SemTag and Seeker: Bootstrapping the Semantic Web via Automated Semantic Annotation Presented by: Hussain Sattuwala Stephen Dill, Nadav Eiron, David Gibson, Daniel Gruhl, R. Guha, Anant Jhingran, Tapas Kanungo, Sridhar Rjagopalan, Andrew Tomkins, John A. Tomlin, Jason Y. Zien IBM Almaden Research Center http://www.almaden.ibm.com/webfountain/resources/semtag.pdf Outline Motivation Goal SemTag Seeker Architecture Phases TBD Results Methodology Design Architecture Environment Conclusion Related and Future work. Motivation Natural language processing is the most significant obstacle in building machine understandable web. To allow for the Semantic Web to become a reality we need: Web-services to maintain & provide metadata. Annotated documents (OWL, RDF, XML, ...). Annotations Current practice of annotation for knowledge identification , extraction & other applications is time consuming needs annotation by experts is complex Reduce burden of text annotation for Knowledge Management Goal To perform automated semantic tagging of large corpora. To introduce a new disambiguation algorithm to resolve ambiguities in a natural language corpus. To introduce the platform which different tagging applications can share. SemTag The goal is to automatic add semantic tags to the existing HTML body of the web. Example: “The Chicago Bulls announced that Michael Jordan will…” Will be: The <resource ref = http://tap.stanford.edu/Basketball Team_Bulls>Chicago Bulls</resource> announced yesterday that <resource ref = “http://tap.stanford.edu/ AthleteJordan_Michael”> Michael Jordan</resource> will...’’ SemTag Uses TAP KB TAP is a public broad, shallow knowledgebase. TAP contains lexical and taxonomical information about popular objects like music, movies, sports, etc. Problem: No write access to original document How do you annotate??? Uses the concept of Label Bureau from PICS (Platform for Internet Content Selection) HTTP server that can be queried for annotation information Separate store of semantic annotation information Example: Annotated Page SemTag Architecture Add to DB Disambiguate windows Tagging Retrieve documents Automatic Manual Tokenize Find Context Spotting determine distribution of terms Learning SemTag Phases 1. Spotting: Retrieve documents from Seeker. Tokenize documents. Find contexts (10 words + label + 10 words) that appears in TAP Taxonomy. 2. Learning: Scan the representative sample to determine distribution of terms at each internal node of the taxonomy. SemTag Phases, cont’d 3. Tagging Disambiguate windows (using TBD). Add to the database. Ambiguities types: Same label appears at multiple locations in TAP ontology. Some entities have labels that occur in context that have no representative in the taxonomy. Training Data: Automatic metadata Manual metadata TBD Methodology Each node has a set of labels. E.g.: cats, football, cars all contain the label Jaguar. Each label in the text is stored with a window of 20 words – the context A spot(l,c) is a label in a context. Each node has an associated similarity function mapping a context to a similarity Higher similarity more likely to contain a reference TBD - Similarity Generate 200k dimensional vector corresponding to context. TF-IDF scheme Each entry of the vector is the frequency of the term occurring at that node divide by corpus frequency of the term. IR Algorithm – Cosine Similarity Vector product of sparse spot vector and dense node vector TBD - Algorithm Some internal nodes very popular: Associate a measurement Mus of how accurate Sim is likely to be at a node. Also Mua, how ambiguous the node is overall (consistency of human judgment) TBD Algorithm: returns 1 or 0 to indicate whether a particular context c is on topic for a node v 82% accuracy on 434 million spots The TBD Algorithm SemTag Results Applied on 264 million pages Produced 550 million labels. Final set of 434 million spots with Accuracy 82%. SemTag Methodology 1. Lexicon generation: Approximately 90 million total words. 1.4 million unique words . Most frequent 200,000 words. 2. Similarity functions: Estimated distribution of terms corresponding to 192 most common TAP nodes to derive fu. SemTag Methodology, cont’d 3. Measurement values: Determined based on 750 relevant human judgments. 4. Full TBD Processing: Applied to 550m spots. 5. Evaluation: Compared TBD results with additional 378 human judgments. Seeker A platform used by SemTag and other increasingly sophisticated text analytics applications. Provides scalable, extensible knowledge extraction from erratic resources. Erratic resources??? Seeker Design Goals Composability Modularity Extensibilty Scalability Robustness Seeker Architecture SemTag Components Indexing Tokens Crawls WEB Storage & Communication Query Processing Annotators Miners Modular & Extensible Scalability & Robustness n/w level APIs Seeker Design To achieve modularity and extensibility SOA (service-oriented architecture) was used where communication among agents is done through a set of language-independent network-level APIs. To achieve scalability and robustness Infrastructure components. Infrastructure Components The Data Store The Indexer Central repository for all data storage. Communication medium. For indexing sequences of tokens. The Joiner Query processing component. Analysis Agents Annotators Performs some local processing on each web page and write back results to the store in form of an annotation. Miners Performs Intermediate processing Looks at the results of spots on many pages in order to disambiguate them. Observation Advantage Other application can obtain semantic annotation from web-available database. Use both human & computer judgments to solve ambiguous data in their TBD algorithm Disadvantage The system require a large amount of storage space to store data. Requires much larger and richer KB to build web scale ontology. Conclusion Automatic semantic tagging is essential to bootstrap the Semantic Web. It’s possible to achieve good accuracy with simple disambiguation approaches. Future Work Develop more approaches and algorithms to automated tagging. Make annotated data public and seeker as a public service. Related Work Systems built as a result of the Semantic Web are divided among two types: Create ontologies – semi automated Page annotation. Examples: Protégé, OntoAnnotate, Anntea, SHOE, … Some AI approaches were used, but, they need a lot of training. Principal tool:Wrapping Some used other NL understanding techniques, example ALPHA. Questions?