Creating a natural language processing model for an information domain

Abstract

Method for generating a natural language processing model (330) for a specific information domain (320), wherein the method comprises: - Deriving (101) a framework of a natural language vocabulary for the specific information domain (320) from a source model (111) of the specific information domain, wherein the framework contains terms from the source model, wherein the source model contains classification hierarchies of the terms, which include objects and attributes; - Creating a dictionary using the basic framework of the natural language vocabulary; - Applying (102) a set of syntax rules that define concepts and relationships; - Extending (103) the framework of the natural language vocabulary based on reference documents (322) from the specific information domain (320) to provide a natural language processing model (330) for the specific information domain (320), wherein extending the framework involves forming clusters and evaluating terms for concepts and relationships, as well as forming an intersection of syntax rules and clusters of concepts and relationships; and - Providing a natural language processing model (330) using the extended framework of natural language vocabulary for the specific information domain (320) to analyze documents in the specific information domain.

Description

AREA OF INVENTION

[0001] This invention relates to the field of generating natural language processing models. In particular, the invention relates to generating a natural language processing model for a specific information domain. BACKGROUND OF THE INVENTION

[0002] The article “An overview and classification of adaptive approaches to information extraction” by Siefkes and Siniakov, published in the Journal on Data Semantics IV, Springer, Berlin, Heidelberg, 2005, pp. 172-212 (DOI: 10.1007 / 11603412_6), describes algorithms and methods in the field of information extraction, in particular adaptive systems that can be adapted to new domains through training or the use of external knowledge sources.

[0003] Modern business analysis methods and processes depend heavily on the information that flows through and in connection with business operations. At the core of this business process information are both transactional data and textual data from sources such as emails, reports, presentations, and instant messaging. This information can be described as information that flows "through" business operations and whose origin lies within them or within a community with verified identities.

[0004] The relative importance of the textual component of this information has been steadily increasing for many years and is now recognized as a very important element. However, due to the difficulties in creating natural language processing (NLP) models to understand this unstructured content, the textual component of the information remains largely unused.

[0005] Furthermore, the advent of Web 2.0 and social media has created an additional deluge of textual data flowing “around” business activities, which can be of great value; the applications cover a very wide range from identifying new markets for products, understanding brand sentiment, and finding influencers, to name just a few.

[0006] The impact that advanced NLP models can have on business lies in the fact that they address one of the fastest-growing needs in the information technology industry. However, due to the complexity of natural language, creating such models is not easy.

[0007] The term Natural Language Processing Model (NLP model) describes an abstracted set of inputs to a text analysis engine, enabling it to extract concepts (named units such as references to a "person," references to a "place," etc.) and relationships between these concepts (e.g., "lives_in"). These "facts" allow the text to be accessed for programmatic use and process automation. Examples of facts in this case would be "John McGrath lives at 123 Main St., Dublin" or "Michael has listed 321 Main St. as his home address."

[0008] Such a model is highly dependent on the domain of the input being processed. There is no one-size-fits-all model, and even for something like recognizing "person," the degree of success (defined by accuracy or recall) will vary across different text sources. For example, a "person" mentioned in medical records is quite different from a "person" mentioned in academic quotations. In the case of medical records, furthermore, subclassifying "person" is crucial for understanding, so that the NLP system can assign the correct semantic subclass to a specific reference to a person—for example, distinguishing between patient and surgeon, general practitioner, and so on.

[0009] Domain-specific NLP models are costly to develop and maintain. There are two well-known approaches to developing such NLP models: 1) manually creating word lists and context rules from scratch, and 2) machine learning-based approaches that require an annotated corpus as input, on which machine learning algorithms are run to "learn" the features that specify the assignment of a semantic class to a given text segment.

[0010] Both known methods are very time-consuming, with the former having the advantage of being explicitly and manually fine-tunable, while the latter is a "black box", which, however, offers a lower entry threshold insofar as only knowledge of the field itself is required for the task of adding annotations, whereas the former also requires some knowledge of the underlying matching techniques.

[0011] Therefore, according to the current state of the art, it is necessary to address the aforementioned problem. SUMMARY OF THE INVENTION

[0012] According to a first aspect of the present invention, a method for generating a natural language processing model for an information domain is provided, comprising: deriving a basic framework of a natural language vocabulary from a source model of the information domain; applying a set of syntax rules.

[0013] From a first perspective, the present invention provides a method for generating a natural language processing model for an information domain, comprising: deriving a basic framework of a natural language vocabulary from a source model of the information domain; applying a set of syntax rules that define concepts and relationships; and extending the basic framework of the natural language vocabulary based on reference documents for the information domain to provide a natural language processing model for the information domain, wherein the extension of the basic framework involves forming clusters and evaluating terms for concepts and relationships.

[0014] Preferably, the present invention provides a method in which the derivation of a basic framework of a natural language vocabulary in the information domain derives preferred terms.

[0015] Preferably, the present invention provides a method in which the application of a set of syntax rules includes taking into account subject, predicate, object and changing the order for the linguistic representation.

[0016] Preferably, the present invention provides a method in which extending the framework includes: selecting a preferred term as a concept or relationship; performing a keyword search for the preferred term in reference documents from the information domain; and providing an ordered set of possible terms for the preferred term.

[0017] Preferably, the present invention provides a method that includes: determining local N-grams; measuring one or more measurable quantities of the N-grams and evaluating the N-grams.

[0018] Preferably, the present invention provides a method in which extending the basic framework for the natural language vocabulary involves deriving further syntax rules based on reference documents from the information domain. The method includes forming intersections of syntax rules and the concepts and relationships grouped into clusters.

[0019] Preferably, the present invention provides a method that includes: using verb structures from linguistic verb classes to control the formation of intersections, which is applied to terms grouped into clusters.

[0020] Preferably, the present invention provides a method in which the extension of the basic framework begins with an initial concept or relationship and iteratively moves outwards through neighboring concepts or relationship links in the source model; and the extended notions of concepts and relationships are refined by increasing evaluations.

[0021] Preferably, the present invention provides a method that includes: determining a deviation from best terms for a concept or relationship using an evaluation threshold.

[0022] Preferably, the present invention provides a method in which the extension of the framework dynamically changes a strategy that repeats in iterations based on results.

[0023] Preferably, the present invention provides a method in which the derivation of a basic framework for a natural language vocabulary is based on more than one source model.

[0024] Preferably, the present invention provides a method in which the derivation of a basic framework for a natural language vocabulary uses freely accessible data to initially populate the basic framework, wherein the ontology classes of the source model are matched with the classes of freely accessible data.

[0025] From another perspective, the present invention provides a computer program that is stored on a computer-readable medium and can be loaded into the internal memory of a digital computer, comprising sections of software code to perform the method as described above when the program is executed on a computer.

[0026] From a broader perspective, the present invention provides a system for generating a natural language processing model for an information domain, comprising: a processor; a scaffold component for deriving a scaffold for a natural language vocabulary from a source model of the information domain; a syntax rule component for applying a set of syntax rules that define concepts and relationships; and an extension component for extending the scaffold for a natural language vocabulary based on reference documents from the information domain to provide a natural language processing model for the information domain, wherein extending the scaffold involves forming clusters and evaluating terms for concepts and relationships.

[0027] Preferably, the present invention provides a system in which the syntax rule component for applying a set of syntax rules includes taking into account subject, predicate, object and changing the order for linguistic representation.

[0028] Preferably, the present invention provides a system in which the extension component for extending the basic framework includes a component for forming clusters of concepts / relationships in order to: select a preferred term as a concept or relationship; perform a keyword search for the preferred term in reference documents from the information domain; and provide an ordered set of possible terms for the preferred term.

[0029] Preferably, the present invention provides a system in which the component for forming clusters of concepts / relationships serves to: determine local N-grams; measure one or more measurable quantities of the N-grams and evaluate the N-grams.

[0030] Preferably, the present invention provides a system in which the extension component for extending the basic framework for the natural language vocabulary includes: a syntax rule generation component for deriving further syntax rules based on the reference documents from the information domain; and an intersection component for forming intersections from syntax rules and the concepts and relationships grouped into clusters.

[0031] Preferably, the present invention provides a system in which the extension component for extending the basic framework starts with an initial concept or relationship and iteratively moves outwards through neighboring concepts or relationship links in the source model; and refines the extended concepts and relationships by increasing evaluations.

[0032] Preferably, the present invention provides a system in which the extension component for extending the basic framework dynamically changes a strategy that repeats in iterations based on results. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] A preferred embodiment of the present invention will now be described merely as an example with reference to the accompanying drawings, which show: Fig. 1 is a flowchart of an embodiment of a method according to a preferred embodiment of the present invention; Fig. 2. A schedule of further details of the procedure by Fig. 1 according to a preferred embodiment of the present invention; Fig. 3 a block diagram of an embodiment of a system according to a preferred embodiment of the present invention; Fig. 4 is a block diagram of a computer system according to a preferred embodiment of the present invention; Fig. 5 a tree representation of a source model according to an aspect of a preferred embodiment of the present invention; Fig. 6 is a relationship representation of a source model according to an aspect of a preferred embodiment of the present invention. DETAILED DESCRIPTION OF PREFERRED FORMATIONS OF THE INVENTION

[0034] It will be clear that, for the sake of simplicity and clarity of illustration, elements in the figures are not necessarily drawn to scale. For clarity, the dimensions of some elements may be exaggerated in relation to others. Furthermore, where deemed useful, reference symbols may be repeated within the figures to indicate corresponding or similar features.

[0035] The following detailed description presents numerous specific details to provide a thorough understanding of the invention. However, those skilled in the art will understand that the present invention can also be implemented in practice without these specific details. In other instances, generally known methods, procedures, and components have not been described in detail so as not to obscure the focus on the present invention.

[0036] This paper describes a method and a system for generating and developing a natural language processing model tailored to a specific information domain. The method leverages the fact that in many cases a domain or industry model already exists—essentially an ontology of the business area itself.

[0037] This provides a cost-effective solution for creating and developing domain-specific natural language processing models from domain or industry models (such as those available and actively developed by IBM's Industry Models Group; IBM is a trademark of International Business Machines Corporation) or other forms of ontology. This initial model or ontology is referred to as the source model. This enables the rapid creation, deployment, and maintenance of a natural language processing model. The natural language processing model can then be used to analyze free-text documents within domain-specific solutions.

[0038] The described procedure and system utilize the structure and information of the domain model to generate an initial NLP model and to increase the number of sparse node labels in the initial NLP model (by definition, the nodes in the initial model can be labeled with a single term, with a preferred term being the recommended best practice) by means of keyword and verb clustering techniques and evaluation techniques.

[0039] The interactions between units within the model itself allow for the collection of textual representations of the concepts it contains or refers to. A number of well-known and simple empirical techniques can be used to match and extrapolate the dictionaries and rules. Without utilizing the environmental nature of the model—that is, without capturing outward-facing links (relationships or predicates) and target node types, and in turn their specifications—there would be a large number of confounding variables in the model.

[0040] Referring to Fig. Figure 1 represents a process flow diagram 100, an embodiment of the described procedure.

[0041] A basic framework of a natural language vocabulary, tailored to a specific information domain, can be derived from a source model. The source model can be any form of ontology, such as an industry or domain model, or another type of ontology within the specific information domain (the ontology could, for example, be a database schema). It is in the nature of these source models to provide the necessary data to derive the basic framework of natural language vocabularies tailored to their chosen information domain (banking, insurance, medicine, etc.). The procedure can utilize more than one source model, including external and publicly available models and data, such as Linked Open Data, WordNet, and so on.

[0042] The basic framework of the natural language vocabulary can be built directly from the source's conceptual model and optionally include synonyms from the business glossaries. Alternatively, the framework can be built using Linked Open Data (a project of W3C Semantic Web Education and Outreach), where the ontology classes in the source model can be checked for compatibility with a class described in the Linked Open Data cloud. This may be the case if the initial framework contains no sample data (only type descriptions), in which case Linked Open Data can be used to initially populate the framework.

[0043] This leads to sparsely populated dictionaries of concept terms and action verbs, which are combined and applied in a simple set of seed syntax rules, taking into account subject, predicate, object, and changing the order for linguistic representation or according to the rules of a language. 102 In some languages, for example, constructions might prevail in which the order 'John saw Mary' (subject-predicate-object) is expressed more naturally than 'saw John Mary'.

[0044] However, since the source models used to derive the initial framework intentionally describe the concepts and relationships in question using a limited set of preferred terms or labels (usually one), the resulting vocabulary or vocabularies must be expanded or "condensed." Concepts and relationships will be discussed separately in the following; however, a relationship can be considered a special type of concept.

[0045] Once such a basic framework exists, it must be expanded or condensed by forming keyword clusters and using other standard techniques with synonyms, verbs, and expressions. 103 Condensation is performed on the basis of a set of reference documents 113 selected from the information domain (e.g., emails, policies, guidelines, documents, reports, etc.). Condensation involves deriving descriptive terms for concepts and relationships from preferred terms in the basic framework and generating syntax rules from the syntax rules used as a starting point.

[0046] The condensed basic framework forms the basis of the NLP model, which is used to process all unstructured information within business activity and to semantically classify all mentions of concepts in documents of the area so that they can be used for text analysis.

[0047] In every medium-sized to large organization, numerous applications operate within the same business area (e.g., multiple payment systems, customer onboarding applications). An organization may have a constantly growing number of products and services, causing the amount of data within the organization to grow exponentially. To address this challenge, common architectures and a shared business language are needed to increase operational efficiency and improve time to market.

[0048] A common business language is needed to address these problem areas in the following activities: cross-departmental processing; application rationalization; gathering project requirements; transferring business operations to IT systems; compliance with legal requirements, etc. A multitude of synonyms exist between the individual business units, and it is very difficult to rationalize them; this problem is further exacerbated when the meaning of new legal requirements needs to be understood and adapted to.

[0049] Industry models include specific resources to meet this need for a common structured business language and therefore provide a way to cost-effectively generate industry-specific NLP models.

[0050] Each industry model offering can include multiple modeling levels for the knowledge of the sector it targets (e.g., banking, insurance, healthcare, etc.). Specifically, they can include: a. • Business glossaries for free text definitions of terms in this industry, including synonyms, acronyms, phrases, etc. b. • Conceptual models for high-level concept / construction definitions and their inherent relationship structures.

[0051] Referring to Fig. Figure 2 presents a process flow chart 200, an embodiment with further processing details of the described procedure.

[0052] An initial preferred term is selected as the term of an initial concept or relationship. 201. The basic framework of natural language vocabulary can be used to initially identify the passages in the text that yield further terminology and syntax, in order to facilitate a condensation process. The preferred term may be a keyword match from the basic framework of natural language vocabulary.

[0053] The preferred term can be used as a keyword search 202 within a set of reference documents taken from the business environment (e.g., emails, formal policies, guidelines, reports, etc.). The keyword search can be case-sensitive to achieve good recall of passages within documents, enabling the collection of a comprehensive vocabulary and syntax. Alternatively, if powerful terms offer value, a case-sensitive keyword search can be used. For example, a case-sensitive keyword search might be useful in some areas or documents where there may be significant overlap and a case-insensitive search would generate too much interference.

[0054] Local n-grams and / or phrases can be considered, that is, all overlapping word sets in which the number of words in the immediate vicinity of the preferred term is 'N'. In the case of concepts, the noun phrases containing the preferred term can be considered. Similarly, the relevant verb phrases for relationships can be considered.

[0055] A measurement can be determined for an N-gram 204. In the simplest embodiment, the frequency of the N-gram is counted, and N-grams are evaluated with a text proximity rating 205, which can of course include the frequency, but would not be limited to this specific measurement.

[0056] There are many possible metrics that could influence the creation of a ranking of potential terms. These can include well-known techniques such as topic-based decomposition, latent semantic analysis, etc., but also the use of synonyms and general resources like WordNet, as long as domain specificity is maintained. (It should be noted that using WordNet can introduce an undesirable tendency toward overgeneralization, but it could be helpful for the results or useful in user-moderated scenarios.)

[0057] Examples of possible measurement quantities include, but are not limited to, the following: The number of tokens for the preferred term (distance): When calculating frequencies, a tf / idf normalization can be applied to the document collection; while it is not necessary for good results, it can offer some improvement in certain cases; When including terms, it may also be necessary to consider resolution regarding simultaneous possible references, e.g., "they" = IBM in the context of a sentence describing an action that the important customers are taking... perhaps = "they will close their account if the interest is not credited by ...".

[0058] An example rule, perhaps: “IBM is an important customer”, would result in the general pattern being enclosed: by an unknown token (meaning outside the natural language vocabulary) or written entirely in CAPITAL LETTERS (“IBM”) or [English] title spelling or in a spelling with internal capitals (CaMeLCase), followed by an inflected form of the verb “to be” (“is”), followed by an optional article (“a”) and an adjective (“more important”) before reference to the preferred term for the concept (“customer”).

[0059] Favoring of unfamiliar or outside-the-vocabulary terms (where outside-the-vocabulary primarily refers to terms not part of the language in question, e.g., English).

[0060] Exclusion of so-called "stop words," e.g., "if," "a," "then." Stop word lists must be defined separately, possibly for each node, but probably for concepts in relation to relationships.

[0061] Similarly for verb relations. A verb relation can be considered a special type of concept and can be represented as such in terms of the node and link structure of the ontology or model.

[0062] Beyond what is included in the description of concepts, verb relational concepts offer the additional possibility of conducting a linguistic verb analysis. Verbs can be divided into groups that exhibit different valency frames. For example, the types of arguments a verb accepts and how they are ordered, which arguments are optional, and the "direction" of the verb; e.g., in "Paul kicked the dog," "kicked" has a subject and an object, and in this case, the direction is from Paul to the dog, i.e., Paul is the subject, and the dog is the object. Therefore, the concept {subject := (person, instance = Paul)} is linked to the concept {object := (animal, instance = the dog)} via a verb relation "kicked."

[0063] The described concept-based n-gram intersection based on an ontology / domain model greatly reduces the need for such detailed verb knowledge. However, where this knowledge is present or available through a processor such as a linguistic deep parser, the described approach can make use of it. Often, such approaches do not perform well in texts with frequent use of domain-specific language. Therefore, combining the described model-based intersection approach with concept-based n-grams, where the concept is considered in its more general form (i.e., also includes verb-relation concepts), can have a positive effect on the overall results and eliminate the need for complex, fundamental linguistic processing, such as deep parsing, which would be error-prone in such domain-specific cases.

[0064] An ordered set of possible terms can be generated for the concept node or relationship in the source model 206, but it will still be subject to quite a lot of interference.

[0065] Once this has been done for an initial concept or relationship, the procedure can be repeated for the next concept or relationship node. 207 In one embodiment, the iteration can move outward through the relationship links in the model and repeat for each of them, generating a set of terms for each concept and relationship and neighboring concepts. Alternatively, the process can jump to neighboring concept nodes before relationship nodes, for example, because of the relative simplicity of phrases consisting more or less of single nouns.

[0066] If the procedure iterates outwards from the initial concept node or initial relationship node in the source model 207, the set of concepts from previous nodes can be refined by increasing the ratings (additive or other weighted function) based on the intersection with the current node.

[0067] For iteration 207, any combination of concept or relationship can be used, including event-driven changes or a change of approach. Alternative strategies can be used, including coordination between heuristic-driven strategies.

[0068] For example, the rate of change in frequencies of potential synonyms can be monitored while the algorithm is running; this can influence which approach is used next, and this can change dynamically depending on the monitoring of the measurements. If the results are not found quickly enough, alternative strategies can be used. If too many terms are found and cross-references need to be created, applying an alternative strategy could help eliminate interference more quickly.

[0069] The subject-predicate-object context is used, and therefore the model has concept nodes (subjects) that are linked to other concept accounts (objects) via relationships (predicates). This leads to a rapid divergence of the best terms for a concept or relationship, and by using a rating threshold, a vocabulary of types can be built.

[0070] This vocabulary forms the basis of the NLP model, which is used to process all unstructured information within business operations and to semantically classify all mentions of concepts in any company documents so that they can be used for any of the current standard benefits of text analysis mentioned above.

[0071] Once many of the explicit labels for concept and relationship examples have been identified, annotations can be added and standard machine learning techniques can be applied to derive syntactic rules based on any exposed feature of the text, e.g., capitalization, word class or category, digits, classified tokens found by regular expressions, such as date / time, URLs, email, etc.

[0072] Adding annotations refers to the process of identifying certain terms (single words or multi-word phrases representing the concepts) in the preceding text, then searching the text for these terms and marking up the locations where they occur. The resulting annotated text can be used as input in a standard machine learning process, which would automatically generate a model that collects the important contextual features that indicate the occurrence of these terms. This can then be applied as an NLP model in itself, but in most cases, it would be better used as a way to generate more possible terms.

[0073] The annotated text then allows either an automatic model to be generated, as in the standard case of machine learning, or a specialist in the field to be presented with a statistical representation of the feature distribution in the environment of annotated instances, so that suitable syntax rules and / or important text features can be recognized and subsequently applied.

[0074] Machine learning can be either fully automated or user-moderated. In the user-moderated case, machine learning can only present the specialist with statistical observations, such as the predominance of a feature pattern that precedes an instance of a certain type.

[0075] For example, patterns such as "Mr." and "Dr." precede references to a person.

[0076] This provides the derivation of the syntax rules, either automated to some extent using various machine learning techniques or using user-moderated hybrid learning, where machine learning is used to discover statistically relevant feature patterns and present them to a user who is a specialist in a particular area.

[0077] The intersection of the derived syntax rules and the set of n-grams derived from a context spanned between any two or more elements of the model (concept or relationship) is used to extend the framework based on the reference documents. This enables the efficient generation of an accurate, high-reach NLP model.

[0078] The use of verb structures from a linguistic classification of verbs can also be used to control the intersection logic that can be applied to the set of N-grams.

[0079] Initial syntax rules can be derived to serve as a starting point (step 102 of Fig. 1) which can be used to initiate the process of generating additional syntax rules. Additional syntax rules can be generated from the reference documents 113. This is similar to condensing the preferred terms from the model's concept definition. A collection of documents containing these annotated concept examples allows us to deduce the syntax that represents these concepts.

[0080] The intersection of the generated syntax rules and the N-grams derived from the context (concept or relationship) spanned between any two or more elements of the model yields good results and enables the efficient production of an accurate, high-reach NLP model.

[0081] A preferred embodiment would include the above process in a "feedback loop", meaning that it can be applied iteratively with the output of itself (e.g., "IBM" is used in the above example, but this time a preferred term "customer" is chosen as the variable, which would likely lead to the discovery of "customer").

[0082] In an alternative embodiment, due to the relative simplicity of phrases consisting primarily of single nouns, the process can jump to adjacent concept nodes before relationship nodes.

[0083] Additionally, any combination of the subject-predicate-object structure derived from the relationships of the model can be used.

[0084] Generating syntax rules can be seen as comparable to generating new terms, the only difference being that more abstract features are taken into account. That is, the term derivation process considers only the actual words and phrases in the text, whereas the derivation of syntax rules considers more abstract features in the text, such as the features referenced in the paragraph above that mentioned machine learning.

[0085] The derivation of syntax rules occurs primarily after the sequence of Fig. The process is complete, including iteration 2. However, the use of initial syntax rules to aid in evaluation shows that the process can branch to generate them and apply another generated set along with the initial, starting set.

[0086] The formation of intersections occurs throughout the entire process, i.e., when evaluating the concepts in the initial iterations and later when learning and testing the validity of the derived rules in the next step.

[0087] Referring to Fig. Figure 3 shows a block diagram 300 illustrating an embodiment of the described method.

[0088] A system 310 is provided for generating an NLP model, which can create and maintain an NLP model for a domain 320. A domain 320 could be, for example, a business domain, medical domain, academic domain, etc. A domain 320 can have a source model 321, such as an industry model or another ontology that uses preferred terminology for the domain 320. The domain 320 can also contain reference documents 322, such as emails, policies, guidelines, documents, reports, etc.

[0089] System 310 for generating an NLP model can include a framework component 311 for deriving a basic framework of a natural language vocabulary from the domain's source model 321. System 310 for generating an NLP model can also include a syntax rule component 312 for combining concept labels and action verbs using a simple set of syntax rules as a starting point.

[0090] System 310 for generating an NLP model can also include an extension component 313 for condensing or extending the basic framework model from reference documents 322 of area 320 with synonymous words, verbs, and expressions by forming keyword clusters and other techniques. Extension component 313 can include a component 314 for forming concept / relationship clusters to expand the scope of concept / relationship labels and a component 315 for generating syntax rules. Extension component 311 can further include an intersection component 316 for creating intersections between the concepts / relationships grouped into clusters and the generated syntax rules to generate an NLP model 330.

[0091] A resulting NLP model 330 can be generated for area 320. NLP model 330 can be used to analyze documents and language used in that area, providing a language analysis. The NLP model can automatically "read" all text in an area and assign it to the appropriate place within the model, which provides a structure for the text to be analyzed.

[0092] Referring to Fig. Figure 4 contains an exemplary system for implementing aspects of the invention: a data processing system 400 suitable for storing and / or executing program code, comprising at least one processor 401 connected directly or indirectly to memory elements via a bus system 403. The memory elements may include: local memory used during the actual execution of the program code, mass storage, and cache memory providing temporary storage of at least a certain amount of program code to reduce the number of times the code needs to be retrieved from mass storage during execution.

[0093] The memory elements can include the system memory 402 in the form of a read-only memory (ROM) 404 and a random access memory 405 (RAM). An input / output basic system (BIOS) 406 can be stored in the ROM 404. The system software 407 can be stored in the RAM 405, which also contains the operating system software 408. The software applications 410 can also be stored in the RAM 405.

[0094] The System 400 can also include primary storage devices 411, such as a magnetic hard disk drive, and secondary storage devices 412, such as a magnetic disk drive and an optical disk drive. The drives and their associated computer-readable media provide non-volatile memory for computer-executable instructions, data structures, program modules, and other data for the System 400. Software applications can be stored on the primary and secondary storage devices 411 and 412, as well as in system memory 402.

[0095] The Computer System 400 can be operated in a networked environment using logical connections to one or more remotely located computers via a Network Adapter 416.

[0096] Input / output units 413 can be connected to the system either directly or via intermediate I / O control units. A user can input commands and data into the system 400 via input devices such as a keyboard, pointing device, or other input devices (for example, microphone, joystick, game controller, satellite dish, scanner, or the like). Output devices can include speakers, printers, etc. A display unit 414 is also connected to the system bus 403 via an interface, for example, a video adapter 415.

[0097] An example is given using the Financial Services Data Model (FSDM) as the source model. The FSDM is a classification model with an enterprise-wide vocabulary used to precisely define the meaning of the many concepts that constitute a financial institution's view of itself and its business environment. It allows all aspects or perspectives belonging to a particular business concept to be expressed in terms of the elements within a classification hierarchy.

[0098] Referring to Fig. Figure 500, a tree diagram, represents the classification hierarchy for the term "Involved Party" (IP) in the FSDM. The rectangles represent objects, and the blocks with rounded corners represent attributes. Attributes are a special type of relationship called "has_a," meaning "Involved Party has_a IP type individual" means "A is an individual." Attributes thus constitute a type of classification. The information in the tree diagram can be used directly at parent nodes or, in special cases, across peer nodes. The entire tree can be used to gather information on the concepts of interest.

[0099] The tree diagram 500 has a root node 501 for "Participant" with subordinate attribute nodes of "Legal Status" 502, "Role" 503 and "Type" 504. The further subordinate nodes of the node "Legal Status" 502 and the node "Role" 503 are generally represented as branches 505 and 506.

[0100] The node "Type" 504 has child nodes for "Organization" 507, "Organizational Unit" 508, and "Individual" 509. The node "Individual" 509 has child attribute nodes represented by "Marital Status" 510 and "Health Status" 511. The node "Marital Status" 510 has child nodes for "Married" 512, "Single" 513, and "Divorced" 514. The node "Health Status" 511 has child nodes for "No Restriction" 515, "Disability" 516, and "Unknown" 517.

[0101] These concepts, represented as nodes, can be related to each other via relationships, as in Fig. 6 shown. Fig. Figure 6 is a schematic representation of 600, in which concepts are linked through relationships. This creates the ability to explain different business scenarios.

[0102] In Fig. 6. An object, for example participant 601, location 602, event 603, and agreement 604, can be linked by relationships represented as arrows 611, 612, 613, 614, and 615. The arrows and the relationships they represent can run in one direction only or in both directions.

[0103] Various relationships 620 to 627 (dashed boxes with lowercase names) can be members of a relationship 611 to 615, which is indicated by a "U" sign.

[0104] Various concepts 630 to 635 (dashed boxes with names beginning with a capital letter) can be members of an object 601 to 604.

[0105] For example, the representation, starting with the party 601, indicates that a party 601 can be an individual 631 or an organization 630, and that the relationship they have to an event 603, such as a transaction 633 or a communication 634, is "initiated" 626. In other words, a party initiates a communication (e.g., John Smith sent an email).

[0106] Example use cases are listed below. Possible applications of NLP and text analysis in general are numerous and wide-ranging, and the following are just a few examples: Example 1

[0107] An insurance provider wants emails to be automatically related to relevant artifacts maintained in data repositories (i.e., customers, policies, claims, etc.).

[0108] An NLP model can be used to convert free text resources such as emails into data input by processing the text contained therein, marking references to relevant information components such as names, policy numbers, claim identifiers (IDs), and then assigning these to existing data records. Example 2

[0109] A model developer must perform a validity check to ensure that a model for banking applications supports the requirements of ISO 20022 / SEPA. SEPA is a new pan-European payment system that was implemented EU-wide between 2007 and 2010. There are specific requirements that financial institutions or payment processing service providers must meet to obtain SEPA registration and be SEPA compliant.

[0110] An NLP model could process the documents containing legal regulations, determine the necessary requirements for the model (concepts, relationships, etc.) and perform a validity check to determine whether a particular model provides the necessary support for these requirements for banking applications. Example 3

[0111] An IBM consultant wants to be able to more easily estimate the costs of upgrading Company X's existing infrastructure for a specific healthcare model.

[0112] Company X has commissioned IBM to reorganize its data infrastructure based on healthcare models provided by industry-specific models. A consultant is tasked with analyzing the existing capabilities of Company X's system and estimating the work involved in implementing an upgrade. An NLP model could be very helpful in analyzing reports on the structure, documentation on the architecture, and other relevant materials from Company X to identify concepts missing from the existing architecture.

[0113] A system for generating a natural language processing model can be provided as a service to a customer over a network.

[0114] The invention can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment comprising both hardware and software elements. In a preferred embodiment, the invention is implemented in software, which includes, but is not limited to, firmware, resident software, microcode, etc.

[0115] The invention can take the form of a computer program product accessible from a computer-usable and computer-readable medium that provides program code for use by or in conjunction with a computer or a system for executing commands. For the purposes of this description, a computer-usable or computer-readable medium can be any device that can contain, store, exchange data with, distribute, or transport the program for use by or in conjunction with the system, device, or unit for executing commands.

[0116] The medium can be an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system (or such device or unit), or a propagation medium. Examples of computer-readable media include semiconductor or solid-state memory, magnetic tape, removable computer disk, random access memory (RAM), read-only memory (ROM), magnetic hard disk, and optical disk. Current examples of optical disks include read-only compact disc (CD-ROM), read / write compact disc (CD-R / W), and DVD.

[0117] Changes and modifications may be made to the foregoing without deviating from the scope of the present invention.

Claims

[1] A method for generating a natural language processing model (330) for a specific information domain (320), the method comprising: - deriving (101) a basic framework of a natural language vocabulary for the specific information area (320) from a source model (111) of the specific information area, wherein the basic framework contains terms from the source model, wherein the source model contains classification hierarchies of the terms that include objects and attributes; - Creating a dictionary using the basic framework of natural language vocabulary; - applying (102) a set of syntax rules that define concepts and relationships; - expanding (103) the basic framework of the natural language vocabulary on the basis of reference documents (322) from the specific information area (320) in order to provide a natural language processing model (330) for the specific information area (320), wherein expanding the basic framework includes forming clusters and evaluating terms for concepts and relationships, as well as forming an intersection of syntax rules and clusters of concepts and relationships; and - Providing a natural language processing model (330) using the extended framework of the natural language vocabulary for the specific information domain (320) for analyzing documents in the specific information domain. [2] The method of claim 1, wherein deriving (101) a skeleton of a natural language vocabulary in the specific information area (320) uses preferred terms. [3] The method of claim 1 or claim 2, wherein applying (102) a set of syntax rules includes considering subject, predicate, object, and changing the order for the linguistic representation. [4] Method according to one of claims 1 to 3, wherein the extension (103) of the basic framework comprises: Selecting (201) a preferred term as a concept or relationship; Performing (202) a keyword search for the preferred term in reference documents (322) from the specific information area (320); and Providing (206) an ordered set of possible terms for the preferred term. [5] A method according to claim 4, comprising: Determining (203) local N-grams; Measuring (204) one or more measurands of the N-grams; and Evaluate (205) the N-grams. [6] Method according to one of the preceding claims, wherein the extension (103) of the basic framework of the natural language vocabulary comprises: Deriving further syntax rules based on the reference documents (322) from the specific information area (320); and Forming an intersection of the syntax rules and the clustered terms for concepts and relationships. [7] A method according to claim 6, comprising: Using verb structures from linguistic verb classes to control the intersection applied to clustered terms. [8] A method according to any one of the preceding claims, wherein extending (103) the framework begins with an initial concept or relationship and iterates outward through neighboring concepts or relationship links in the source model; and refines the extended notions of concepts and relationships by increasing the number of evaluations. [9] A method according to claim 8, comprising: Determine a deviation from best terms for a concept or relationship using a score threshold. [10] A method according to any one of the preceding claims, wherein extending (103) the framework dynamically changes a strategy repeated in iterations based on results. [11] Method according to one of the preceding claims, wherein the derivation (101) of a basic framework of a natural language vocabulary is based on more than one source model (111). [12] Method according to one of the preceding claims, wherein deriving (101) a skeleton of a natural language vocabulary uses freely accessible data to initially populate the skeleton, wherein the ontology classes of the source model are compared with the classes of freely accessible data. [13] A computer program stored on a computer-readable medium and loadable into an internal memory of a digital computer, comprising software code sections for, when executed on a computer, carrying out the method according to any one of claims 1 to 12. [14] A system (310) for generating a natural language processing model (330) for a specific information domain (320), comprising: a processor; a framework component (311) for deriving (101) a framework of a natural language vocabulary from a source model (321) of the specific information domain (320), wherein the framework contains terms from the source model (321), wherein the source model (321) contains classification hierarchies of the terms that include objects and attributes, and wherein the system (310) is adapted to form a dictionary using the framework of the natural language vocabulary; a syntax rule component (312) for applying (102) a set of syntax rules that define concepts and relationships; an extension component (313) for extending (103) the basic framework of the natural language vocabulary on the basis of reference documents (322) from the specific information domain (320) in order to provide a natural language processing model (330) for the specific information domain, wherein extending the basic framework includes forming clusters and evaluating terms for concepts and relationships and forming an intersection of syntax rules and clusters of concepts and relationships, and wherein the system (310) is adapted to provide a natural language processing model (330) using the extended basic framework of the natural language vocabulary for the specific information domain for analyzing documents in the specific information domain. [15] The system (310) of claim 14, wherein the syntax rule component (312) for applying (102) a set of syntax rules includes considering subject, predicate, object, and changing the order for the linguistic representation. [16] The system (310) of claim 14 or claim 15, wherein the extension component (313) for extending (103) the framework includes a component (314) for forming clusters of concepts / relationships for: Selecting (201) a preferred term as a concept or relationship; Performing (202) a keyword search for the preferred term in reference documents (322) from the specific information area (320); and Providing (206) an ordered set of possible terms for the preferred term. [17] The system (310) of claim 16, wherein the component (314) for forming concept / relationship clusters is for: Determining (203) local N-grams; Measuring (204) one or more measurands of the N-grams; and Evaluate (205) the N-grams. [18] System (310) according to one of claims 14 to 17, wherein the extension component (313) for extending (103) the basic framework of the natural language vocabulary includes: a component (315) for deriving further syntax rules based on the reference documents (322) from the specific information area (320); and an intersection component (316) for forming an intersection of the syntax rules and the clustered terms for concepts and relationships. [19] The system (310) of any one of claims 14 to 18, wherein the extension component (313) for extending the framework starts at an initial concept or relationship and iterates outward through neighboring concepts or relationship links in the source model; and refines the extended notions of concepts and relationships by increasing the number of evaluations. [20] The system (310) of any one of claims 14 to 17, wherein the extension component (313) for extending (103) the framework dynamically changes an iteratively repeated strategy based on results.

Images