180 likes | 333 Views
A Flexible Workbench for Document Analysis and Text Mining. Jon Atle Gulla, Terje Brasethvik and Harald Kaada Norwegian University of Science and Technology Norway. Outline:. Why a linguistic workbench? How does it work? How to use it? How did we use it?. Index. Retrieve. Docs.
E N D
A Flexible Workbench for Document Analysis and Text Mining Jon Atle Gulla, Terje Brasethvik and Harald Kaada Norwegian University of Science and Technology Norway Outline: • Why a linguistic workbench? • How does it work? • How to use it? • How did we use it? Gulla, Brasethvik and Kaada
Index Retrieve Docs Modified Query Result page Building Search Engines • Need to handle syntactic and morphological variation in documents: • language identification, text categorization, stemming/lemmatization, stopwords • Want to modify query to improve search result • stemming/lemmatization, spell-checking, query reformulation with ontologies/dictionaries, grammatical analysis, phrasing, anti-phrasing [FAST search engine (www.alltheweb.com)] Gulla, Brasethvik and Kaada
Text Ontology Extracting Information From Text • Structuring knowledge from text • tagging, compounds, grammatical analysis, ontological interpretation, regular expressions, patter recognition Minimal recursion semantics representations Database [Deep Thought EU project] Gulla, Brasethvik and Kaada
Manual labor Statistical & linguistic analyses Domain doc. coll. Ontology Constructing Ontologies • Want to extract prominent concepts/relations from text • tagging, compounds, NP recognition, term frequencies, stopwords, language identification [Brasethvik & Gulla, DKE, 38/1, 2001] Gulla, Brasethvik and Kaada
A simple expandable workbench for planning and running sequences of linguistic/statistical text analysis techniques Common Challenges • How to combine linguistic/statistical techniques for document analysis? • Many combinations feasible • Not clear what to use under which circumstances • How to support the experimental use of techniques? • Make use of existing techniques • Add new ones • Parameterize techniques • Run techniques in different orders Gulla, Brasethvik and Kaada
parameters transform or add input text output text Workbench Concept • Each technique is a component: • parameters to govern behavior • dependencies with other components • Workbench • manages components as building blocks • users can define an analysis as a chain of building blocks • no programming involved as long as appropriate components are available on the network Gulla, Brasethvik and Kaada
Workbench Concept Job = input text collection + sequence of parameterized online components Library of components = components available on the network Result = XML representation of documents, all (temporary) results Gulla, Brasethvik and Kaada
Workbench Architecture • Components: • Each component a web service • Programmed in any language (Java, Perl, Python, C) • Add to or transform input text document(s) • Execution of jobs: • Workbench keeps track of techniques that are available and coordinates their execution • All communication with XML-RPC • All temporary files stored in DOXML format for later inspection Gulla, Brasethvik and Kaada
The Principle of Adding Information kliniske undersøkelser Tagging Lemmatization Phrase detection Gulla, Brasethvik and Kaada
How to Use Workbench? • Set up techniques as web services with XML-RPC interface on some networked computers • Tell the workbench where to find them • Define job: • Specify document(s) to run job on • Select components and set parameters • Decide order of components • Run job Gulla, Brasethvik and Kaada
Selecting a Component Gulla, Brasethvik and Kaada
Defining a Job Iver’s document analysis job consists of 5 techniques Gulla, Brasethvik and Kaada
How did we use it? • KITH: Norwegian Center of Medical Informatics • Editorial responsibility for creating and publishing ontologies for medical domains • Traditional approach: • Workshops with experts • Manual process • New approach • Generate concept/relation candidates for health school ontology based on KITH’s document collection on the topic • 2.79 MB collection of documents Gulla, Brasethvik and Kaada
The KITH Ontology Construction Job Gulla, Brasethvik and Kaada
Extracted Prominent Concepts Gulla, Brasethvik and Kaada
Extracted concept relationships Gulla, Brasethvik and Kaada
KITH Evaluation • KITH case • 10 components used to extract concept candidates from document collection • 99 of 111 concepts in KITH’s existing ontology found • New concepts detected • Considerable faster than traditional manual approach • Workbench results included in KITH’s experimental ontology-driven IR system: www.volven.no Gulla, Brasethvik and Kaada
Conclusions • Presented a light-weight and expandable workbench for document analysis and text mining • Easy to set up, easy to use • Limited functionality • Future work: • Add more components to library • Allow more advanced job structures (choices, iterations, etc.) Thank you! Gulla, Brasethvik and Kaada