Method and system for discovering knowledge from text documents using associating between concepts and sub-concepts

[0001] The invention relates generally to the field of natural language processing, text mining, and knowledge discovery. In particular, it relates to a method and a system for extracting and discovering knowledge from text documents.

[0002] Due to the recent advancement of information technology and the growing popularity of the Internet, a vast amount of information is now available in digital form in both the Internet and the Intranet environments. Such availability of information has provided many opportunities. In the commercial world for example, online information is an advantageous source of business intelligence that is crucial to a company's survival and adaptability in a highly competitive environment. Unfortunately, a user in this situation is usually faced with too much information and too little knowledge that is useful or actionable knowledge. The processes of extracting and discovering knowledge, or knowledge extraction and discovery, from text documents or the like textual data are thus very important tasks of considerable application potential and impact.

[0003] Conventional methods and systems of knowledge extraction and discovery from text documents typically focus on the extraction of information or meta-data from free-text documents. Meta-data, which are condensed and typically semi-structured representations of text content, can be considered as the raw form of knowledge and are essentially facts specified in the text documents. Meta-data do not include knowledge that is not mentioned explicitly in the text. In addition, there is usually too much information extracted by the conventional methods and systems and it is a painstaking process for a user to organize and discover wisdom from the extracted information.

[0004] Specifically, in U.S. Pat. No. 6,076,088 and U.S. Pat. No. 6,263,335, both entitled “Information Extraction System and Method using Concept-Relation-Concept (CRC) Triples” by Paik et al, systems are proposed for building subject knowledge bases in the form of Concept-Relation-Concept (CRC) triples from text documents. The systems can acquire new knowledge by automatically identifying new names, events, or concepts from text documents.

[0005] In International Patent Publication No. WO 01/01289 entitled “Semantic Processor and Method with Knowledge Analysis of And Extraction from Natural Language Documents” by Tsourikov et al, the use of natural language processing methods are proposed for the extraction of Subject-Action-Object (SAO) tuples from text documents upon a user request. The methods further include normalization and organization of SAO triplets into Problem Folders with Action-Object (AO) portions as the name of the folders containing a list of subjects. In International Patent Publication No. WO 01/82122 entitled “Expanded Search and Display of SAO Knowledge Based Information” by Tsourikov et al, the methods proposed by Tsourikov et al in WO 01/01289 are extended by proposing methods for normalizing SAO triplets through paraphrasing AOs.

[0006] A critical problem associated with the foregoing proposals lies in the common and attendant inability of the proposed systems and methods to derive new or hidden knowledge from text documents that is often the critical differentiating factor in gaining an edge over competitors.

[0007] There is therefore a need for a method and a system for knowledge extraction and discovery from text documents for addressing such a problem.

[0008] In accordance with a first aspect of the invention, there is provided a method for discovering knowledge from text documents, the method comprising the steps of:

• • [0009] extracting from text documents semi-structured meta-data, wherein the semi-structured meta-data includes a plurality of entities and a plurality of relations between the entities;
  • [0010] identifying from the semi-structured meta-data a plurality of key entities and a corresponding plurality of key relations;
  • [0011] deriving from a domain knowledge base a plurality of attributes relating to each of the plurality of entities relating to one of the plurality of key entities for forming a plurality of pairs of key entity and a plurality of attributes related thereto;
  • [0012] formulating a plurality of patterns, each of the plurality of patterns relating to one of the plurality of pairs of key entity and a plurality of attributes related thereto;
  • [0013] analyzing the plurality of patterns using an associative discoverer; and
  • [0014] interpreting the output of the associative discoverer for discovering knowledge.

[0015] In accordance with a second aspect of the invention, there is provided a computer program product comprising a computer usable medium having computer readable program code means embodied in the medium for discovering knowledge from text documents, the computer program product comprising:

• • [0016] computer readable program code means for extracting from text documents semi-structured meta-data, wherein the semi-structured meta-data includes a plurality of entities and a plurality of relations between the entities;
  • [0017] computer readable program code means for identifying from the semi-structured meta-data a plurality of key entities and a corresponding plurality of key relations;
  • [0018] computer readable program code means for deriving from a domain knowledge base a plurality of attributes relating to each of the plurality of entities relating to one of the plurality of key entities for forming a plurality of pairs of key entity and a plurality of attributes related thereto;
  • [0019] computer readable program code means for formulating a plurality of patterns, each of the plurality of patterns relating to one of the plurality of pairs of key entity and a plurality of attributes related thereto;
  • [0020] computer readable program code means for analyzing the plurality of patterns using an associative discoverer; and
  • [0021] computer readable program code means for interpreting the output of the associative discoverer for discovering knowledge.

[0022] In accordance with a third aspect of the invention, there is provided a system for knowledge discovery from free-text documents, comprising:

• • [0023] means for extracting semi-structured meta-data from the free-text documents;
  • [0024] means for identifying key entities and key relations from the semi-structured meta-data;
  • [0025] a knowledge base that defines the attributes of entities;
  • [0026] means for formulating patterns based on the key entities and the attributes of entities related to the key entities; and
  • [0027] means for analyzing the patterns for knowledge.

[0039] The foregoing problem is addressed by a method and a system described hereinafter for transforming meta-data or the like information extracted from text documents in a domain and thereby discovering knowledge that is new or previously undiscovered in the extracted information and the text documents.

[0040] A method and a system for knowledge discovery according to embodiments of the invention described hereinafter relate to the discovery of new or hidden knowledge from free-text documents in a domain. According to the embodiments, semi-structured meta-data are first extracted from unstructured free-text documents. The semi-structured meta-data typically comprise entities as well as the relationships between the entities known as relations. The embodiments involve the use of a domain knowledge base dependent on taxonomy, a concept hierarchy network, ontology, a database, or a thesaurus, from which attributes or the like features of the entities can be obtained. The knowledge discovery method or system then uncovers the hidden knowledge in the relevant domain by analyzing the relationships between the attributes and the entities, mapping from an attribute space to an entity space. A domain refers to an application context in which the embodiments operate or function. Relevant fields of application include knowledge management, business intelligence, scientific discovery, bio-informatics, semantic web, and intelligent agents.

[0041] With reference to FIG. 1, using a knowledge discovery method a knowledge discovery system 10 in accordance with an embodiment of the invention is described, comprising a meta-data extractor 20, an intermediate meta-data store 30, a meta-data filter 40, a domain knowledge base 50, a meta-data transformer 60, an associative discoverer 70, a knowledge interpreter 80, and a user interface 90.

[0042] The meta-data extractor 20 allows a user of the knowledge discovery system 10 to extract semi-structured meta-data from free-text documents, which can be stored in the meta-data store 30 on a permanent or temporary basis. The semi-structured meta-data can be in the Noun-Verb-Noun form as commonly referred to in the field of Natural Language Processing, or in the form of Concept-Relation-Concept (CRC) triples as proposed in U.S. Pat. No. 6,263,335, or in the form of Subject-Action-Object (SAO) tuples as proposed in International Patent Publication No. WO 01/01289, or in other known forms. A noun, a concept, a subject, or an object, is an entity or the like individual element in a domain. For example, an entity can be a company, a product, a person, a protein, and etc.

[0043] For purposes of brevity of description hereinafter, an entity is exemplified by a concept of which an attribute is a sub-concept. However, the embodiments of the invention are not restricted to application to concepts and sub-concepts but include application to other types of entities and attributes of the entities.

[0044] To identify knowledge of interest, the meta-data filter 40 identifies key concepts and relations based on the occurrence frequencies of the key concepts and users' preferences. The meta-data transformer 60 then converts a concept to a plurality of sub-concepts based on the domain knowledge base 50. For example in the case of companies, for which the concept is a company profile, the sub-concepts may be related to the company profile, such as CEO profile, countries of operations, business sectors, financial ratios and etc. The domain knowledge base 50 may be dependent on a conventional relational or object-oriented database, taxonomy, a thesaurus such as WORDNET, and/or a conceptual model such as one described in “Concept Hierarchy Memory Model”, published in “International Journal of Neural Systems”, Vol. 8 No. 3, pp. 437-446 (1996). The features or attributes of the concepts or sub-concepts as specified in the domain knowledge base 50 can be predefined manually or generated automatically through other term extraction or thesaurus building algorithms known in the art.

[0045] The associative discoverer 70 may embody a statistical method, a symbolic machine-learning algorithm, or a neural network model, capable of supervised and/or unsupervised learning. The neural network may comprise, for example, an Adaptive Resonance Theory Map (ARTMAP) system, such as one described in “Fuzzy ARTMAP: A Neural Network Architecture for Incremental Supervised Learning of Analog Multidimensional Maps, published in “IEEE Transactions on Neural Networks”, Vol. 3, No. 5, pp. 698-713 (1992), or an Adaptive Resonance Associative Map (ARAM) system, such as one described in “Adaptive Resonance Associative Map”, published in “Neural Networks”, Vol. 8, No. 3, pp. 437446 (1995).

[0046] The user interface 90 may comprise a graphical user interface, keyboard, keypad, mouse, voice command recognition system, or any combination thereof, and may permit graphical visualization of information groupings.

[0047] The knowledge discovery method can be executed using a computer system, such as a personal computer or the like processing means known in the art. The knowledge discovery system 10 can be a stand-alone system, or can be incorporated into a computer system, in which case the user interface 90 can be the graphical or other user interface of the computer system and the domain knowledge base 50 may be in any conventional recordable storage format, for example a file in a storage device, such as magnetic or optical storage media, or in a storage area of a computer system.

[0048] An exemplary implementation of an embodiment of the invention is a business intelligence system in which concepts may refer to companies, products, and people, relations may refer to launch events, key hires and etc, and sub-concepts may refer to company profile, product features, and personal profile stored in an enterprise database or taxonomy. The knowledge discovery system 10 in this context may serve to uncover hidden relations among company profiles and, as an example, stock price performance, or hidden relations among personal profile and company.

[0049] Yet another exemplary application of an embodiment of the invention is a scientific discovery system in which concepts may refer to genes, plants, or diseases, relations may refer to protein interaction, localization, or disorder association, and sub-concepts may refer to DNA sequences, plant attributes and etc. The knowledge discovery system 10 in this context can be used to uncover hidden relations between, as an example, DNA sequences and a specific disease, in term of identifying key DNA segments that have a strong link to the disease.

[0050] With reference to FIG. 2A, an instance of the knowledge discovery method in accordance with a further embodiment of the invention is described. In the knowledge discovery method, a meta-data extraction step 110 first scans text documents to extract meta-data, for example, in the form of concept-relation-concept 3-tuples. A meta-data filtering step 130 next removes irrelevant 3-tuples by focusing on the prominent or key concepts and relations that the user deems as important. Next, a meta-data transformation step 140 reads an external domain knowledge base 50 dependent on taxonomy, ontology, a database, or a concept hierarchy network and derives the sub-concepts for the concepts related to a key concept. A pattern formulation step 150 then forms training samples, each consisting of a vector representing sub-concept and a vector representing a key concept. An associative discovery step 160 subsequently processes vector pairs and learns the underlying associations in the form of ARAM clusters. A knowledge interpretation step 170 further extracts knowledge in the form of IF-THEN rules from ARAM.

[0051] With reference to FIG. 2B, an application of the knowledge discovery method for discovering the hidden relationships between product attributes and diseases, such as cancer, is described for illustrative purposes. The meta-data extraction step 110 scans consumer product reports to extract meta-data consisting of associations between product attributes and diseases. The meta-data filtering step 130 removes irrelevant 3-tuples by focusing on the causal relation between consumer products and cancer. The meta-data transformation step 140 reads external product knowledge dependent on taxonomy, ontology, or a database, and derives the product attributes for each product. The pattern formulation step 150 forms training samples consisting of a vector representing product attributes and another vector representing cancer. The associative discovery step 160 processes vector pairs and learns the underlying associations in the form of ARAM clusters. The knowledge interpretation step 170 extracts knowledge in the form of IF-THEN rules from ARAM. For this application, examples of the rules discovered may be the association of specific manufacturers with hazardous (cancer causing) products, and the identification of product ingredients, or combinations thereof, which consistently led to cancer.

[0052] The knowledge discovery method is described in greater detail hereinafter by way of exemplary methods or models for implementing each processing step.

Meta-Data Extraction

[0053] With reference to FIG. 3, which illustrates an exemplary flow diagram of a meta-data extraction process, the meta-data extractor 20 assumes that the input documents are in plain text format. If the input documents are in a different format, the required conversion is first performed in a pre-processing step 102. For example, menu bars or formatting specifications such as HTML tags as found on web pages are removed to keep only the main body of the content. Text in image formats is converted to plain text using an optical character recognition system. Speech audio signals are converted to text using a speech recognition system. Captions and transcriptions in video are converted to text using a character recognition system and a speech recognition system respectively.

Name Entity (NE) Recognition

[0054] A name entity recognition step 104 next extracts all entities such as person names, company names, dates and numerical data (e.g. 10,000) from the preprocessed texts and creates an index. The index stores the frequency of each entity together with the identities of the source documents. NE recognition also identifies different variations of the same entity, e.g. “George Bush”, “Bush, George” and “Mr. Bush”, through a co-reference resolution algorithm. All variations of an entity are reduced into a standard form and annotate the documents using this standard form for further processing.

NVN 3-Tuples Extraction

[0055] An NVN 3-tuples extraction step 106 then parses sentences and derives syntactic 3-tuples in the form of (NC, VC, NC), where NC is a Noun Clause and VC is a Verb Clause. NC and VC are extended forms of Noun Phrase and Verb Phrase respectively, defined by the regular expressions
NC:−(ADJP)?(NP)⁺((IN|VBG)*(ADJP|NP))*
VC:−(ADVP)?(VP)⁺(IN|ADVP)*(VP)*
where ADJP is an adjective, NP is a noun phrase, IN is a preposition, VBG is a Verb gerund, ADVP is an adverb, and VP is a verb phrase.

[0056] NVN 3-tuples offer a low-level representation for capturing the agent/verb/object relationships in text. As a first step in the extraction, text is tagged using a Part-of-Speech (POS) tagger. Then a rule-based algorithm is employed to extract the NVN 3-tuples. As an example, consider the sentence: “Bill Gates released Windows XP, an operating system for PC's”. The rule
{[NC1][VC1][NC2], [DT][NC3]}→(NC1,VC1,NC2), (NC2,IS-A,NC3)
when invoked on sentence would result in the NVN 3 tuples:

• • [0057] (Bill Gates, released, Windows XP), and
  • [0058] (Windows XP, IS-A, operating system for PC's).

[0059] The set of the parsing rules can be constructed and validated based on a collection of documents. The rules aim to take care of a large proportion of the major sentence forms. In situation when no rule is found to match with the sentence, a special rule is made use to extract only the (NC,VC,NC) 3-tuple that appears at the beginning of the sentence. In the above example, this corresponds to (Bill Gates, released, Windows XP). Though this is approximate, often it could extract the main action conveyed by the sentence.

Sense Disambiguation

[0060] Next, a sense disambiguation step 108 identifies the specific meanings of noun clauses (NC) and verb clauses (VC) through the use of WordNet. For every word, WordNet distinguishes between its different word senses by providing separate synsets and associating a sense with each synset. The context of the words in a NC/VC is then used to compute a distance measure and pick the correct word sense for the NC/VC.

Clustering and Unification

[0061] In a clustering and unification step 110, the disambiguated NCs are grouped through a clustering algorithm, as is known in the state of the art, based on their similarities. By this, the knowledge recovery method would also take care of most of the entities that have not been reduced to the standard form during NE Recognition due to inaccuracies in NE co-reference resolution.

[0062] In addition, the verb clauses are unified according to the respective meanings. For example, “causes”, “leads to”, and “results in” are all different forms of expressing a causal relation. Clustering of noun clauses and unification of verb clauses complete the meta-data extraction step 110 by transforming the syntactic based NVN 3-tuples to semantic based Concept-Relation-Concept (CRC) 3-tuples type of meta-data representation.

Meta-Data Filtering

[0063] As the meta-data extraction step 110 in FIG. 2A may produce too many CRC 3-tuples, the meta-data filtering step 130 allows a user to focus on subsets of the 3-tuples by identifying key concepts and relations for the purpose of knowledge discovery. Key concepts and relations can be provided directly by the user. For example, the user can identify “cancer” as a key concept if he/she is interested in discovering factors related to cancer. Alternatively, key concepts/relations can be identified automatically through simple statistical methods, as are known in the state-of-the-art. For example, a concept/relation can be referred to as a key concept/relation if it is contained in more than half of the CRC 3-tuples extracted. This approach enables a user to discover important concepts/relations previously unknown to the user.

Meta-Data Transformation

[0064] Given a set of CRC 3-tuples, (A₁,R,B), (A₂,R,B), . . . , and (A_n,R,B), produced by the meta-data extractor 20, where B is a key concept and R is a key relation identified by the meta-data filter 40, and A₁, A₂, . . . , A_n denotes the concepts that are related to B under relation R, the meta-data transformation step 140 first obtains the sub-concept representation of A₁, A₂, . . . , A_n from the domain knowledge base 50. The domain knowledge base 50, whose main purpose is to provide a sub-level representation for the key concepts identified, may be dependent on a conventional relational or object-oriented database, taxonomy, a thesaurus (such as WORDNET), and/or a conceptual model. Without loss of generality, each concept (A₁) can be represented by a M-dimensional vector of attributes or features,
A_i=(a_i1, . . . , a_iM)  (1)
where a_ij is a real-valued number between zero and one, indicating the degree of presence of attribute j in concept A_i. Note that in the above representation, the relations are position specific. In other words, (A,R,B) does not necessarily imply (B,R,A). For each concept A_i, the pattern formulation step 150 formulates an example consisting of the sub-representation of A_i and the associated concept B in the form of ({a_i1, a_i2, . . . , a_iM}/B) for processing in the associative discovery step 160 described below.
Associative Discovery

[0065] With reference to FIG. 4, there is provided one such neural network-based associative discoverer 70 as described in “Adaptive Resonance Associative Map”, published in “Neural Networks”, Vol. 8 No. 3, pp. 437446 (1995), which is an example of the associative discoverer 70. As described in the article cited above, ARAM is a family of neural network models that performs incremental supervised learning of clusters (pattern classes) and multidimensional maps of both binary and analog patterns. An ARAM system can be visualized as two overlapping Adaptive Resonance Theory (ART) modules consisting of two input fields F₁^a 71 and F₁^b 72 with a cluster field F₂ 73. For the knowledge discovery method described herein, the input field F₁^a 71 serves to represent the attribute vector A and the input field F₁^b 72 serves to represent the concept vector B. Each F₂ cluster node j is associated with an adaptive template vector w_j^a and a corresponding adaptive template vector w_j^b for learning the mapping from attributes to concepts. Initially, all cluster nodes are uncommitted and all weights are set equal to 1. After a cluster node is selected for encoding, it becomes committed.

[0066] With reference to FIG. 5, given an input vector A with an associated input vector B, the system first searches for an F₂ cluster J encoding a template vector w_j^a and a template vector w_j^b paired therewith that are closest to the input vectors A and B, respectively, according to a similarity function. Specifically, for each F₂ cluster j, the clustering engine calculates a similarity score based on the input vectors A and B, and the template vectors w_j^a and w_j^b, respectively. An example of a similarity function is given below as the category choice function, eqn. (2). The F₂ cluster that has the maximal similarity score is then selected and indexed at J.

[0067] With reference to FIG. 6, the system performs template matching to verify that the template vectors w_j^a and w_j^b of the selected cluster J match well with the input information vectors A and B, respectively, according to another similarity function, e.g. eqn. (3) below. If so, the system performs template learning to modify the template vectors w_j^a and w_j^b of the F₂ cluster J to encode the input vectors A and B, respectively. Otherwise, the cluster is reset and the system repeats the process until a match is found. The detailed algorithm is given below.

[0068] The ART modules used in ARAM may be of a type that categorizes binary patterns, analog patterns, or a combination of the two patterns (referred to as “fuzzy ART”), as is known in the art. Described below is a fuzzy ARAM model composed of two overlapping fuzzy ART modules.

[0069] Fuzzy ARAM dynamics are determined by the choice parameters α^a>0 and α^b>0; the learning rates β^a in [0,1] and β^b in [0,1]; the vigilance parameters ρ^a 74 in [0,1] and ρ^b 75 in [0,1]; and a contribution parameter γ in [0,1]. The choice parameters α^a and α^b control the bias towards choosing a F₂ cluster whose template vectors have a larger norm or magnitude. The learning rates β^a and β^b control how fast the template vectors w_j^a and w_j^b adapt to the input vectors A and B, respectively. The vigilance parameters ρ^a and ρ^b determine the criteria for a satisfactory match between the input vectors A and B and the template vectors w_j^a and w_j^b, respectively. The contribution parameter γ controls the weighting of contribution from the F₁^a and F₁^b fields when selecting an F₂ cluster.

[0070] With reference to FIG. 7, the dynamics of the associative discoverer 70 is described by way of a flow diagram. Given a pair of F₁^a and F₁^b input vectors A and B, for each F₂ node j, a category choice process 202 computes the choice function T_j as defined by
T_j=γ|A^w_j^a|/(α^a+|w_j^a|)+(1−γ)|B^w_j^b|/(α^b+|w_j^b|)  (2)
where, for vectors p and q, the fuzzy AND operation is defined by (p^q)_i=min (p_i,q_i), and the norm is defined by |p|=Σ_i p_i. The system is said to make a choice when at most one F₂ node can become active. The choice is indexed at J by a select winner process 204 where T_J=max {T_j: for all F₂ nodes j}.

[0071] A template matching process 206 then checks if the selected cluster represents a good match. Specifically, a check 208 is performed to verify if the match functions, m_J^a and m_J^b, meet the vigilance criteria in their respective modules:
m_J^a=|A^w_J^a|/|A|≧ρ^a and m_J^b=|B^w_J^b/|B|≧ρ^b.  (3)

[0072] Resonance occurs if both criteria are satisfied. Learning then ensues, as defined below. If any of the vigilance constraints is violated, mismatch reset 212 occurs in which the value of the choice function T_J is set to 0 for the duration of the input presentation. The search process repeats, selecting a new index J until resonance is achieved.

[0073] Once the search ends, a template learning process 210 updates the template vectors w_J^a and w_J^b, respectively, according to the equations:
w_J^a(new)=(1−β^a)w_J^a(old)+β^a(A^w_J^a(old))  (4)
and
w_J^b(new)=(1−β^b)w_J^b(old)+β^b(B^w_J^b(old)  (5)
respectively. For efficient coding of noisy input sets, it is useful to set β^a=β^b=1 when J is an uncommitted node, and then take β^a<1 and β^b<1 after the cluster node is committed. Fast learning corresponds to setting β^a=β^b=1 for committed nodes.

[0074] At the start of each input presentation, the vigilance parameter ρ^a equals a baseline vigilance ρ^a. If a reset occurs in the cluster field F₂, a match tracking process 214 increases ρ^a until it is slightly larger than the match function m_J^a. The search process then selects another F₂ node J under the revised vigilance criterion. An exemplary setting for the other parameters is as follows: α^a=α^b=0.1, β^a=β^b=1, ρ^a0.5, ρ^b=1, and γ=0.5.

[0075] Through its supervised learning procedure, the associative discoverer 70 learns hidden relationship in terms of mapping between the attribute set {a_i1, a_i2, . . . , a_in} and the concept B. The knowledge uncovered may be in the form of {f_i1, f_i2, . . . , f_in}→B where {f_i1, f_i2, . . . , f_in},a subset of {a_i1, a_i2, . . . , a_in}, indicates the key features that are related to B.

Knowledge Interpretation

[0076] As the last step of the knowledge discovery method, the Knowledge Interpretation step 170 extracts symbolic knowledge in the form of IF-THEN rules from an ARAM. There is provided one such rule extraction algorithm as described in “Rule Extraction: From Neural Architecture to Symbolic Representation”, published in “Connection Science”, Vol. 7, No. 1, pp. 3-26 (1995). Referring to FIG. 8, in a fuzzy ARAM network, each cluster node in the F₂ field roughly corresponds to a rule. Each node has an associated weight vector that can be directly translated into a verbal description of the antecedents in the corresponding rule. Each such node is also associated to a weight template vector in the F₁^b field, which in turn encodes a prediction. Learned weight vectors, one for each F₂ node, constitute a set of rules that link antecedents to consequences. The number of rules equals the number of F₂ nodes that become active during learning.

[0077] However, large databases typically cause ARAM to generate too many rules to be of practical use. To reduce the complexity of fuzzy ARAM, a rule pruning procedure aims to select a small set of rules from trained ARAM networks based on their confidence factors. To derive concise rules, an antecedent pruning procedure aims to remove antecedents from rules while preserving accuracy.

Rule Pruning

[0078] The rule pruning algorithm derives a confidence factor for each F₂ cluster node in terms of its usage frequency in a training set and its predictive accuracy on a predicting set Usage and accuracy roughly corresponds to support and confidence as used in the field of associative rule mining, respectively. The confidence factor identifies good rules with nodes that are frequently and correctly used. This allows pruning of ARAM to remove rules with low confidence. Overall performance is actually improved when the pruning algorithm removes rules that were created to handle misleading special cases.

[0079] Specifically, the pruning algorithm evaluates a F₂ cluster j in terms of a confidence factor C_j:
C_j=γU_j+(1−γ)A_j,  (6)
where U_j is the usage of node j, A_j is its accuracy, and γ in [0,1] is a weighting factor.

[0080] For a cluster j that predicts outcome k, its usage U_j equals the fraction of training set patterns with outcome k coded by node j (F_j), divided by the maximum fraction of training patterns coded by any node J (F_j):
U_j=F_j/max{F_j}.  (7)

[0081] For a cluster j that predicts outcome k, its accuracy A_j equals the percent of predicting set patterns predicted correctly by node j (P_j), divided by the maximum percent of patterns predicted correctly by any node J (P_J) that predicts outcome _k:
A_j=P_j/max{P_J: node J predicts outcome k}.  (8)

[0082] After confidence factors are determined, clusters can be pruned from the network using one of following strategies:

[0083] Threshold Pruning—This is the simplest type of pruning where the F₂ nodes with confidence factors below a given threshold τ are removed from the network. A typical setting for τ is 0.5. This method is fast and provides an initial elimination of unwanted nodes. To avoid over-pruning, it is sometimes useful to specify a minimum number of clusters to be preserved in the system.

[0084] Local Pruning—Local pruning removes clusters one at a time from an ARAM network. The baseline system performance on the training and the predicting sets is first determined. Then the algorithm deletes the cluster with the lowest confidence factor. The cluster is replaced, however, if its removal degrades system performance on the training and predicting sets.

[0085] A variant of the local pruning strategy updates baseline performance each time a cluster is removed. This option, called hill-climbing, gives slightly larger rule sets but better predictive accuracy. A hybrid strategy first prunes the ARAM systems using threshold pruning and then applies local pruning on the remaining smaller set of rules.

Antecedent Pruning

[0086] During rule extraction, a non-zero weight to an F₂ cluster node translates into an antecedent in the corresponding rule. The antecedent pruning procedure calculates an error factor for each antecedent in each rule based on its performance on the training and predicting sets. When a rule makes a predictive error, each antecedent of the rule that also appears in the current input has its error factor increased in proportion to the smaller of its magnitudes in the rule and in the input vector. After the error factor for each antecedent is determined, a local pruning strategy, similar to the one for rules, removes redundant antecedents.

Quantizing Weight Values

[0087] When learning analog patterns or with slow learning, ARAM learns real-valued weights. In order to describe the rules in words rather than real numbers, the feature values represented by weights w_j^a are quantized. A quantization level Q is defined as the number of feature values used in the extracted fuzzy rules. For example, with Q=3, feature values are described as low, medium, or high in the fuzzy rules. Quantization by truncation divides the range of [0,1] into Q intervals and assigns a quantization point to the lower bound of each interval; i.e., for q=1, . . . , Q, let V_q=(q−1)/Q. When a weight w falls in interval q, the algorithm reduces the value of w to V_q. Quantization by round-off distributes Q quantization points evenly in the range of [0,1], with one at each end point; i.e., for q=1, . . . , Q, let Vq=(q−1)/(Q−1). The algorithm then rounds a weight w to the nearest Vq value.

[0088] The embodiments of the invention are preferably implemented using a computer, such as the general-purpose computer shown in FIG. 9, or group of computers that are interconnected via a network. In particular, the functionality or processing of the knowledge discovery system and method of FIGS. 1 to 8 may be implemented as software, or a computer program, executing on the computer or group of computers. The method or process steps for acquiring, sharing and managing knowledge and information within an organization are effected by instructions in the software that are carried out by the computer or group of computers. The software may be implemented as one or more modules for implementing the process steps. A module is a part of a computer program that usually performs a particular function or related functions. Also, a module can also be a packaged functional hardware unit for use with other components or modules.

[0089] In particular, the software may be stored in a computer readable medium, including the storage devices described below. The software is preferably loaded into the computer or group of computers from the computer readable medium and then carried out by the computer or group of computers. A computer program product includes a computer readable medium having such software or a computer program recorded on it that can be carried out by a computer. The use of the computer program product in the computer or group of computers preferably effects an advantageous system for acquiring, sharing and managing knowledge and information within an organization in accordance with the embodiments of the invention.

[0090] The system 28 is simply provided for illustrative purposes and other configurations can be employed without departing from the scope and spirit of the invention. Computers with which the embodiment can be practiced include IBM-PC/ATs or compatibles, one of the Macintosh™ family of PCs, Sun Sparcstation™, a workstation or the like. The foregoing is merely exemplary of the types of computers with which the embodiments of the invention may be practiced. Typically, the processes of the embodiments, described hereinafter, are resident as software or a program recorded on a hard disk drive (generally depicted as block 29 in FIG. 16) as the computer readable medium, and read and controlled using the processor 30. Intermediate storage of the program and any data may be accomplished using the semiconductor memory 31, possibly in concert with the hard disk drive 29.

[0091] In some instances, the program may be supplied to the user encoded on a CD-ROM or a floppy disk (both generally depicted by block 29), or alternatively could be read by the user from the network via a modem device connected to the computer, for example. Still further, the software can also be loaded into the computer system 28 from other computer readable medium including magnetic tape, a ROM or integrated circuit, a magneto-optical disk, a radio or infra-red transmission channel between a computer and another device, a computer readable card such as a PCMCIA card, and the Internet and Intranets including email transmissions and information recorded on websites and the like. The foregoing is merely exemplary of relevant computer readable mediums. Other computer readable mediums may be practiced without departing from the scope and spirit of the invention.

[0092] In the foregoing manner, a system and a method for transforming meta-data or information extracted from text documents and thereby discovering knowledge that is new or previously not mentioned in the extracted information and the original documents are described. Although only a number of embodiments of the invention are disclosed, it will be apparent to one skilled in the art in view of this disclosure that numerous changes and/or modification can be made without departing from the scope and spirit of the invention.