Towards Acquiring Case Indexing Taxonomies From Text
Contents
1. is a repair instruction After step 3 all statements from the text are available in logical form We propose to induce a classifier to extract features represented in logical form To do this we will obtain training data by asking a user to classify sampled sentences as features or non features from a given text document The predicates and arguments in the corresponding logical forms will serve as the features and we will select an appropriate learning algorithm through an analysis of the learning task We will then use the trained classifier to extract features from all the text Our proposed approach differs from SMILE Briininghaus and Ashley 1999 in that FACIT performs feature extraction and case indexing while SMILE performs only the latter Also FACIT s indexing strategy differs greatly see step 6 FACIT s feature extraction method differs from IE approaches that use patterns templates e g Weber et al 1998 or induction to extract cases from text Rather than using a large library of domain specific IE patterns FACIT uses one or more classifiers to extract features Furthermore as we illustrate in step 5 features extracted by IE techniques do not lend themselves to generalization Thus FACIT requires less feature engineering and will likely yield systems that have higher recall and precision performance than shallow NLP approaches in situations where the characterizing features are not known a priori 5 Feature or2. Solution _ Clean the print head See page 9 8 Figure 3 Example semi structured case text from the Multipass domain document CSSI 1998 Transforming text into its logical form involves a two step process that includes syntactic parsing and semantic interpretation Syntactic parsing assigns part of speech and sentence structure using a grammar and a lexicon The sentence structure represents the grammatical structure comprising a hierarchical relationship between the terms and phrases of a sentence For example Figure 4 shows the parse of the sentence Data from computer is not printed Although this example has implicit tabular structure FACIT does not require any implicit structure to extract case indices S OBJ NP VP KE VP Data from computer is not printed N P N V V Figure 4 A syntactic parse for a printing domain sentence Syntactic parsers are categorized as either shallow or deep Deep parsers search for and enumerate all potential parses based on the grammar and lexicon e g CMU s Link Parser Link 2003 Depending on a sentence s length and complexity it may have thousands of parses which can yield considerable sentence structure ambiguity that must be resolved by semantic interpretation Generating all parses and selecting a valid parse among them can be computationally expensive However all potential parses must be considered to ensure that the valid parse will be found Shallow parsers
3. INSTANCE_OF DATA data_1 AND INSTANCE_OF COMP_INSTR computer_1 AND NOT PRINTED HUMAN data_1 PRINTING_INST AND NATIVE OF computer_1 data_1 Figure 5 Logical form derived from the parse in Figure 4 Sentences with different grammatical structure but the same meaning must have or must be reducible to the same logical form For example the following sentences would yield the same logical form or meaning as the sentence in Figure 4 except for new variable instantiations Data sent from the printer to the computer is not printed Data is not printed by the printer Multipass is not printing data from the computer A large amount of surface syntactical variation of NL text is eliminated by transforming it into its logical form Furthermore predicate calculus operations are applicable to logical forms thereby enabling symbolic reasoning We propose to use these operators for generalizing and selecting features in step 5 of FACIT FACIT creates the logical form from the syntactic parse as follows 1 Look up senses in the lexicon All concepts indexed by the terms in the sentence are retrieved from the LKB For the sentence in Figure 4 our approach retrieves the concepts DATA COMP_INSTR and NOT for the terms data n computer n and not d Also retrieved are OCCUR_LOCATED and NATIVE_OF for the term from p and PRINT_EVENT for the term print v which was obtained by morphologically parsing the term printed v Clearly c
4. 5500 manual reveals new terms and concepts for updating the lexicon For example these include noun terms and phrases such as Multipass C 5000 sheet feeder or automatic document feeder power cord and interface cable Also compound terms like sheet feeder can be compositionally interpreted as an INSTRUMENT PART for feeding sheets CDW s output is used to manually create new concepts as exemplified in Figure 2 which can be related to existing concepts within the SO SOs can represent considerable lexical and domain knowledge using inheritable objects and relations For example Figure 2 shows that MULTIPASS C 5500 is a type of PRINTING_INSTRUMENT that human agents use to print INFORMATION and is an artifact made by the organization CANON_INC It includes PARTs such as MULTIPASS C 5500 lt SHEETFEEDER lt terms lt Multipass C 5500 n terms lt sheet feeder n ADF n gt unit n gt type_of lt FEEDING_INSTR gt type_of lt PRINTING _INSTR gt attributes lt attributes lt SIZE this unspecified SIZE this unspecified COLOR this unspecified gt COLOR this unspecified gt constituents lt constituents lt PART this PAPER _GUIDE gt PART this SHEETFEEDER behaviors lt PART this OP_PANEL FEED ACT HUMAN PAPER this PART this POWER CORD gt MULTIPASS C 5500 gt behaviors lt creative events lt PRINT_ACT HUMAN INFO MAKE ACT CANON_INC this gt gt this gt MOVE_ACT HUMAN INFO
5. g Weber et al 1998 Briininghaus and Ashley 1999 2001 4 Implementation Status In this section we summarize the status of the Java tools we have implemented for each of FACIT s steps 1 Updating a semantic lexicon We implemented the SO representation the Sublanguage Ontology Editor and the CDW to populate and update sublanguage ontologies An SO implements a lexical representation in an XML structured repository that includes our extensions to GL theory Gupta and Aha 2003 We are currently extending SOs to support both syntactic and semantic morphological processing that will increase the robustness of semantic interpretation CDW processes English text to help a knowledge engineer update SOs with domain specific terms and their related concepts CDW discovers concept elements such as the terms phrases acronyms and abbreviations for indexing SO concepts CDW currently operates in a standalone mode We will integrate these tools to increase the efficiency of SO updating tasks 2 Syntactic parser We implemented JLink an interface to CMU s easily available Link Parser Link 2003 which supports syntactic parsing We integrated this with our Java implementation of Brill s 1995 part of speech tagger to help efficiently select better parses We may later replace this with a suitable probabilistic parser 3 Semantic Interpreter Our SO driven semantic interpreter is a preliminary implementation that operates on JL
6. gt creative events lt MAKE ACT CANON_INC this gt gt Legend CONCEPTS Slotnames lt values gt tinherited slot part of speech e g n Noun Figure 2 Two MULTIPASS sublanguage ontology concepts SHEETFEEDER which in turn includes a part PAPER GUIDE The domain knowledge acquired during this phase will be the basis for feature organization step 5 In addition the noun term unit was added as a synonym for MULTIPASS Other senses of unit that are irrelevant to the selected application can be suppressed to prevent unnecessary ambiguity resolution overhead in step 3 In technical domains nouns representing components parts and names account for the majority of lexicon updates 2 Syntactically parse the source text Although case extraction methods often assume each document contains a case this depends on the document type For example troubleshooting manuals often contain information in a tabular format from which cases must be constructed prior to encoding This preprocessing step not shown in Figure 1 must be performed prior to syntactic parsing Thus we developed the Document Extraction Workbench to help manually extract cases into arbitrarily complex structures For example it can be used to create a case with fields corresponding to a source document s column headings Figure 3 displays a preprocessed input for FACIT Problem Data from the computer is not printed Cause The print head unit may need cleaning
7. solve similar new decision problems Aamodt and Plaza 1994 Conversational CBR CCBR is a CBR methodology that engages a user in a question answer dialog to retrieve cases Aha et al 2001 It has been successfully deployed in many help desk and troubleshooting applications Taxonomic CBR enhances CCBR by exploiting features organized into taxonomies to shorten user adaptive conversations and improve case retrieval performance Gupta 2001 Gupta et al 2002 A key challenge for applying CBR is acquiring cases from text documents e g manuals reports logs which is a focus of Textual CBR Ashley and Lenz 1998 For each case these systems must determine or be told which of the predefined features to use as indices When the mapping of indexing features to text cases is simple a bag of words approach can be used to automate feature extraction e g Burke et al 1997 However when this mapping is complex these features must be manually identified e g Weber et al 1998 Br ninghaus and Ashley 2001 Unfortunately this approach fails when the features are not known a priori as is true for complex troubleshooting applications and many other domains To our knowledge Copyright 2004 American Association for Artificial Intelligence www aaai org All rights reserved this feature acquisition problem has not been addressed previously The problem is further compounded by Taxonomic CBR s need to organize features into subsum
8. use statistical memory based e g Zavrel and Daelemans 1999 and or data based techniques to efficiently return one or a few top ranked parses but they return only constituent phrases and a partial syntactic structure This fast technique has been used in information retrieval and IE applications whose needs can be met by a shallow parse output However using shallow parsing for feature extraction and assignment is problematic because The likelihood of finding a valid parse can be unacceptably low Jt shifts and increases the burden of knowledge engineering to the development of IE patterns which provide limited domain knowledge and cannot be effectively reused to aid similarity assessment Because case index acquisition can be an off line process in our domains we use deep parsing Furthermore FACIT eliminates the use of domain specific IE patterns As shown later the domain knowledge acquired and stored in LKBs such as an SO can be effectively reused for similarity assessment To this end we have adapted the Link Parser Link 2003 to perform deep parsing It degrades gracefully when presented with ill formed text by allowing broken links or structures 3 Semantically interpret the text Semantic interpretation transforms the grammatical form or the syntactic parse into a logical form which uses predicate argument structures to represent the meaning of sentences contained in the text as propositions see Figure 5
9. Q Lenz et al 1998 IEs Indexing Assistant Concept Taxonomy None Baudin amp Waterman 1998 Minor amp Hiibner 1999 SMILE Briininghaus amp Ashley 1999 SMILE AutoSlog Briininghaus amp Ashley 2001 S IEs SE SE E SE SE IDS Yang et al 2003 None Generative None FACIT Gupta amp Aha this paper text documents Source refers to the source documents While most methods use lexicons that are sense enumerative FACIT s lexicon is generative see Section 3 Several methods use patterns templates to assign indexing features while more sophisticated methods e g SMILE automatically construct these patterns In contrast FACIT does not use patterns to extract features Most methods use a bag of words approach to represent concepts in the text An exception is propositional patterns Brininghaus and Ashley 2001 which provide an abstraction mechanism In contrast FACIT uses canonical logical forms to represent text that may contain features this enables reasoning on the information content of the source documents Most methods use an attribute value feature index representation such as a set of fields in a template although case retrieval networks Lenz et al 1998 have also been used In contrast FACIT derives feature subsumption taxonomies in addition to features Invariably developers and or experts serve multiple roles during the feature assignment process This is also true for FACIT b
10. Towards Acquiring Case Indexing Taxonomies From Text Kalyan Moy Gupta and David W Aha gt ITT Industries AES Division Alexandria VA 22303 Navy Center for Applied Research in Artificial Intelligence Naval Research Laboratory Code 5515 Washington DC 20375 surname aic nrl navy mil Abstract Taxonomic case based reasoning is a conversational case based reasoning methodology that employs feature subsumption taxonomies for incremental case retrieval Although this approach has several benefits over standard retrieval approaches methods for automatically acquiring these taxonomies from text documents do not exist which limits its widespread implementation To accelerate and simplify feature acquisition and case indexing we introduce FACIT a domain independent framework that combines deep natural language processing techniques and generative lexicons to semi automatically acquire case indexing taxonomies from text documents FACIT employs a novel method to generate a logical form representation of text and uses it to automatically extract and organize features In contrast to standard information extraction approaches FACIT s knowledge extraction approach should be more accurate and robust to syntactic variations in text sources due to its use of logical forms We detail FACIT and its implementation status 1 Introduction Case based reasoning CBR is a general methodology for retrieving and reusing past experience to
11. ganization generalization To organize features into subsumption taxonomies we propose the following procedure First it will perform a pair wise comparison of each feature to examine a potential subsumption relation i e a member or subset relation between two logical sentences Russell and Norvig 1995 These features are expressed in logical forms that include concepts and relations from the SO We assume that each feature is a complex sentence of literals connected by connectives as shown in Figure 5 Initially we will consider only conjunctive expressions The background knowledge to induce taxonomies is in the SO which uses three types of lexical relations to assess subsumption 1 s_a_type_of This standard relation is the basis for multiple inheritance in a SO and applies to both entities and events For example the entity PRINTING _INSTR is_a_type_of INSTRUMENT and the event TURN is_a_type_of MOVE 2 Constituent This includes a family of is_a_part_of relations that applies to entities such as INCLUDE WHOLE _PART and SET MEMBER 3 Is_a_subevent This includes hierarchical and temporal relationships between events For example two subevents of PRINT ACT are FEED PAPER ACT and MOVE PRINTHEAD ACT We will use additional background knowledge as needed to assess subsumption relations For example domain specific implication rules such as NOT EVENT gt PROBLEM EVENT implies that if an event does not occur then there
12. ink output Informal tests show its interpretation to be accurate and its speed to be fast for text of reasonable complexity such as for Navy Lessons Learned System documents 4 Feature extractor Not yet implemented 5 Feature organizer Not yet implemented 6 Case creation and indexing Our Document Extraction Workbench tool is useful for manually annotating the primary components of a case from documents It uses a handy drag and drop interface to provide XML markup of arbitrary complexity However we have not yet automated this process or implemented case indexing tools 5 Conclusion and Future Work Documents such as manuals logs and reports are a primary source of cases for CBR applications Case acquisition systems have predominantly focused on case indexing but have ignored the important task of feature acquisition In this paper we introduced FACIT a semi automated knowledge extraction framework that uses deep NLP techniques to perform feature acquisition and relational feature organization when the source documents are relatively unstructured and the ability to detect subtle nuances in language is crucial to extraction performance FACIT uses domain specific generative ontologies to create logical form representations for text documents of arbitrary complexity We showed how these could be used to semi automatically extract features and their relations without using a priori patterns In the future we will develop imp
13. is a problem with the event This permits the conclusion that for example the statement printing Problem subsumes the statement Data from computer is not printed To assess this subsumption relation our procedure will generalize the logical form of the statement Data from computer is not printed by reducing the conjuncts to NOT PRINTED HUMAN DATA PRINTING_INST and then applying the rule to obtain the further generalization PROBLEM PRINT_ EVENT which is a logical form for the statement Printing problem We will address where and how such background knowledge will be acquired and stored in our future research efforts After all potential subsumption relations are identified in a matrix directed graphs each representing a taxonomy shall be automatically constructed and presented to the domain expert for verification 6 Assigning indices to cases Indexing a taxonomic case involves assigning one or more leaves from distinct feature taxonomies Step 4 provides the logical form of features applicable to the cases Using the feature taxonomies as a reference FACIT will select only the most specific distinct features applicable to a case to encode it If a most specific feature in the case is not a leaf from one of the taxonomies then the case shall be brought to a domain expert s attention for review and correction This process of case indexing significantly differs from those that assign predefined features e
14. lement and evaluate components for each of FACIT s steps and complete our existing components We will evaluate the accuracy of generating logical forms by processing a variety of source text and investigate various machine learning techniques for extracting features from logical forms Finally we will formalize the algorithm for subsumption detection and assess the impact of implication rules for this task As FACIT is not yet fully implemented we cannot present evidence for our claims When completed it could best be compared with other domain independent approaches that semi automatically perform feature extraction organization and assignment from text documents However we are not aware of any other approach that addresses this complete set of problems Acknowledgements Thanks to Rosina Weber Karl Branting and our reviewers for excellent suggestions on an earlier version of this paper References Aamodt A amp Plaza E 1994 Case based reasoning Foundational issues methodological variations and system approaches AJ Communications 7 39 59 Aha D W Breslow L A amp Munoz Avila H 2001 Conversational CBR Applied Intelligence 14 1 9 32 Ashley K amp Lenz M Eds 1998 Textual case based reasoning Papers from the AAAI workshop Technical Report WS 98 12 Madison WI AAAI Press Baudin C amp Waterman S 1998 From text to cases Machine aided text categorization for capturing bu
15. national Conference on CBR pp 219 233 Vancouver BC Canada Springer Gupta K M amp Aha D W 2003 Nominal concept representation in sublanguage ontologies Proceedings of the Second International Workshop on Generative Approaches to the Lexicon Technical Report Geneva Switzerland University of Geneva School of Translation and Interpretation Gupta K M Aha D W amp Moore P 2004 Automatically organizing indexing taxonomies from acquired features Manuscript submitted for publication Gupta K M Aha D W amp Sandhu N 2002 Exploiting taxonomic and causal relations in conversational case retrieval Proceedings of the Sixth European Conference on Case Based Reasoning pp 133 147 Aberdeen Scotland Springer Knight K amp Luk S 1994 Building a large knowledge base for machine translation Proceedings of the American Association of Artificial Intelligence pp 773 778 Seattle WA AAAI Press Lenz M Hiibner A amp Kunze M 1998 Textual CBR In M Lenz B Bartsch Sp rl H D Burkhard amp S Wess Eds CBR technology From foundations to applications Berlin Springer Link 2003 The link parser application program interface API http www link cs cmu edu link api Minor M amp H bner A 1999 Semi automatic knowledge acquisition for textual CBR In R Feldman amp H Hirsh Eds Text mining Foundations techniques and applications Papers from the IJCAI 99 Worksh
16. oncepts represented using predicate argument relations are necessary for deriving logical forms Therefore LKBs that do not support such representations cannot be directly used 2 Resolve semantic ambiguity Semantic ambiguity results when multiple concepts are retrieved for a term Heuristics can be used to resolve these ambiguities For example OCCUR_LOCATED and NATIVE_OF are both retrieved for the term from p In this case heuristics select the concept NATIVE_OF because a larger proportion of its arguments are instantiated 3 Resolve syntactic ambiguity When multiple parses are semantically interpreted the instantiation and predicate argument binding differ among them FACIT selects the parse s that has the most predicate argument bindings as the valid one Therefore syntactic ambiguity resolution takes place during the semantic interpretation step We implemented a preliminary version of a semantic interpreter that operates with our SO and the output of the Link Parser 4 Extract features from the logical form Case features in a troubleshooting application are abnormal states and or observations pertaining to a piece of equipment For example statements such as Data from the computer is not printed Printout curls and Printout does not match the paper size are abnormal conditions in a printer troubleshooting domain whereas the statement Make sure the computer and the application are configured correctly
17. op Unpublished manuscript Pustejovsky J 1995 The generative lexicon Cambridge MA MIT Press Russell S amp Norvig P 1995 Artificial Intelligence A modern approach Englewood Cliffs NJ Prentice Hall Weber R Martins A amp Barcia R M 1998 On legal texts and cases In Ashley amp Lenz 1998 Yang C Orchard R Farley B amp Zaluski M 2003 Automated case base creation and management Proceedings of the Sixteenth International Conference on Industrial amp Engineering Applications of Artificial Intelligence and Expert Systems Loughborough UK Springer Zavrel J amp Daelemans W 1999 Recent advances in memory based part of speech tagging Technical Report 9903 Tilburg Netherlands Tilburg University Faculty of Arts Computational Linguistics and AI Induction of Linguistic Knowledge Group
18. oubleshooting manuals maintenance logs and failure modes analysis Developers select the case material translate it into language for end user consumption and encode it into a representation for use by the CBR system Unfortunately these processes are mostly manual are minimally supported by editors and require a significant amount of skill effort and time This complicates the wide spread application of CBR and is exacerbated by the use of increasingly sophisticated CBR methodological variants such as Taxonomic CBR Thus case index acquisition can be significantly accelerated by using software tools that assist with identifying extracting and transforming content from text sources to identify their indices Several researchers have developed methods for assigning indices to text documents as summarized in Table 1 but none appear to have developed index extraction methods for domain independent unstructured Table 1 Characterizing approaches for assigning indices to text documents Legend AV Attribute value CRN Case retrieval network DSF Developer supplies features DSL Developer supplies lexicon DSS Developer supplies similarity information ESV Expert supplies values for attributes EW Expert validation FT Free text IEs Information entity pairs ProPs Propositional patterns SE Sense enumerative SST Semi structured text None FAQ Finder Burke et al 1997 Prudentia None Weber et al 1998 FAII
19. ption taxonomies Gupta 2001 Fortunately this feature acquisition and organization problem can be addressed by knowledge extraction techniques Cowie and Lehnert 1996 which aim to deduce knowledge artifacts such as rules cases and domain models from text However knowledge extraction involves significantly more complex natural language processing NLP methods than do traditional information extraction IE approaches In this paper we introduce knowledge extraction methodologies in the form of a domain independent framework for feature acquisition and case indexing from text FACIT FACIT uses deep NLP techniques with a generative lexicon for semantic interpretation to enable robust interpretation of previously unseen text documents Gupta and Aha 2003 In a forthcoming paper we present evidence that standard IE techniques perform poorly in comparison to FACIT s knowledge extraction techniques on this feature organization task Gupta et al 2004 We next describe related work on acquiring case indices from text We then introduce FACIT illustrating its processes with an example Finally we report on FACIT s implementation status and discuss future research ideas 2 Methods for Indexing Text Cases Several engineering processes for acquiring high quality cases exist e g Gupta 1997 Developers typically consult documented knowledge and subject matter experts e g for an equipment diagnosis application they may rely on tr
20. s i 3 Semantically Interpret the Parsed Text Ea gt e Morphologically analyze tagged text Lexicon Identify senses Disambiguate senses Source Logical Form 4 Extract features y gt 5 Organize generalize features 2 Syntactically Parse the Source Text Tag text with part of speech Analyze sentence structure y Feature Taxonomies y gt 6 Assign case indices y Taxonomic Cases Figure 1 The FACIT framework processes and steps explicitly listing all potential senses of a term Instead a small set of powerful operators generates them on demand from their context of use GL supports strong compositionality and can derive senses of previously unseen term combinations The effort required to update GLs is sublinear and comparatively marginal We developed several extensions to GL theory and implemented these in a representation called Sublanguage Ontology SO Gupta and Aha 2003 We also developed software tools that support the development and maintenance of SOs including the Sublanguage Ontology Editor which allows users to edit new and existing concepts and the Concept Discovery Workbench CDW which supports the semi automatic acquisition of concepts from text documents This greatly simplifies and accelerates ontology updating Using the CDW to discover terms from the Multipass C
21. siness reengineering cases In Ashley amp Lenz 1998 Brill E 1995 Transformation based tagger V1 14 http www cs jhu edu brill RBT1_14 tar Z Briininghaus S amp Ashley K D 1999 Bootstrapping case base development with annotated case summaries Proceedings of the Third International Conference on Case Based Reasoning pp 59 73 Seeon Germany Springer Briininghaus S amp Ashley K D 2001 The role of information extraction for textual CBR Proceedings of the Fourth International Conference on Case Based Reasoning pp 74 89 Vancouver BC Canada Springer Burke R D Hammond K J Kulyukin V Lytinen S L Tomuro N amp Schoenberg S 1997 Question answering from frequently asked questions files Experiences with the FAQ Finder system AI Magazine 18 1 57 66 Cowie J amp Lehnert W 1996 Information extraction Communications of the ACM 39 1 80 91 CSSI 1998 Canon Multipass C 5500 User Manual Canon Computer Systems Inc Felbaum C Ed 1998 WordNet An electronic lexical database Cambridge MA MIT Press Gupta K M 1997 Case base engineering for large scale industrial applications In B R Gaines amp R Uthurusamy Eds Artificial Intelligence in Knowledge Management Papers from the AAAI Spring Symposium Technical Report SS 97 01 Stanford CA AAAI Press Gupta K M 2001 Taxonomic conversational case based reasoning Proceedings of the Fourth Inter
22. st applicable sense among them Domain independent LKBs e g WordNet Felbaum 1998 Sensus Knight and Luk 1994 have poor coverage for domain specific text applications For example WordNet covers only 25 6 of the terms from our naval training exercises domain Gupta et al 2002 Its coverage of senses is likely to be even lower because it lacks domain specific senses for known terms Consequently selected lexical resources must include domain specific terms and senses Thus issues of concern include the lexicon choice and the effort required to update it Semantic lexicons can be categorized as either sense enumerative e g WordNet Sensus or generative Pustejovsky 1995 Enumerative lexicons which require listing every sense of a term or phrase in the lexicon have weak lexical semantics few impoverished relation types between concepts weak compositionality cannot derive the meaning of an unlisted phrase from its constituent terms and large sense ambiguities Thus the effort to update such lexicons increases linearly with the number of unknown terms and phrases In contrast generative lexicons GLs include rich well principled semantics can express an unlimited set of relations and do not require Lexical Resource Taxonomic Case Acquisition from Text Development at Source Text e g Troubleshooting Manual l Fault Reports 1 Update the Semantic Lexicon Domain specific concept
23. ut to a much lesser degree we assume a domain expert will provide feedback only on whether sampled sentences contain features of interest In summary FACIT is the first index acquisition methodology to use a generative semantic lexicon and a logical form representation for extracted features This permits it to identify feature relations and thus generate feature subsumption taxonomies 3 Acquisition and Indexing Framework FACIT updates a semantic lexicon and uses it for syntactic and deep semantic interpretation to create a complete and valid logical form representation which is a set of Templates Induced classifier Induced Rules and ProPs Templates DSL WordNet Synset Sense Tagged Bag of Words DSF DSL DSF DSS DSL DSF ESV None Dictionary Terms Bag of Words AV Factors sentences represented in a predicate argument structure FACIT extracts features from the logical form to index cases We next describe and illustrate FACIT s six steps see Figure 1 by processing example sentences from the troubleshooting chapter of Canon s Multipass C 5500 printer user manual CSSI 1998 Steps 2 4 implement a knowledge extraction process DSL DSF DSL DSF ESV DSL DSF ESV DSL DSF EV DSL EV 1 Update the semantic lexicon NLP systems employ lexical knowledge bases LKBs to look up potential senses concepts associated with a term and use a disambiguation technique to select the mo
Download Pdf Manuals
Related Search