System and method for automating the generation of an ontology from unstructured documents

The invention pertains to systems and methods for creation of ontologies. More particularly, the invention pertains to domain independent systems and methods that can automatically process a corpus of unstructured documents into one or more ontologies.

An ontology is a model of the important entities and relationships in a domain. Ontologies have been used in capturing the semantics of a system. Much of the focus in developing semantics and ontologies has been on handcrafting them. That may be appropriate for high level concepts that are domain-independent and which will be used by virtually all domain specific ontologies, but there is a need for a less time-consuming approach to generating domain ontologies, where extensive knowledge is already captured in unstructured texts.

Machine processable domain knowledge has been captured in a variety of formats ranging from expert systems to procedural code to UML models. Regardless of the representation paradigm, the knowledge has in general been hand-coded. On the other hand, large quantities of knowledge have been captured in unstructured texts such as conference proceedings, technical papers, books, and more recently web pages and electronic documents. The latter have generally been written by subject matter, or, domain experts while the former have traditionally required a tight collaboration between domain experts and knowledge engineers or programmers. It has generally been difficult or impossible to obtain the former from the latter by automated, computer-assisted means.

There are a number of ontology extraction tools mentioned in the literature including:

ASIUM—is a system based on clustering and cooperative methods that is targeted to extracting taxonomic relationships. ExtrAKT—extracts ontology information from structured database like Prolog knowledge bases. OntoLT—is a Protégé plugin for ontology extraction from annotated text. OntoLearn—learns domain concepts and detects taxonomic relationships among them to produce a domain concept forest. TextToOnto is part of KAON tool suite. TextToOnto extracts terms that can potentially be included in the ontology as concepts and it also does rule extraction either on the basis of proximity of two terms or by looking for common patterns like “term like term” to identify hierarchy relationships. XRA—uses software reverse engineering approach to extract an initial ontology from given data sources and their application programs. DOODLE-OWL—is a system for on-the-fly ontology construction. This system relies heavily on user interaction in this construction.

In view of the large quantities of information available in unstructured hard copy texts, web pages and electronic documents, there is a continuing need to be able to extract concepts or relationships from existing, unstructured documents. In view of the existing volumes of unstructured documentation and the rate of increase, domain independent automated methods and systems are preferable. It would also be preferable if required human involvement could be minimized.

FIG. 1 is a block diagram of a system which embodies the invention;

FIGS. 2A-2J are flow diagrams that illustrate processing in accordance with the invention;

FIGS. 3A-3U are exemplary screens that illustrate aspects of the processing of FIGS. 2A-2J ; and

FIGS. 4A-4B illustrate link extraction and ontology processing in accordance with the present invention of an exemplary medical document.

While embodiments of this invention can take many different forms, specific embodiments thereof are shown in the drawings and will be described herein in detail with the understanding that the present disclosure is to be considered as an exemplification of the principles of the invention, as well as the best mode of practicing same, and is not intended to limit the invention to the specific embodiment illustrated.

A system which embodies the invention can automatically process, identify and extract semantic domain information from technical unstructured documents. Extracted information is structured as an ontology to provide a semantic representation.

The identification and extraction of relevant information is domain independent. Documents which incorporate concepts from a variety of different domains can be structured into different, domain dependent ontologies using common, domain independent, processing in accordance with the present invention.

In one aspect of the invention; phrases of a predetermined maximum length can be extracted from a document. Core noun phrases can then be extracted. The relevant nouns or noun phrases in the document are automatically extracted, independent of the domain of the document. The extracted nouns and noun phrases represent the key concepts in the document.

Links can be extracted into a file. The information extracted from the document can then be structured into an ontology.

In yet another aspect of the invention, phrase concepts can be added to create a new ontology or added to an existing ontology. Core noun phrase concepts can be added followed by link concepts.

Embodiments of the invention make it possible to extract semantic models (ontologies) from a vast repository of unstructured text, which documents much of human knowledge. Drawing on this repository as well as subject matter expertise, the present methodology and system aid in the construction and maintenance of an ontology by the Subject Matter Expert (SME).

In another aspect of the invention, an initial ontology can be generated automatically. Subsequently, the concepts and links that were extracted but not integrated directly into the ontology can be presented to the SME in a drag-and-drop environment so that they can be easily folded into the ontology.

In a disclosed embodiment of the invention, information is automatically extracted from unstructured documents. An initial ontology can be automatically created using non-domain specific processing. A workflow environment can be provided which uses known ontology tools (such as publicly available software, PROTÉGÉ from Stanford Medical Informatics) for visualization, comparison and maintenance.

FIG. 1 illustrates a system 10 in accordance with the invention. Unstructured documents, web pages and the like, all without limitation, generally indicated at 12 are input to and processed by one or more programmable processors 16 . Text in various forms can be stored in database(s) 18 . Documents 12 can be combined into one file and processed, or, stored in separate files and processed all without limitation.

Processing can be carried out by software recorded on a computer readable medium, having various functions, generally indicated at 20 . Representative executable modules include software 22 that acquires input material such as documents, web pages, e-mails and the like all without limitation, in any available form from all available sources. Acquired input material can be stored in database(s) 18 for subsequent processing via database interface software 24 .

Acquired input materials can be initially processed by text extraction software 26 as discussed subsequently. Records output by extraction software 26 can then be automatically incorporated into an ontology created and/or maintained by domain independent ontology creation software 28 . Structured records can be stored in database(s) 18 .

The subject matter expert, SME, can interact with the software 28 and structured records in database(s) 18 via graphical user interface software 30 and keyboard/display devices 32 a, b . The SME can resolve issues left open by software 28 .

It will be understood that the software modules such as 22 - 30 can be executed by one or more processors which can communicate via one or more computer networks such as the Internet. The particular hardware/software configuration for a selected implementation is not a limitation of the invention.

FIGS. 2A-2J are flow diagrams that illustrate processing in accordance with the invention. FIGS. 3A-3J are exemplary screens that illustrate various aspects of the processing illustrated in FIGS. 2A-2J .

As those of skill in the art will understand, the processing illustrated in the flow diagrams of FIGS. 2A-2J could be implemented in the system 10 in various embodiments. For example and, without limitation, processor 16 and intake software 22 could be located adjacent to a corpus or plurality of documents to be processed. The intake software 22 could communicate with the extraction software 26 via one or more computer networks. Similarly, the ontology creation software 28 could communicate with the databases 18 also via one or more computer networks. Ontology review software 28 a can be used by the SME to review and/or compare ontologies.

FIG. 2A illustrates a domain independent method of processing 100 in accordance with the invention. A user logs in, step 102 , via a logon screen, FIGS. 3A and 3B . In a step 104 a user selects a function to either create a new session or select an existing session as illustrated in the screen of FIG. 3C . A user elects whether to extract concepts from text documents, add concepts to an ontology or review and compare ontologies as illustrated in the screen of FIG. 3D .

In a step 106 the user has elected to extract concepts from text documents. In step 108 the user has selected an input text directory, as illustrated on the screen of FIG. 3E , and has selected a text file a portion of which is displayed on the right half of the screen of FIG. 3E .

Subsequently, in step 110 on the screen of FIG. 3E , the user also sets phrase concept cutoff percentages and phrase length parameters. The cutoff percentage specifies a degree of frequency required to keep a phrase in a sorted list. Also in the screen of FIG. 3E the user can elect to extract phrase concepts from the selected text file, illustrated in step 112 . The phrase extraction process is illustrated in more detail in the flow diagram of FIG. 2B .

Subsequent to extracting the phrases from the selected documents, the user can, in step 114 , using the screen of FIG. 3F set a percentage cutoff for core noun phrases. The selected percentage cutoff establishes a degree of frequency required to keep the respective core noun phrases in a sorted list. Those that fail to exhibit the selected degree of frequency are discarded. The process of extracting core noun phrases can then be selected in the screen of FIG. 3F step 116 . Additional details of the process of extracting core noun phrases are illustrated in the flow diagram of FIG. 2C .

The goal of the process of extracting core noun phrases, step 116 is to identify textual elements that incorporate any number of adjectives, including zero, followed by nouns. Processing is implemented by use of files or lists of words, such as adjective words, barrier words and verb words. The barrier words and verbs are words that are not likely to be part of a core noun phrase. The adjective words are words that are usually used as adjectives.

The list of barrier words includes those words which force a separation of phrases. The adjective words list includes those words which primarily appear as adjectives. A word in this list may appear only as an adjective. A list of stop words can be provided. These are words which are to be ignored.

Link extraction is initiated in step 118 . As illustrated in the screen of FIG. 3G a user can establish cutoff percentages for subjects, and objects. Additionally in step 120 gaps, maximum words between subject and predicate, can be set by the user as well as maximum words between the predicate and an associated object. The user using the screen of FIG. 3G can initiate the link extraction process, step 122 . Details of the link extraction process are illustrated in FIG. 2D .

The goal of the link extraction process is to extract relationships and links of the form “subject-predicate-object”. These three components are driven by word lists where the subject and object words are generated from the input documents to be processed. Each extraction has a confidence level associated with it. A publicly available lexical database like WordNet organizes English words into sets of synonyms called synsets and WordNet is utilized in generating these word lists.

Completion of the extract links processing, step 122 , results in a file with an unstructured “flat” representation of the important concepts and relationships from the document. In step 116 , the processing to extract core noun phrases can be expected to load all information which is extracted from the document into the respective file.

Subsequent steps relate to establishing or expanding an ontology, steps 130 - 138 , to provide a structured representation of the information extracted from the document. Subsequently, in steps 140 - 148 various ontologies can be reviewed or compared.

In step 130 a user can select a tab to add previously extracted concepts to an ontology. The screen of FIG. 3H illustrates both a concepts directory and an ontology directory which can be selected by the user, step 132 . Additionally, the user can select as a desired action adding single-sense phrase concepts, core noun phrase, concepts or links.

The display of FIG. 3H illustrates contents of the phrase file, the phrases extracted in step 112 . In the upper right of the display of FIG. 3H three columns are presented. These are frequency, normalized frequency and the phrase. Phrases are sorted by frequency.

The phrases illustrated on the display of FIG. 3H are those which were extracted in step 112 from the document illustrated on FIG. 3E . In step 134 a new ontology can be created by selecting as the desired action, step one namely, adding single-sense phrase concepts. Additional processing details of step 134 are illustrated in the block diagram of FIG. 2E .

Subsequent to creating a new ontology, step 134 the user can add core noun phrases to that ontology in step 136 . The display of FIG. 3I illustrates on the upper right therein the core noun phrases file created in step 116 . The core noun phrases have been extracted from the document illustrated in FIG. 3E . The user specifies the core noun phrase file via selecting the appropriate entry in the lower right portion of that display.

The core noun phrases file illustrated in the upper right of FIG. 3I once again has three columns, frequency, normalized frequency and the core noun phrase. The phrases are sorted by frequency of occurrence.

The link concepts from the file previously created in step 122 can then be added to the ontology, step 138 . The display of FIG. 3J illustrates in the upper right a portion of the link text file from step 122 . The illustrated portions of the file are the confidence column and “subject” sorted by confidence value. Additional details of the processing of step 138 are illustrated in the flow diagram of FIG. 2J .

In step 140 a user can elect to review and compare ontologies, as illustrated in the screen of FIG. 3K . When so elected, the user in step 142 can select an ontology directory and file along with specifying whether an ontology is to be viewed or compared with another ontology.

In step 144 , as illustrated in the screens of FIGS. 3L , 3 M user can view a graphical representation of an ontology, which includes a previously created ontology based on an unstructured document. Additionally, the user can, as illustrated in the screen of FIG. 3N , view extracted links in the ontology. Finally, in the screen of FIG. 30 the user can view captured instance data.

Subsequently, in step 146 the user can compare two ontologies. The screens of FIGS. 3P-3U illustrate various aspects of the comparison. In FIG. 3P an exemplary second file is illustrated which will be compared with the previously structured file. In FIG. 3Q a user can select the “compare” tab or button as the next step in the process. In the screen of FIG. 3R a second ontology can be selected for comparison purposes. Finally, in the screen illustrated in FIG. 3S a comparison of the two ontologies is illustrated via a tree view. In the screen of FIG. 3T a comparison of the two ontologies is illustrated with a table view. Finally, in the screen of FIG. 3U a comparison of the links of the two ontologies is illustrated.

The flow diagram of FIG. 2B illustrates processing 200 for extracting phrases, step 112 . An exemplary text fragment 202 a is provided as well as exemplary phrases there in, 202 b , which can be extracted by the processing 200 .

In initial step 204 a plurality of input documents 12 can be combined into one file. Then, in step 206 all text between delimiters such as <tag>, </tag> is deleted. Remaining text is all converted to a lowercase representation.

In step 208 total word count is set to zero and list L 1 is set to empty. In step 210 the available text is split into sentence fragments or chunks based on barrier characters. Barrier characters include “.!?;” so that extracted phrases will not span any of the barrier characters.

If all sentence fragments have been processed in step 212 , then flow goes to step 214 where for each entry in the list L 1 the normalized frequency is determined. In step 216 all entries in List L 1 having a normalized frequency value below the preset cutoff limit are deleted. Finally in step 218 the list L 1 is sorted based on normalized frequency values to produce the extracted phrases output file and the process returns to step 114 , FIG. 2A .

If all sentence fragments have not been processed, step 212 the next sentence fragment is split into words based on non-word characters, step 222 . In step 224 the total word count is incremented by the number of words in the respective sentence fragment.

In step 226 all words which incorporate one or more numeric digits are discarded and the list L 2 is set to empty. In step 228 all sequences of words up to the user pre-specified maximum phrase length are identified as a phrase and added to the list L 2 .

In step 232 a determination is made as to whether all entries in list L 2 have been processed. If not, in step 234 the next phrase in list L 2 is selected.

In step 236 an evaluation is made as to whether the phrase is incorporated in the stop word list. Exemplary stop words include:

Because The A An About

Additional stop words could be incorporated as would be understood by those of skill in the art. If the phrase is in the stop words list, we check if all entries are processed, step 232 . Otherwise, in step 240 a determination is made as to whether the phrase is present in list L 1 . If not, in step 242 the phrase is added to list L 1 with a frequency value of 1. If the phrase is already in list L 1 , in step 244 the frequency value of that phrase in list L 1 is incremented by 1.

The flow diagram of FIG. 2C illustrates processing 300 for extracting core noun phrases, step 116 . Region 302 of FIG. 2C illustrates examples of identified core noun phrases from the text fragment 202 a . Further, as will become apparent from the subsequent discussion, the core noun phrase processing proceeds without prior knowledge on the part of the methodology 100 as to the relevant concepts of the text being processing. As a result, processing 100 is substantially domain independent.

In addition to be able to extract concepts from the subject documents, without prior identification thereof, acronyms or abbreviations can also be extracted from the text and treated as concepts. An example acronym is TCTO which appears in the list of extracted core noun phrases 302 . Similarly, “prosuper” while not in any English dictionary is a concept in the domain of the text fragment 202 a of FIG. 2B and can be extracted by the processing 300 .

In a step 306 previously extracted phrases in the CombinedFile.txt, step 112 are processed, if not previously stored. Text between tags is dropped and remaining text is converted to lower case, as in step 206 .

In step 308 total word count is initialized to zero and list L 1 is set to empty or null. Text is split in to sentence fragments based on barrier characters step 310 , see step 210 .

In step 312 a determination is made as to whether all sentence fragments have been processed. If yes, in step 314 , for each entry in the list L 1 normalized frequency is determined. In step 316 all entries in list L 1 with a normalized frequency below the user preset cutoff are deleted. Finally in step 318 the list L 1 of remaining entries is sorted based on normalized frequency and the process returns to step 118 , FIG. 2A .

In the event that the all sentence fragments have not been processed, step 312 , in step 322 each sentence fragment is split into words based on non-word characters. In step 324 the total word count is incremented by the number of words in the sentence fragment. In step 326 all words with numeric digits are discarded and list L 2 is set empty.

In step 328 core noun phrases are identified within the sentence fragment by starting from the left most word in the fragment. A core noun phrase corresponds to the maximal phrase “PHR1 PHR2” in sentence fragment where all words in PHR 1 are in the adjective word list 330 a . The words in PHR 2 are not in the adjective word list 330 a, the barrier word list 330 b or the verb word list 330 c . This processing can identify multiple core noun phrases from a single sentence fragment.

Where all entries in the list L 2 have not been processed, step 332 , the next phrase in the list L 2 is selected in step 334 , and then it is determined if the phrase is in list L 1 in step 336 . If not, that phrase is added to the list L 1 with a frequency value of 1, step 338 . Otherwise, the frequency value of that phrase in list L 1 is incremented by 1, step 340 .

The flow diagram of FIG. 2D illustrates processing 400 of link extraction step 122 . On FIG. 2D extracted links 402 are illustrated for the text fragment 202 a.

In an initial step 404 subject and object word lists 404 a , 404 b are generated from the file CombinedFile.txt.CNP.csv, from step 318 , based on previously established subject and object cutoff parameters. These subject and object word lists are domain specific and have been extracted from the document file. The ability to recognize and extract subject words and object words from the respective document(s) contributes to the versatility and flexibility of the link extraction process 122 .

Steps 406 , 410 correspond to steps 306 - 310 as previously discussed. If in step 412 all sentence fragments have been processed, the links of list L 2 are sorted based on descending confidence step 414 .

In step 416 a determination is made as to whether all entries on the list L 2 have been processed. If so, the process returns to step 130 . If not, the next entry is selected in step 418 .

In step 420 a determination is made as to whether “Seq” (phrase from subject thru object) has an odd number of parenthesis. If so, the processing goes to step 416 . Otherwise, in step 422 , a determination is made as to whether the components “subject”, “predicate” and “object” are already represented in list L 1 as part of a prior extraction. If so, processing continues to step 416 . Otherwise in step 424 , the entry is added to CombinedFile.txt.link.text and processing continues with step 416 .

In the event that all sentence fragments have not been processed step 412 the next sentence fragment is selected and the list L 2 is set to empty or null, step 428 . Step 430 specifies fragment processing to identify instances of “subject”, “predicate” and “object”.

Subjects and objects are elements that match an entry in the subject file 404 a and object file 404 b , respectively. As noted above the subject word file for 404 a and object word file 404 b are based on extracted subjects and objects from the core noun phrase file, step 404 . These elements have domain specific characteristics. WordNet also organizes the word senses in various categories called “Category ID”. If for some sub-sequence of the predicate, all the verb senses have the same category ID then that category ID is associated with the extracted link. The WordNet category ID is represented in the VerbCategorized.dat file 404 c.

For example, relative to exemplary phraseology 402 the word “prosuper” would appear in the subject word file 404 a having been extracted from the core noun phrase file. The number of words appearing between subject and predicate; and between predicate and object must be less than or equal to the gap cutoff parameters previously set. Other details are specified in step 430 .

Details of the step 134 , generating initial ontology from single-sense phrases in a file, are illustrated in the flow diagram of FIG. 2E of processing 500 . Addition of phraseology from the text fragment 202 a to the ontology is illustrated in 502 of FIG. 2E .

In step 504 a determination is made as to whether all entries in the CombinedFile.txt.PhrFreq.csv file, step 218 of FIG. 2B have been processed. If so, a determination is made in step 506 as to whether all entries in the instance list have been processed. If so, processing continues in step 136 of FIG. 2A , if not the next entry in the instance list is selected in step 508 . In step 510 , a determination is made as to whether a term exists in the ontology of which this entry can be an instance. If so, it is added as an instance and processing continues in step 506 . Otherwise, processing goes to step 506 .

When all of the entries in the combined file, step 504 have not been processed, the next entry is selected for processing, step 516 . If the next entry as an instance type in WordNet, step 518 , the entry E is added to the instance list step 520 . If not, an evaluation is made step 522 as to whether the entry E exists in the ontology. If not, an evaluation is made in step 524 as to the whether the entry E exists in WordNet. If not, in step 526 an evaluation is made as to whether the entry E is the same as its stemmed version. If it is, in step 528 entry E is set to be the stemmed version of E.

An evaluation is made in step 530 as to whether entry E has a single-sense which is a noun. If so the synset associated with entry E and its parent hierarchy are added to the ontology file as in step 532 .

Details of the step 136 , add core noun phrase concepts to ontology, are illustrated in the flow diagram of FIG. 2F , processing 600 . A determination is made in step 604 as to whether all entries in the core noun phrases file, step from 116 , have been processed. If not the next phrase in the file is processed step 606 .

A determination is made as to whether the phrase already exists in the ontology, step 608 . If not, an evaluation is made as to whether the phrase is a multi-word phrase step 610 . If not the phrase is added to list PossConcept, step 612 .

If the phrase is a multi-word phrase, each combination of elements P 1 P 2 are processed where P 1 and P 2 are separated by a space, step 616 . An evaluation is made as to whether element P 2 exists in the ontology, step 618 . If so, element P 1 is associated with P 2 in a list named PossOntSubClass, step 620 . Otherwise, in step 622 element P 1 is associated with P 2 in a list named PossConceptSubClass.

In the event that all entries in the file have been processed, step 604 , the respective lists PossOntSubClass, PossConceptSubClass and PossConcept are processed respectively in'steps 624 , 628 and 630 . Details of steps 624 , 628 and 630 are illustrated respectively in the flow diagrams of FIGS. 2G , 2 H, 2 I which respectively illustrate processing 700 - 1 , 700 - 2 and 700 - 3 . The details of the processing of FIGS. 2G , 2 H, 2 I would be readily understandable by those of skill in the art after a review thereof. No further discussion of those figures is needed.

Subsequent to processing the PossConcept list 630 , previously extracted links, step 122 are added to the ontology, step 138 . Details of the processing of step 138 are illustrated in the flow diagram of FIG. 2J , processing 800 .

A determination is made, step 804 as to whether any additional entries exist from the link file which need processing. If not, processing continues with step 140 .

Where there are additional entries, the next entry is selected, step 806 . A determination is made as to whether or not the predicate ‘Pred’ in the entry exists in the ontology, step 808 . If not an evaluation is made as to whether the predicate has a single sense which is a verb in WordNet, step 810 . If not an instance is created of LinksToReview in the ontology, step 812 . If desired a subject matter expert could review the created instance.

Alternately, in step 814 a WordNet word sense is created for each word in the verb synset in the ontology. Finally, in 816 an object property is created for the verb synset and all of its parents in the ontology.

Ontology processing and structuring of the document is completed as the execution of step 138 . If desired, the ontology can be reviewed and compared with other ontologies as discussed previously, step 140 .

FIGS. 4A and 4B illustrate results of processing a Chronic Cough Management Guideline. The links extracted are shown in FIG. 4A and a portion of the generated ontology is shown in FIG. 4B . These are generated using the processing previously discussed of FIG. 2A . The example of FIG. 4A-4B illustrates the flexibility and the domain independent characteristics of the processing 100 of FIG. 2A .

From the foregoing, it will be observed that numerous variations and modifications may be effected without departing from the spirit and scope of the invention. It is to be understood that no limitation with respect to the specific apparatus illustrated herein is intended or should be inferred. It is, of course, intended to cover by the appended claims all such modifications as fall within the scope of the claims.