particular patient. The method also includes taking one or more data items and applying an expression to process those (one or more) data items into derived attributes. Derived attributes are easier to handle. This is because derived attributes extract higher-level (higher-level) more important information from the original data item, thus reducing the data to be interpreted and making it easier to handle. The data pre-processing method includes a grouping operation that groups related data items into one or more data subsets. Here, each subset group is called an aggregate data item. Continuing with the current example, the grouping operation can collect values for a particular subset of related biomarkers and calculate a statistical value, such as a maximum value, for each subset. Thus, text generation methods and systems need only consider a single data value for each group instead of having to interpret individual biomarkers, significantly reducing the number of data values to consider. To do.

  When certain data items are complex, such as a patient's text history or other text data, the mapping operation may look for patterns in the text (“regular expressions”) and map these patterns to a set of keywords. it can. Thus, instead of having to interpret a large free text data item, the method and system for generating text need only consider a remarkably simple “standard” representation of this text item. Numerous variations in medical history can result in the same simple standard representation, again allowing easier interpretation that allows interpretation using significantly fewer rules and decision points.

  Another example of a mapping is that instead of assigning a text pattern to a keyword, a complex text data item is converted to a number of simpler “atomic” data items, ie the value of each atomic data item is a Boolean value (“ Assign to data items that are true (true) or "false" (yes or no), finite enumeration ("a", "b", "c"), or numeric. An example of an atomic data item assigned from a complex medical history may be a data item called “diabetes status” having a value of “true” or “false”. Another example may be a data item called “diabetic drug” having the enumerated values “biguanide”, “meglitinide”, or “sulfonylurea”. In this way, selected important concepts included in the medical history are extracted and expressed in another standard manner.

  In all these examples, complex data is preprocessed into simpler data items to facilitate interpretation.

In one embodiment of the invention, an aggregate data populator device or tool (eg, a database structure) receives a plurality of data items. Here, each data item corresponds to one result of a plurality of examinations, for example. In this example, the results of multiple tests are
(A) Investigation of the patient's condition (eg, does the patient have a particular form of disease or allergy)
(B) an audit of a large amount of data (eg, what is required when determining whether a ticket should be reissued), or (c) a large amount of information to extract or reach a decision Used in essentially any analysis where complex data items (including enumerated data and numeric data in text reports) need to be analyzed.

  Referring to FIG. 5 and the example of protein biomarker test, test values of a plurality of protein biomarker tests are grouped into a plurality of subset data groups (aggregated data items). In other words, each aggregated data item is populated from a collection of individual protein biomarker test values (step 502).

In this embodiment of the text generation system, the device (aggregated data item populator) includes information in the form of a predetermined data structure that associates various types of data items with the appropriate aggregated data item (s). . This data structure allows the device to populate a given aggregated data item using one or more of the received data items by applying various rules for processing the received data. In other words, the aggregate data item populator populates related aggregate data items by mapping personal data items to related aggregate data items. An “aggregated data item populator” includes a set of rules that determine how personal data items should be mapped. Personal data items (including primary and derived attributes) are mapped to one aggregated data item by name, type, value, or by membership in another set. In other words, aggregate data items are populated with personal data items according to collective membership. In the current example, each data item in one of the aggregated data items is, for example, an associated biomarker for a particular disease or allergy. Using the example airfare table, each data item in one of the aggregated data items can be, for example, an associated condition for ticket reissue.

  In another embodiment (see step 902 in FIG. 9), the step of preprocessing the data includes a method of extracting data expressed in free-form text (described in more detail later in this document; step 902). . For the purposes of this section, one embodiment of a system and method for generating text is a means for extracting data expressed in various ways, including free-form text, using text condenser attributes. including. The data items so extracted are then sent to other data items (eg numeric data items related to individual test results received by the system, individual items of report / record data, eg credit cards (Expiration date and ticketing date) are processed in the same way.

  Referring to step 702 of FIG. 7, the additional aggregated data item may then be supplied with data by other rules that operate on the aggregated data item. The additional aggregate data item may include, for example, a data item having a significant value. In that case, additional rules are applied to the additional aggregated data items. An example of a rule may include determining whether the number of important data items in the additional aggregated data item has exceeded a threshold. The result of the rule may indicate a positive test result, in which case appropriate text is generated that reports a positive or other test result. Text may be generated flexibly on a case-by-case basis without using rules for each case by using aggregated data items.

  Referring to FIG. 9, a method of generating text that includes extracting data (eg, clinical memos, airfare charts, real estate advertisements) that is expressed in free-form text from different sources (step 902). Another embodiment 900 is shown. This method allows the analysis of free text blocks containing relevant information expressed in various ways. Information extracted from the free text (eg, numeric data or other information) is analyzed along with other data to arrive at a conclusion or decision (step 904). For example, clinical notes may contain important information expressed in free text and must be interpreted in conjunction with pathological examinations and demographic data.

  In the ticketing environment, the inventors' first attempt to solve the problem arising from the need to interpret free text generates a derived attribute called "Text Normalization Attribute" (TNA). It included that. TNA converts free text into a series of key terms. “Key Term” is a unique code representing a piece of free text. The key terms may include variable elements, such as currency values. Several variants of free text fragments may be mapped to a single key term. The mapping from free text to a series of key terms provides a standard representation of the free text.

  TNA allowed each key term to be defined according to its numerous forms, i.e. by another phrase of the key term. The output of the derived attribute was a string of “condensed” or “normalized” text consisting of key terms extracted from free text. FIG. 1 shows a user interface displaying a typical text block and its “normal form” defined by the TNA. A TNA is essentially a map of regular expressions to keywords, as shown in FIG.

  Related key terms were listed as a table, and for each keyword there was a list of matching regular expressions. The raw text was then converted into a list of tokens by searching for the (positionally) closest match from the current search position. Here, multiple matches starting from the same position are selected according to the match length. A built-in matcher converts a currency value, such as “AUD 75”, into a special monetary value token that can be considered a key term with variable elements.

  The normalized text was analyzed to extract the desired value (one or more), for example, the monetary value (75) of the transaction and the “value” (AUD) of the currency. The syntax used in the experiment conducted by the inventors to extract the desired value was the syntax in a proprietary ripple-down condition language that uses a text regular expression pattern matching algorithm. . FIG. 3 shows a user interface screen displaying two examples of comments with embedded variables used to extract currency and value from normalized text.

TNA was tried by building a knowledge base, where comments were a variable expression that gave the cost of reissuing a ticket for a given reason. In almost all cases,
Amount = {Amount in code matching "CX BT FOR MV $" in NormCat} Currency = {Currency in code matching "CX BT FOR MV $" in NormCat}
The conditions for adding a comment like
NormCat contained the code sequence “CX BT FOR MV $”.

  In essence, the same matching sequence had to be written three times, ie twice in the variable comment and once in the condition to add it.

  Using this text normalization process, the inventors were able to build a knowledge base that successfully analyzed most of the fare schedules studied in one country, such as Australia, but data from the most complex fare schedules Some enhancements were needed to extract the. However, there was a problem that meant it was difficult to maintain a knowledge base, especially when new keywords or key terms had to be added for another country's fare schedule. Similar problems have arisen in other situations, for example, when the interpretation of a patient's test results requires the inclusion of clinical notes from two or more clinicians.

  The problem with TNA can be explained as follows.

A. Sensitivity to changes in extracted information As a result of adding new keywords to the TNA, variables in comments and conditions in rules may no longer be evaluated as intended. For example, if the fare table is text. . . BEFORE DEPARTURE BUT WITHIN 24 HOURS OF SCHEDULED LED TIME CHARGE AUD 75 FOR CANCELATION. . . (If you are not departing and within 24 hours of your scheduled departure time, you will be charged AUD 75 for cancellation)
Is included.

  This text includes the keywords “BEFORE DEPARTURE”, “CANCELATION”, and “FOR”. These keywords are synonyms (alternate or regular alternative expressions) of the key terms listed in FIG. TNA maps regular expressions to related keywords.

If TNA replaces “BEFORE DEPARTURE” with “BT”, “FOR” with “FOR”, “CANCELATION” with “CX”, and also replaces the built-in match of monetary values The text (ie the output of the derived attribute, which is a condensed text string)
BT MV <AUD, 75> FOR CX
It becomes.

This normalized text satisfies the condition code sequence “BT MV $ FOR CX”.

Here, if we decide that the phrase “WITHIN 24 HOURS OF SCHEDULED LIGHT TIME” (within 24 hours of scheduled departure time) is important and needs to be captured, we have new keywords for this, for example “ "W24HFT" must be added. Our normalized text is now
BT W24HFT MV <AUD, 75> FOR CX
It becomes.

  However, the new normalized text no longer satisfies the original condition because “W24HFT” is present in the code sequence. That is, the addition of a new key term can easily cause the TNA to perform a different evaluation than intended.

  The exact same problem occurs when key terms are removed from the text normalization process.

B. Comment and Condition Redundancy As outlined in the example above, the same matching sequence had to be used three times to extract values and currency from normalized text. This was inefficient in terms of processing time, as well as the time required by the user to create comments and conditions, and ultimately made knowledge base maintenance more difficult than necessary.

C. Sensitivity to keyword name change For example, if it is decided to change a keyword from “BT” to “Before Travel” (before departure), the variables and conditions that used this keyword are no longer applicable. This is similar to problem A, but the keyword name change is a cosmetic change and is more easily avoided. On the other hand, adding new keywords or deleting existing keywords is a more fundamental change to the text normalization process.

  Thus, previous attempts to solve the problem of pre-processing data in free-form text suffer from the inability to deal with modified keywords and inefficiencies in defining comments and conditions. It was. This limitation was found when attempting to address the problem in the context of the ticket issue example and log file example outlined above.

  Now consider the following log file fragment, taking an example of an IT support service.

A text normalization process that uses TNA filters these log entries,
PM DC
Reduce to. Here, the first log entry is encoded as “PM” and the third is encoded as “DC” (WARNING Cold not disconnect client (warning does not disconnect the client)), and the second and fourth (information Entries (for provisioning) are ignored.

A rule indicating a false positive (ie, insignificant) DC alarm may use the condition code sequence “PM DC” included.

However, if the TNA is changed to include a new item, such as a backup (BCK) event, the resulting normalized text is
PM BCK DC
It becomes.

  The condition indicating that the DC alarm was a false positive is then no longer correctly evaluated.

  Therefore, the same TNA restrictions as described in the previous ticketing example apply here as well.

  The embodiment of FIG. 9 provides a new tool (known as “Text Condenser Attribute” or TCA) that incorporates both Key Terms and Key Concepts. By combining both key terms and key concepts into a single tool, the problems caused by adding or deleting keywords are overcome. Since the keyword is an object shared by both the term and the concept, the name can be changed without affecting the rule applied to the aggregated data item (for example, step 702 in FIG. 7). Furthermore, since the tool includes the extraction of the key concept as the derived attribute itself, there is less need to iterate the conditions and variables.

  FIG. 10 shows an exemplary user interface of a TCA. “Attribute” or “Primary Attribute” is one of the basic elements of rule conditions or other expressions. Each attribute has a name and an associated value or possibly a series of values (such as when multiple values in time series are associated with the attribute). The attribute represents a data value element of the rule condition, for example, a lower (low level) data item such as a single allergen marker, or a higher (high level) aggregate data item such as a pollen item. Other elements of the rule condition are arithmetic, text, logical, or other expressions that associate an attribute with its value to form a Boolean expression. For example, the rule condition “some pollen is high” includes the attribute “pollen” (aggregated data item) and the logical expression “some X is high”, where the variable “X” is replaced with the value pollen. It has been.

  A “case” is a collection of attributes and their values that are presented to the expert system for interpretation. The preprocessor takes complex cases, ie cases with a large number of attributes, attributes with a large amount of free-form text data, or attributes with a long sequence of data items, and aggregate data items (higher attributes, or “derived” "Attributes" are added to a case to reduce the complexity of the case and make it easier and more generic to use in rule conditions and interpretation reports.

The text summarizer attribute (TCA) is such a derived attribute. The Text Summarizer attribute defines a set of keywords (or “key terms”) along with a set of key concepts or “Derived Matches” (see FIG. 11). Each key concept or derivative match is
(A) Target (actually another derived attribute in the knowledge base or expert system)
(B) Extraction formula (determines the value of the derived attribute for the matched format)
(C) List of “Matching Forms” (matching format is a series of key terms (key term sequence))
Consists of.

The present embodiment performs TCA evaluation on a block of text as follows. That is,
(A) The text is normalized to a series of keywords (sequence of keywords).
(B) The normalized text is parsed by each of the derived matches to provide a value for the key concept. For each derived match, the longest of the matching formats (if any) is picked. This is known in the literature as a “greedy” pattern match.
(C) For each derived match for which a matching form finds a match, a predetermined expression for its associated derived match is applied, which becomes the attribute value for the derived attribute corresponding to that key concept. .

  Consider an example of analyzing an air fare table in which all related expressions are “$ (1)”. This is interpreted by the systems 1, 3 (see FIGS. 4 and 6) as “return the first monetary token found in the matched text”. We will see another extraction formula later.

  The above process can be applied across all of the samples in the sample sequence for the attribute referenced in the case (eg, “category” in the ticket reissue example above). A “sample sequence” is an ordered and timed list of values for a given attribute. Each value in the sample sequence is associated with a date and time. In this way, the TCA generates a sample sequence for the TCA and for each of the associated derived attributes. A sample sequence containing at least one non-blank value is introduced into the case. FIG. 12 shows an exemplary case with a value for the attribute “Category”, where the value for TCA is called “TCA” and the value for the derived attribute is , “CxBt”, “CxAt”, “RiOb1”, and “RiRtc”.

The use of TCA allows the problems associated with the use of TNA described above to: (a) allow keywords to be added and deleted safely;
(B) reduce redundancy in comments and conditions;
(C) The problem is solved by enabling the renaming of keywords.

A. Keywords can be safely added and deleted When defining matching formats in derived matches (ie, key concepts), the user is able to match the raw text to be matched so that some example raw text accompanies each matching format. You are prompted to provide an example text (see FIG. 13). The example provided by the user must give a match to the matching format. If the user makes a change to a keyword (eg, by adding a keyword) so that the normalized example no longer matches the matching format, the user will (eg, a derived match indicated in a different color) Be warned about this (by or by some other means of alerting the user).

  For example, when a keyword “-” having a phrase “-” for matching (matching) is added to a set of keywords, as shown in FIG. To recover derived matches, new keywords must be deleted, or some of the matching formats need to be changed to match the normalized versions of those examples.

  Thus, the example in a derived match is similar to the cornerstone case in a ripple down knowledge base in that it provides the context for the definition of its key concept.

B. Less verbosity of comments and conditions Derived attributes of TCA can be used directly on variables in comments and conditions. These conditions just insist on the existence of a derived attribute in the case, eg
The availability of CxBt could be used to add the comments shown in FIG.

C. Keyword names can be changed Since keyword names are only object references shared by keywords and matching forms in derived matches, by including the derived match in the TCA along with the keywords, the system You will be resistant to changes. Therefore, for example, when the keyword “BT” is renamed to “BeforeTravel”, our collation format is automatically updated as shown in FIG.

Other benefits of TCA Different extraction formulas The illustrated example of a derived match shows the extraction of monetary values from the normalized text of a fare table. There may be other additional types of information that need to be extracted from free-form text, such as fare schedules. An example (in the ticketing scenario) is
(A) whether a key phrase (key phrase) has occurred and / or (b) date.

  Continuing with the example of ticketing, if the matching format includes one or more dates (these appear automatically as keywords, as well as monetary values), use the formula "@ (i)" The i th date can be extracted. To handle key phrases, we use the expression “?” To indicate that the derived attribute should take the value “True” (True) if there is a match. An example of a TCA that uses both of these equations (ie, extracts a date and a Boolean value) is shown in FIG. FIG. 19 shows how these Boolean and date values for derived attributes appear in this exemplary case.

Tool tips (Tooltips)
If a user sees a derived attribute and its value in a case, the user may not know why it is there. That is, it may not be clear what part of the raw text it represents. To assist in this regard, in one embodiment, the raw text that resulted in the derived attribute and its matching value is provided as a tool hint (as illustrated in FIG. 20).

  For large reports consisting of several report sections (each with an optional heading), the order in which these report sections are presented depends on the end user (eg, physician, airline examining issued ticket, real estate) It is an important factor for professionals or buyers / sellers. That is, the end user wants to see the most important report section near the top of the report. However, what makes one report section more important than another depends on the particular case being interpreted. Therefore, it is beneficial to order specified report sections using rules that operate on the data in each case. Some other report sections, such as the summary report section that is always at the top of the report, must have a fixed placement. Thus, the user may be able to define a mix of both fixed and variable report section ordering.

  Allergy reporting is an area where variable report section ordering may be required. There are at least five separate report sections corresponding to comments on pollen, food, mite, mold, and animal allergen test results. If food allergy test results are most important for a given patient, the food report section should come before the other four, and so on. The report section corresponding to the least important test result should be placed after the others. In addition, there is a fixed report section, a summary report section located at the beginning of the report, and a recommendation report section usually located at the end of the report.

As a result, the system provides operator 38 with a means to define a “derived attribute” for each variable report section using a rule syntax that assigns values corresponding to the desired report section ordering. . In the above example of allergy, for example, there are five derived attributes of “pollen_order”, “food_order”, “mite_order”, “mold_order”, and “animal_order”. pollen_order is defined as the highest value among all pollen data items, and the same applies to other items. The derived attribute “pollen_order” is associated with the pollen report section. For each case, the values of the five derived attributes are calculated and the corresponding report sections are ordered according to these values. For example, the case has the following data items and values: turf = 50, birch = 20, (pollen)
Wheat = 5, Soybean = 15, (Food)
Mold = 2
Tick = 1
Cat = 62, dog = 49 (animal)
The report section has the following order: animal, pollen, food, mold, tick.

In some embodiments, the system may provide at least one of the following: That is,
・ Ripple-down rule system as basic technology to manage very large knowledge base required ・ Automatic validation and reflexive testing using encoded output from knowledge base A mechanism for generating encoded information that is machine language instructions, such as controlling a workflow engine that controls, for example, a laboratory workflow such as: Natural language syntax for constructing rule conditions (syntax)
• Insert variables into comments that are evaluated by the particular case being interpreted. Variables may be defined using aggregate data items.

Inference Like a human expert, an expert system is not a large set of data that may or may not be applicable to a rule, but a form that can apply rules to data that has features to which the rule can be applied. Complex data needs to be reduced. Thus, rule-based systems require a program that selects the relevant rules. For example, we need a program that determines which rules have conditions that match the attributes of the presented case (s).

  An inference engine is such a program that applies the appropriate rules to a case and thus performs tasks that use the rules to infer a result, such as a “decision”.

  This inference is often based on rules that are appropriate in some circumstances. However, in other environments, the rules need to be changed. Thus, there is a continuous iteration of knowledge-based rules. The applicability of one or more rules to a case presented for interpretation determines the accuracy of the interpretation of the results.

  Examination of the interpretation generated by the knowledge base and possible corrections determine a set of “rejected” cases. These are presented to experts who maintain the knowledge base, accepting the original interpretation, or otherwise generate new rules so that the corrected interpretation is given to subsequent cases by the knowledge base.

  Like a human expert, a rule-based expert system needs to have data presented to the expert system regarding these features so that inferences can be made from the relevant important features. When inferring from complex raw data, not only is the number of specific rules needed unmanageable, but how new ones are encountered in the presented complex data even after the rules are built Cannot be interpreted. Inference from more generalized features in the data makes the interpretation process itself more generally applicable, and therefore more robust.

  In high transaction environments, expert systems can play a vital role in using human expertise to provide rapid interpretation of raw data. For example, pathology laboratories may need to provide interpretive reports on tens of thousands of patients per day, far beyond the human capacity of a small number of pathologists that can be employed at the laboratory. is there.

Multi-level reasoning Multi-level reasoning is a process of continuously generating higher level abstract concepts (called super concepts) in the interpretation process.

Higher level abstractions Intermediate conclusions are used to provide higher level abstractions that are later used by other rules in the inference process. The intermediate conclusions themselves do not appear in the final interpretation, but do affect the rules used to determine the final interpretation.

In a preferred embodiment,
1. Attribute level, and There are two stages of multilevel reasoning at the rule level.

Multi-level inference at the attribute level Attributes may be defined in terms of other attributes according to an expression or expression (eg, arithmetic expression, Boolean expression, set membership operation).

  These are called “derived” attributes as opposed to “primary” attributes, which are attributes received from an online information system.

  For example, a knowledge-based task is to flag all suspicious transactions for human professionals for further consideration and not to flag non-suspicious transactions. Flagging non-suspicious transactions wastes human expert time.

  Examples of transactions analyzed by the knowledge base include the following three transactions (transactions).

  The primary attributes in these financial transactions in a case may consist of “company”, “country”, “amount”, and “previous flag”.

Derived attributes are defined as abstractions that make the resulting rules more general and more maintainable. For example, the following derived attributes are defined:
1. "Oil industry". This is true if the transaction involves a company in the oil industry. Baxon is one belonging to the group of oil companies.
2. "Non-convention country". This is true if the transaction involves a specific set of countries that are not members of a specific international treaty. Uginta and Tigeria are included in this set.
3. "Border line amount". This is true if the transaction is close to the minimum so that it automatically falls under surveillance. In this example, an amount between $ 9500 and $ 9999 is considered a borderline.
4). “The number of previous border lines”. This is the total number of transactions that have been in the borderline within the past week.

  When these derived attributes are evaluated by a multilevel inference process, this case becomes:

This multi-level inference in this example requires two inference analyzes to define all attributes.
1. The first analysis makes it possible to determine the attributes “oil industry”, “non-conventional country” and “border line amount”. Following this,
2. The second analysis makes it possible to calculate the attribute “number of previous border lines”.

  Another example of multi-level reasoning is provided by the applicant and inventors of the present invention in Australian Patent Application 2010904545 (incorporated herein by reference).

  Derived attributes that use multi-level inference application at the attribute level are specified with respect to a single expression or expression, so that only those derived attributes that can be derived from the primary attribute can be determined.

Multilevel inference at the rule level
1. 1. derived attributes extracted from primary attributes (as shown above), and It is made possible by the conclusion of the knowledge base extracted by the rule base.

  At the first level of rule-based reasoning, the conclusion is given by one or more rules defined in terms of attributes (primary and / or derived attributes).

  At the second level, the conclusion may be given using a rule that incorporates the first level conclusion in the rule condition.

  This multi-level inference at the rule level may then be applied to the third and subsequent levels as needed.

In the financial example above, the user uses the rule condition: • “Border Line Amount” is true • “Non-Convention Country” is true • “Previous Border Line Number” is 2 or more The first level conclusion may be defined as “suspicious transaction”. In this case, all these conditions are true.

  The reason that "suspicious transaction" is defined as a conclusion rather than just another derived attribute is that this "suspicious transaction" is a more complex concept than that defined using a single formula or expression.

  For example, there can be many completely different scenarios that indicate “suspicious transactions”. Each scenario is specified by a different set of rule conditions, all of which give the conclusion “suspicious transaction”.

  Conversely, there may be cases where there appears to be very similar to a “suspicious transaction” scenario, but there is a slight difference that makes it not suspicious.

The user now has a rule condition: • “suspicious transaction”, and • “no previous flag”
May be used to define a second level of conclusion “checking required”.

  Again, these conditions are both true for this case. The conclusion "no previous flag" is included to avoid flagging transactions that have already been flagged, thereby avoiding duplicate inspection of transactions, the purpose of which is again human. This is to avoid wasting specialist time.

  The user can also define a third level of conclusion called “explanation”. The conclusion is, “This transaction for <amount> is suspicious. Last borderline transaction from company <company> has been <number of previous borderlines> times and current transaction is from a non-convention country <country>. Because there is a value. The rule condition that gives this conclusion is simply “checking”.

  The purpose of this conclusion is to provide a human inspector with a text explanation of why this case was flagged for examination by the knowledge base. Variables in comments (indicated by <> symbols) are evaluated from specific attribute values in this case.

The user must use the rule conditions “Inspection Required” and “Oil Company”
Can also be used to define a third level of conclusion called the “queue” with the value “oil trade”. The purpose of this conclusion is to send this case to a specific queue that is examined by an expert who is experienced in a specific type of transaction (in this case, a transaction by an oil company).

  Once these conclusions are evaluated by a multilevel inference process, the interpretation of this case and its rule base is as follows:

  Once all attributes have been defined, three inference passes are required to determine all these conclusions.

  In the first pass, the conclusion “suspicious transaction” is given. In the second pass, “need inspection” is given. In the third pass, “Queue” and “Description” are given.

  By applying multi-level inference at the rule level to a knowledge base that uses the Ripple Down Rules methodology, a concept such as “suspicious transaction” is considered, and all scenarios that show suspicious transaction are considered and suspicious It is possible to conclude after excluding all scenarios that do not show a deal. This gives a broader view than that provided by one or more derived attributes using only multi-level reasoning.

  Multi-level reasoning allows the knowledge base to be defined using concepts that are at the level of abstraction appropriate for the expert field.

  At the attribute level, multi-level reasoning allows a more or less arbitrary set of primary attributes received from an online information system to be pre-processed into derived attributes that represent concepts relevant to the field.

  Preprocessing at the attribute level can also reduce data complexity, for example, by extracting concepts from large blocks of free text or by filtering and grouping many related attributes into higher-level attributes. it can.

  At the conclusion level, multi-level reasoning defines that the higher level abstraction represents a concept that a human expert would normally use when understanding this field and justifying the interpretation of a particular scenario. to enable.

  Using the RippleDown methodology to make a conclusion is that these higher level abstractions have been refined over time to provide not only a natural representation of potentially complex concepts, but also very accurate and clear. It means that you can also provide simple expressions.

  Build a large, sophisticated, and useful knowledge base with these abstractions using a set of appropriate abstractions specific to the domain of the case, coupled with ripple-down built-in cornerstone case verification procedures (discussed separately) And professional maintenance work becomes possible.

Example 1
The first application is a leukemia report knowledge base in which diagnosis is performed using hundreds of tests whose values are determined by a microarray of hundreds of protein expression markers or gene expression markers. Experts build a diagnostic report knowledge base that identifies comments on multiple subsets (subsets) of related markers, diagnoses corresponding to this pattern, and those significant subsets listed in text reports for referral physicians May be.

  The array test results are provided as a plurality of data items and value pairs as input to the knowledge base. In this example, these data items are labeled CD1 to CD100 for identification. These labels represent 100 elements for the array. Real-world examples may include hundreds of markers.

  In this example, if one of the data values is less than 50, it means that there is no expression of the antibody corresponding to the marker for the patient sample. A value greater than 50 may be significant (depending on the value of other markers). Marker values higher than 100 indicate significant expression.

  The diagnosis of a particular variety of leukemias can be estimated from the value of a designated subset of 100 data values.

  For example, a diagnosis of B cell chronic lymphocytic leukemia (B-CLL) can be deduced from the fact that at least two of CD1, CD2, CD3, CD4 and CD5 show significant expression. This diagnosis is supported by the fact that any of CD6, CD7, CD8, CD9 and CD10 shows significant expression, but these do not diagnose BCLL by itself.

  In addition, the diagnosis of acute myeloid leukemia (AML) can be estimated from the fact that at least two of CD11, CD12, CD13, CD14 and CD15 show significant expression. This diagnosis is supported by the fact that any of CD16, CD17, CD18, CD19 and CD20 shows significant expression, but these do not diagnose AML by itself.

Data is supplied to the five aggregated data items as specified by the following structure example using the received data items.
1. “BCLL diagnosis” is populated by data items CD1, CD2, CD3, CD4, CD5.
2. “BCLL support” is populated by data items CD6, CD7, CD8, CD9, CD10.
3. “AML diagnosis” is populated by data items CD11, CD12, CD13, CD14, CD15.
4). “AML support” is populated by data items CD16, CD17, CD18, CD19, CD20.
5. “Leukemia” is populated by the aggregate data items “BCLL diagnosis”, “AML diagnosis”, “BCLL support”, “AML support”.

  FIG. 21 shows an outline of an aspect of a structure that gives a hierarchical relationship between these data items and 15 aggregated data items. In some embodiments, when the lower hierarchy of the structure is populated, the value or property in the upper hierarchy is calculated. The structure may be stored in the memory 20 of the device 1, the hard drive 2 (or other data storage unit) or the like and interpreted by the CPU 4.

The following ranges: (a) “No detection” defined as constant 50, and (b) “High” defined as constant 100.
May also be defined.

In order to represent significant data items in each set, the following rules 1. “Significant BCLL diagnosis”. This is populated by the rule “BCLL diagnosis in range [> high]”.
2. “Significant BCLL support”. This is populated by the rule “BCLL diagnosis in range [> no detection]”.
3. “Significant AML diagnosis”. This is populated by the rule “AML diagnosis in range [> high]”.
4). “Significant AML support”. This is populated by the rule “AML diagnosis in range [> no detection]”.
By applying, further aggregated data items are populated.

BCLL diagnosis comment is the following pseudo-text "Pan-B cell antigen expression in <significant BCLL diagnosis>, co-expression in <significant BCLL support>, typical for B cell chronic lymphocytic leukemia (B-CLL) . "
Given by.

  The variable “<significant BCLL diagnosis>” is an instruction for enumerating the names and values of data items of significant BCLL, and the same applies to the variable <significant BCLL support>. In this embodiment, the listed names and values are ordered in descending order of the data item values so that the most significant attributes are listed first.

The BCLL diagnostic rule determines the generation of BCLL diagnostic comments by “number of significant BCLL diagnoses ≧ 2” and “number of significant BCLL support ≧ 1”.
Trigger like this.

  In other words, a comment is generated when two or more data items exist in the set “significant BCLL diagnosis” and one or more data items exist in the set “significant BCLL support”.

As a second comment example, an AML diagnostic comment is co-expressed in the following pseudo-text "Positive <significant AML diagnosis, as name>", consistent with AML antigen expression, <significant AML support, as name> The possibility of M2 classification is suspected. "
Given by.

  The data item “<significant AML diagnosis, as name>” is an instruction to the knowledge base for enumerating only the names of significant AML data items, and the same applies to the variable <significant AML support, as name>. is there.

  In this embodiment, the listed names and values are ordered in descending order of data item values so that the most significant data items are listed first, even if the values are not shown in the comments.

The AML diagnostic rules that trigger the generation of AML comments are:
“Number of significant AML diagnoses ≧ 2” and “Number of significant AML support ≧ 1”
It may be.

  A comment is generated when there are two or more data items in the set “significant AML diagnosis”. This means that there are two or more data items with a value greater than 100 in the set “AML diagnosis” and at least one data item with a value greater than 50 in the set “AML support”. To do. Consider the test results of samples from patient “A” as follows:

These results are sent to one embodiment of the knowledge base. This knowledge base has aggregated data items: • Significant BCLL diagnoses are evaluated as “CD5, CD3 and CD4” • Significant BCLL support is evaluated as “CD7 and CD8” • Significant AML diagnosis and significant AML support is Both are evaluated as null (empty) and evaluated, and expression is evaluated.

  Thereafter, the knowledge base interprets according to the rules determined as described above. Since there are three elements in the “significant BCLL diagnosis” set and two elements in the “significant BCLL support” set, the BCLL rule is applicable in this case.

The knowledge base evaluates the variables in the BCLL comments “<significant BCLL diagnosis>” and “<significant BCLL support>” and then the evaluated comments “CD5 (260), CD3 (190) and CD4 (150 ) Pan-B cell antigen expression, CD7 (90) and CD8 (60) co-expression, typical for B cell chronic lymphocytic leukemia (B-CLL). "
give.

  For the second example, consider the following patient “B” test results.

These results are sent to the knowledge base. The knowledge base initially aggregated data items: • Significant BCLL diagnosis and significant BCLL support both evaluate as null (empty) • Significant AML diagnosis evaluates as “CD12 and CD11” • Significant AML support is , “CD19, CD20 and CD18”.

  The knowledge base then interprets according to the above rules. Since there are two elements in the “significant AML diagnosis” set and three elements in the “significant AML support” set, the AML rule is applicable in this case.

The knowledge base then commented: “Matches AML antigen expression based on positive CD12 and CD11, co-expressed on CD19, CD20 and CD18. The possibility of M2 classification is suspected.”
give.

Example 2
Another application is an allergy report knowledge base that can result in 500 or more possible IgE tests. The task of allergy specialists is which subset of tests performed has a significant outcome value for the patient, which test values may not be significant, and which tests need to be tracked Advising the referring physician about (including which of the 500 possible tests should be performed together as a follow-up).

  One solution is as follows.

  • Group all possible data item names into multiple aggregated data items based on discipline specific rules (such as significant pollen attributes) and case specific rules To do.

  Use any of the characteristics of the aggregate data item as the basis for further rules and / or comments. For example, when the number of elements of a certain aggregate data item exceeds a certain number, or when the set includes a specific element, a specific comment is given.

  • Use one or more aggregated data items as variables in comments. For example, “dog, cat and peanut allergies are significant”. In this comment, the phrase “dog, cat and peanut” is an assessment of an aggregate data item consisting of significant allergens for this case. The general form of this comment can be “{significant allergen} allergy is significant”, where the set {significant allergen} itself is defined by the rules.

  • Optionally, include the value of the data item in the comment. For example, “dog (102.3), cat (56.4) and peanut (43.5) allergies are significant”.

  • Properly order data items within aggregated data items that appear in comments. For example, order the attribute values in the case in descending order so that the most important attribute appears first in the comment.

  Automatically format data items into naturally structured sentences that are consistent with the rest of the report. For example, if three attributes are significant, the set format is “dog, cat and peanut allergy”, and if only two attributes are significant, the set format is “dog and cat allergy. It is good also as.

  -Allows grouping of data items to be defined based on previous groupings of data items. For example, the new aggregate data item can be the difference, sum, product, or any set operation between one set and another set. This makes it possible to define a hierarchical structure consisting of a plurality of sets. For example, the difference between the set “appropriate examination” and the set “instructed examination” may identify a set of appropriate examinations that have not yet been indicated.

• Be able to define comments that appropriately use either personal data items or aggregated data items that contain their attributes. For example, if the list of personal data items is too long for comments in that particular case, the term “food allergy” is used instead of “peanut, soy, milk, egg and peach allergy”.

  Similarly, it is possible to appropriately define a comment that uses a super aggregate data item name instead of a subset (subset) name. For example, if the list of individual aggregated data items is too long for comments in that particular case, the term “inhalation allergy” is used instead of “pollen, animal, mold, allergy”.

Pseudotext The following table defines some of the characteristics of the pseudotext described above, where s, t, x, y, z are whether they are primary items or derived data items. Regardless, it refers to a data item.

  The letters X, Y, Z particularly refer to aggregated data items (a type of derived data item). The letters a and b refer to numbers or named constants. The letter p refers to a Boolean expression.

Example 3
The third application example interprets the ticket fare table,
(A) Departure city and destination city (b) Knowledge base used to determine conditions under which tickets can be reissued, such as how much penalty can be applied.

  The second knowledge base is used to interpret the reissue ticket to determine the actual fee paid as well as the departure and destination cities.

  The third knowledge base interprets the output of the other two knowledge bases to determine whether the reissue ticket complies with fare table conditions.

  Considering the first knowledge base for the time being, the free-form text fragment in the fare table can be as follows:

A Text Condenser Attribute (TCA) is used to pre-process this fare table fragment. Some keywords in TCA are:
(A) “CANCELATION FEE” (cancellation fee) and “CX” for other similar phrases
(B) “BEFORE DEPARTURE” (before departure) and “BT” for other similar phrases
(C) “<AUD n>” representing a monetary value having a variable amount of money such as “AUD110” or “AUD75”.
It is.

  The key concept in the TCA is a derived attribute “CancellationFeeBeforeTravel”, which is defined as having a variable monetary value N derived from this series of keywords “CX BT <AUD N>”. In this example, the attribute “CancellationFeeBeforeTravel” has a numerical value 110.

  The knowledge base has rules for outputting this and other fare table conditions as conclusions in a standardized format such as “CancellationFeeBeforeTravel = 110”. These standardized forms are entered into a third knowledge base, which outputs the second knowledge base output and fare table conditions that summarize the actual navigation and fees paid for the reissue ticket. Compare. The third knowledge base then interprets the fare schedule conditions and the summarized details of the reissue ticket to determine whether the reissue ticket has been reissued according to the fare schedule conditions.

Example 4
A fourth application is a knowledge base that is used to monitor a series of log file entries and determine whether and when to send alerts to IT support staff.

  Expert systems can play a fundamental role in assisting IT support staff in monitoring and interpreting log files. The expert system considers this very large free-form text data to identify specific alarms, warnings, or trends that the IT support staff considers important and indicate that any preventive or corrective action by the IT support staff is permitted. Can help to extract from.

  A more specific difficulty when interpreting log files is the problem of “false positives”, ie alarms that are not really important or less important than other alarms.

  For example, the first “out of memory” alert may be significant, but the same alert repeated subsequently may only point to a problem that has already been warned (false positive); It may be a completely new problem (true positive). If there is a long time difference between these two alarms, the second alarm is more likely to be true positive.

  Another example is the case of a “disk thrashing” alarm due to a high rate of page faults. This is important, but in some cases it is only a symptom of a more important problem, which may be an out of memory condition. In this case, the presence of an “out of memory” alert prior to the “disk thrashing” alert means that the second alert is not as important as if there was no preceding “out of memory” alert.

  Another example is the “client disconnect failure” alert. For most of the day, this is a significant alarm that allows immediate investigation by IT support staff. However, the alert was logged at a specific time of the day, for example, at 2 am, when the system is known to be offline due to daily preventive maintenance (PM). In some cases, the alert may not be significant.

  Rules that determine the significance and appropriate response of an alarm must consider not only the types of alarms but also their order, frequency, timing relative to each other, and even absolute time stamps.

  Once the significance of the alarm is determined, a decision regarding the appropriate action to take can be made by the expert system. Therefore, in order for the expert system to interpret the log file, it is necessary to perform preprocessing so as to classify individual log entries or a sequence of log entries into a sequence of key terms or key concepts. Thus, complex and essentially free-form text log files are reduced to a normalized format from which simpler and higher-level atomic data items are extracted and used in rule conditions can do.

  An example of a potential alarm situation when the log entry indicates that the user (client) cannot be disconnected is shown. However, if this alert occurs after the system has started preventive maintenance, the potential alert is considered a false positive and no alert is sent to the IT support staff.

  Consider the following example of a series of log file entries.

In order to pre-process these log file entries prior to interpretation by the knowledge base, we build a TCA. The keywords in TCA are
(A) “PM” for “Preventive Maintenance” (preventive maintenance) and other similar phrases
(B) “WARNING” (Warning) and “WARN” for other similar phrases
(C) “Could not connect client” (could not disconnect the client) and “DC” representing other similar phrases
It is.

  The TCA value for these log file entries is in normalized text format “PM WARN DC”. One of the key concepts of this TCA is a derived attribute “Alert” (alarm). This is defined to have the Boolean value “true” when the normalized text format includes “WARN” (also in this example).

  The knowledge base has a rule for adding a workflow action “Send alert email” (send an alert email) when the value of the attribute Alert is “true”. This workflow operation includes the email address of the IT support staff to be notified, the email header provides a summary of the alert, and the email body describes the details of the alert.

  Another key concept of TCA is the derived attribute “FalsePositive” (false positive). This is defined to have the Boolean value “true” when the normalized text format includes “PM WARN DC” (also in this example).

  The knowledge base has another subsequent rule to delete the workflow action “Send alert email” if the value of the attribute FalsePositive is “true”.

  In the pre-processing stage, these two derived attributes Alert and FalsePositive are added to the case, both having the value “true”.

  During the knowledge base reasoning phase, an alert workflow action is added to the interpretation by a rule with the condition “Alert is true” (alert is true). However, the alert workflow action is deleted by a subsequent rule that has the condition “FalsePositive is true” (false positive is true), and the net result is that the alert workflow action does not exist in the interpretation and therefore IT support staff No alarm email will be sent to you.

  Having described the embodiments so far, it will be appreciated that some embodiments may have some of the following advantages.

  Multiple data items (including text results presented as free-form text in some embodiments) can be processed for the purpose of generating a single text report.

  The report presents the data items that are important to each case, in the appropriate order, in a linguistically natural syntax.

  The number of specific report variations is essentially infinite due to the number of possible subsets of data items and the number of possible orderings within each subset.

  The number of specific rule conditions that determine a particular report is also essentially infinite, due to the number of attributes and their value patterns in the case.

  Nevertheless, since the rules are based on derived attributes that are the result of the preprocessing stage, the expert can build and maintain the knowledge base using a manageable number of rules.

  An expert system is provided that can manage multiple attributes and corresponding multiple report variations.

  It will be understood that the invention shown in the specific embodiments can be subject to many variations and / or modifications without departing from the spirit and scope of the invention as broadly described. Accordingly, the embodiments of the invention are to be considered in all respects only as illustrative and not restrictive.

DESCRIPTION OF SYMBOLS 1 System 2 Computer readable medium 3 System 4 Central processor 8 Data item 10 Data receiver 11 Data transmitter 12 Remote site 14 Network 20 Memory 22 Data source 24 Aggregated data item 26 Text information 28 Screen 30 Printer 32 Workstation 36 Receiver 37 Information system 38 users

Claims (20)

A method for generating information from a plurality of personal data items , performed by a knowledge-based system for inferring conclusions , comprising:
(A) a population step of populating aggregated data items using at least one of the plurality of personal data items ,
Each personal data item includes original information including an attribute and a value that are identifiers of the personal data item,
The aggregated data item is a form of a derived attribute;
The derived attribute is a conclusion estimated by a rule-based knowledge base and represents a transformation from a set of personal data items to a single data item having a value;
The value of the derived attribute is attribute-to-value for each personal data item in the set of personal data items such that the derived attribute forms a single data item suitable for inference by a rule-based knowledge base. An aggregate value including mapping, wherein the single data item retains the original information for each of a plurality of the personal data items and queries from the entire knowledge base to extract information about the personal data items Receive
The population step ;
(B) an application step of applying a rule applied by a rule-based knowledge base to infer to the aggregated data item,
The rule so that a single rule can query the plurality of personal data items as a single data item rather than relying on multiple rules for each personal data item or a combination of the plurality of rules. Are performed on the set of personal data items, (i) query, (ii) iteration, (iii) identify subset, (iv) identify specific personal data items, (v) sort, ( vi) comparing a set of personal data items with another set of personal data items, and (vii) a set operation that includes one or more of the other set operations.
The applying step;
(C) using the aggregate data items and a generation step of generating information,
The method of generating information is performed by a rule-based knowledge base that generates information by applying one or more of the rules to at least one of the aggregated data items ;
The information generated is (i) belongs to one or more of text information and (ii) machine instructions,
The generating step is such that the rule-based knowledge base can generate information on a plurality of personal data items by applying a rule including a set operation to a derived attribute.
i. Including in the information an identifier of one or more personal data items that populate the aggregated data items;
ii. Sub-step of including in the information the value associated with one or more personal data items that populate the aggregated data item;
Including one or more of
Method. The method of claim 1, wherein the generating step further comprises a sub-step of determining an order of identifiers for one or more of (a) personal data items and (b) aggregated data items . The method according to claim 1 or 2 , wherein the text information is grammatically correct. The method according to claim 1, wherein the one or more rules are applied by the rule-based knowledge base when generating information from at least a part of a ripple down rule knowledge system. The plurality of personal data items includes at least unstructured data items having a value consisting of one or more of (a) free-form text and (b) a sequence of free-form values . 5. The method according to any one of 4. The derived attribute is a text summarizer attribute;
The text summarizer attribute is free so that the unstructured data items are processed from complex data items to one or more simple data items each having a value for ease of interpretation. One or more regular expressions in the formatted text,
(A) a standard representation of the free form text that is a sequence of keywords in a normalized form;
(B) a number of atomic data items, each atomic data item representing a key concept enumerated in the free-form text data item, and the value of each atomic data item is (i) a Boolean A number of atomic data items that are one of values, (ii) a finite enumeration, and (iii) a numeric value;
Mapping to one or more of
The method of claim 5. To make interpretation easier,
(A) populating aggregated data items with at least one of a plurality of derived attributes;
(B) applying one or more rules including one or more set operations to the aggregated data item;
(C) mapping regular expressions in free-form text to one or more of the standard expressions or the multiple atomic data items;
One or more of the steps occur during the iteration,
The method of claim 6 . A method for generating information from a plurality of personal data items, performed by a knowledge-based system for inferring conclusions, comprising:
(A) an application step of applying a rule to an aggregate data item to infer,
The personal data item comprises original information including an attribute and a value that are identifiers of the personal data item;
The aggregated data item is a form of a derived attribute;
The derived attribute is a conclusion estimated by a rule-based knowledge base and represents a transformation from a set of personal data items to a single data item having a value;
The derived attribute value is an attribute-to-value mapping for each personal data item in the set of personal data items such that the derived attribute forms a single data item suitable for inference by a rule-based knowledge base. The single data item holds the original information about each of the plurality of personal data items, and queries from the entire knowledge base to extract information about the personal data items. receive,
The applying step;
(B) an evaluation step of evaluating the result of one or more rules using one or more aggregated data items, each including one or more personal data items,
The one or more rules are applied by a rule-based knowledge base;
The single rule can be queried as a single data item rather than relying on multiple rules or combinations of rules for each personal data item. The above rules are performed on a set of personal data items: (i) query, (ii) iteration, (iii) specify subset, (iv) specify specific personal data items, (v) sort , (Vi) comparing a set of personal data items with another set of personal data items, and (vii) a set operation including one or more of the other set operations.
The evaluation step;
(C) a generating step for generating the information according to the result,
The generating step is performed by a rule-based knowledge base that generates information by applying one or more of the rules to at least one of the aggregated data items;
The generated information belongs to one or more of (i) text information and (ii) machine instructions;
The generating step is such that the rule-based knowledge base can generate information on a plurality of personal data items by applying a rule including a set operation to a derived attribute.
    i. Including in the information an identifier of one or more personal data items that populate the aggregated data items;
    ii. Sub-step of including in the information the value associated with one or more personal data items that populate the aggregated data item;
Including one or more of
The generating step;
Including methods. A method for generating information from a plurality of personal data items, performed by a knowledge-based system for inferring conclusions, comprising:
(A) a receiving step of receiving a conceptual representation of information including an interpretation portion representing an operation on an aggregate data item including a plurality of personal data items,
Each personal data item includes original information including an attribute and a value that are identifiers of the personal data item,
The aggregated data item is a form of a derived attribute;
The derived attribute is a conclusion estimated by a rule-based knowledge base and represents a transformation from a set of personal data items to a single data item having a value;
The value of the derived attribute is attribute-to-value for each personal data item in the set of personal data items such that the derived attribute forms a single data item suitable for inference by a rule-based knowledge base. An aggregate value including mapping, wherein the single data item retains the original information for each of a plurality of the personal data items and queries from the entire knowledge base to extract information about the personal data items Receive
The receiving step;
(B) an application step of applying a rule applied by a rule-based knowledge base to infer to the aggregated data item,
The rule so that a single rule can query the plurality of personal data items as a single data item rather than relying on multiple rules for each personal data item or a combination of the plurality of rules. Are performed on the set of personal data items, (i) query, (ii) iteration, (iii) identify subset, (iv) identify specific personal data items, (v) sort, ( vi) comparing a set of personal data items with another set of personal data items, and (vii) a set operation that includes one or more of the other set operations.
The applying step;
(C) a generation step of generating the information from the interpretation part,
The information is generated by applying one or more rules to at least one of the aggregated data items;
The generated information belongs to one or more of (i) text information and (ii) machine instructions;
The generating step is such that the rule-based knowledge base can generate information on a plurality of personal data items by applying a rule including a set operation to a derived attribute.
    i. Including in the information an identifier of one or more personal data items that populate the aggregated data items;
    ii. Sub-step of including in the information the value associated with one or more personal data items that populate the aggregated data item;
Including one or more of
The generating step;
Including methods. A system for generating information from a plurality of personal data items,
  (A) a computing device;
  (B) a preprocessor that is executable in the computing device and populates aggregated data items using at least one of the plurality of personal data items;
Each personal data item includes original information including an attribute and a value that are identifiers of the personal data item,
The aggregated data item is a form of a derived attribute;
The derived attribute is a conclusion estimated by a rule-based knowledge base and represents a transformation from a set of personal data items to a single data item having a value;
The value of the derived attribute is attribute-to-value for each personal data item in the set of personal data items such that the derived attribute forms a single data item suitable for inference by a rule-based knowledge base. An aggregate value including mapping, wherein the single data item retains the original information for each of a plurality of the personal data items and queries from the entire knowledge base to extract information about the personal data items Receive
The preprocessor;
(C) a rule-based knowledge base that infers by applying rules to the aggregated data items,
The rule so that a single rule can query the plurality of personal data items as a single data item rather than relying on multiple rules for each personal data item or a combination of the plurality of rules. Are performed on the set of personal data items, (i) query, (ii) iteration, (iii) identify subset, (iv) identify specific personal data items, (v) sort, ( vi) comparing a set of personal data items with another set of personal data items, and (vii) a set operation that includes one or more of the other set operations.
The rule-based knowledge base;
(D) an information generator that generates information using the derived attribute;
With
The information generator forms at least part of a decision support system;
The generated information belongs to one or more of (i) text information and (ii) machine instructions;
The generation of the information is such that the rule-based knowledge base can generate information on a plurality of personal data items by applying a rule including a set operation to a derived attribute.
i. Including in the information an identifier of one or more personal data items that populate the aggregated data items;
ii. Sub-step of including in the information the value associated with one or more personal data items that populate the aggregated data item;
Including one or more of
A system comprising: The system of claim 10, wherein the preprocessor can iteratively construct derived attributes that use conclusions previously generated in an iterative process. 12. A system according to claim 10 or 11, wherein the information generator can iteratively construct conclusions using rules that use previously defined conclusions in an iterative process. The derived attribute is
(A) Aggregated data item and
(B) a text summarizer attribute;
(C) extracting any one or more superordinate concept data items from a plurality of data items, thereby reducing any data complexity and any other result of data preprocessing
The system according to claim 10, wherein the system is one or more of the following. A receiver for receiving the plurality of personal data items;
The receiver is one of (a) a handheld electronic device and (b) another computing device having processing power;
The system according to any one of claims 10 to 13. (A) Aggregated data item,
(B) one or more rules,
(C) other derived attributes;
(D) one or more conclusions,
(E) any other result of data preprocessing that extracts one or more superordinate concept data items from a plurality of data items, thereby reducing data complexity;
(F) any other conceptual representation used in defining the interpretation part in the generation of the information
The system according to any one of claims 10 to 14, further comprising a builder capable of processing one or more of the above. 15. A system according to any one of claims 10 to 14, further comprising a transmitter for sending the generated information to one or more of (a) a machine and (b) a recipient. A system for generating information from a plurality of personal data items,
(A) a rule-based knowledge base that infers by applying one or more rules to aggregated data items,
The personal data item comprises original information including an attribute and a value that are identifiers of the personal data item;
The aggregated data item is a form of a derived attribute;
The derived attribute is a conclusion estimated by a rule-based knowledge base and represents a transformation from a set of personal data items to a single data item having a value;
The value of the derived attribute is attribute-to-value for each personal data item in the set of personal data items such that the derived attribute forms a single data item suitable for inference by a rule-based knowledge base. An aggregate value including mapping, wherein the single data item retains the original information for each of a plurality of the personal data items, and queries from the entire knowledge base to extract information about the personal data items Receive
The rule-based knowledge base;
(B) an evaluator for evaluating the result of the one or more rules,
The single rule can be queried as a single data item rather than relying on multiple rules or combinations of rules for each personal data item. The above rules are executed on the set of personal data items: (i) query, (ii) iteration, (iii) identification of subset, (iv) identification of specific personal data items, (v) Sorting, (vi) comparing a set of personal data items with other sets of personal data items, and (vii) a set operation including one or more of the other set operations.
The evaluator;
(C) an information generator for generating information according to the result,
The information is generated by applying the one or more rules to at least one aggregated data item;
The generated information belongs to one or more of (i) text information and (ii) machine instructions;
The generation of the information is such that the rule-based knowledge base can generate information on a plurality of personal data items by applying a rule including a set operation to a derived attribute.
    i. Including in the information an identifier of one or more personal data items that populate the aggregated data items;
    ii. Sub-step of including in the information the value associated with one or more personal data items that populate the aggregated data item;
Including one or more of
The information generator and
A system comprising: A system for generating information,
  (A) a receiver for receiving a conceptual representation of information including an interpretation portion representing an operation on an aggregate data item including a plurality of personal data items,
Each personal data item includes original information including an attribute and a value that are identifiers of the personal data item,
The aggregated data item is a form of a derived attribute;
The derived attribute is a conclusion estimated by a rule-based knowledge base and represents a transformation from a set of personal data items to a single data item having a value;
The value of the derived attribute is attribute-to-value for each personal data item in the set of personal data items such that the derived attribute forms a single data item suitable for inference by a rule-based knowledge base. An aggregate value including a mapping, wherein the single data item holds the original information for each of a plurality of the personal data items, and from the knowledge base as a whole to extract information about the personal data items Get an inquiry,
The receiver;
(B) a rule-based knowledge base that infers by applying rules to the aggregated data items,
The rule so that a single rule can query the plurality of personal data items as a single data item rather than relying on multiple rules for each personal data item or a combination of the plurality of rules. Are performed on the set of personal data items, (i) query, (ii) iteration, (iii) identify subset, (iv) identify specific personal data items, (v) sort, ( vi) comparing a set of personal data items with another set of personal data items, and (vii) a set operation that includes one or more of the other set operations.
The rule-based knowledge base;
(C) an information generator for generating the information from the interpretation part,
The information is generated by applying one or more rules to at least one of the aggregated data items;
The generated information belongs to one or more of (i) text information and (ii) machine instructions;
The generation of the information is such that the rule-based knowledge base can generate information on a plurality of personal data items by applying a rule including a set operation to a derived attribute.
    i. Including in the information an identifier of one or more personal data items that populate the aggregated data items;
    ii. Sub-step of including in the information the value associated with one or more personal data items that populate the aggregated data item;
Including one or more of
The information generator and
A system comprising: A computer program comprising instructions for controlling a computer to carry out a method according to the method of claim 1. A computer readable medium providing a computer program according to the computer program of claim 19.