Ken Samuel

Dialogue Act Tagging with Transformation-Based Learning

For the task of recognizing dialogue acts, we are applying the Transformation-Based Learning (TBL) machine learning algorithm. To circumvent a sparse data problem, we extract values of well-motivated features of utterances, such as speaker direction, punctuation marks, and a new feature, called dialogue act cues, which we find to be more effective than cue phrases and word n-grams in practice. We present strategies for constructing a set of dialogue act cues automatically by minimizing the entropy of the distribution of dialogue acts in a training corpus, filtering out irrelevant dialogue act cues, and clustering semantically-related words. In addition, to address limitations of TBL, we introduce a Monte Carlo strategy for training efficiently and a committee method for computing confidence measures. These ideas are combined in our working implementation, which labels held-out data as accurately as any other reported system for the dialogue act tagging task.

Translating owl and semantic web rules into prolog: Moving toward description logic programs

We are researching the interaction between the rule and the ontology layers of the Semantic Web, by comparing two options: 1) using OWL and its rule extension SWRL to develop an integrated ontology/rule language, and 2) layering rules on top of an ontology with RuleML and OWL. Toward this end, we are developing the SWORIER system, which enables efficient automated reasoning on ontologies and rules, by translating all of them into Prolog and adding a set of general rules that properly capture the semantics of OWL. We have also enabled the user to make dynamic changes on the fly, at run time. This work addresses several of the concerns expressed in previous work, such as negation, complementary classes, disjunctive heads, and cardinality, and it discusses alternative approaches for dealing with inconsistencies in the knowledge base. In addition, for efficiency, we implemented techniques called extensionalization, avoiding reanalysis, and code minimization.

Integration Workbench: Integrating Schema Integration Tools

A key aspect of any data integration endeavor is establishing a transformation that translates instances of one or more source schemata into instances of a target schema. This schema integration task must be tackled regardless of the integration architecture or mapping formalism. In this paper we provide a task model for schema integration. We use this breakdown to motivate a workbench for schema integration in which multiple tools share a common knowledge repository. In particular, the workbench facilitates the interoperation of research prototypes for schema matching (which automatically identify likely semantic correspondences) with commercial schema mapping tools (which help produce instance-level transformations). Currently, each of these tools provides its own ad hoc representation of schemata and mappings; combining these tools requires aligning these representations. The workbench provides a common representation so that these tools can more rapidly be combined.

Protein Name Tagging Guidelines: Lessons Learned

Interest in information extraction from the biomedical literature is motivated by the need to speed up the creation of structured databases representing the latest scientific knowledge about specific objects, such as proteins and genes. This paper addresses the issue of a lack of standard definition of the problem of protein name tagging. We describe the lessons learned in developing a set of guidelines and present the first set of inter-coder results, viewed as an upper bound on system performance. Problems coders face include: (a) the ambiguity of names that can refer to either genes or proteins; (b) the difficulty of getting the exact extents of long protein names; and (c) the complexity of the guidelines. These problems have been addressed in two ways: (a) defining the tagging targets as protein named entities used in the literature to describe proteins or protein-associated or -related objects, such as domains, pathways, expression or genes, and (b) using two types of tags, protein tags and long-form tags, with the latter being used to optionally extend the boundaries of the protein tag when the name boundary is difficult to determine. Inter-coder consistency across three annotators on protein tags on 300 MEDLINE abstracts is 0.868 F-measure. The guidelines and annotated datasets, along with automatic tools, are available for research use.

Applying Semantic Rules to Achieve Dynamic Service Oriented Architectures

As the complexity and tempo of world events increase, command and control (C2) systems must move to a new paradigm that supports the ability to dynamically modify system behavior in complex, changing environments. Historically, the behavior of Department of Defense (DoD) C2 systems has been embedded in executable code, providing static functionality that is difficult to change. We propose the use of semantic models to represent system behaviors abstracted from procedural code, and we demonstrate that this provides a well-defined foundation for dynamic service oriented architectures. This paper describes an implementation that models a military convoy traveling through an unsecured area under changing conditions. The W3C standard Web Ontology Language (OWL) was used to describe the battlespace domain, and the proposed W3C Semantic Web Rule Language (SWRL) was used to capture recommended operating procedures for convoys in theater. In our experiment, two sets of rules were used: one set models rules of engagement for favorable visibility conditions on the battlefield, and the other models rules of engagement for poor visibility conditions. Ontologies and rule sets were translated into integrated knowledge bases that could be consulted as a service to derive alerts and recommendations for the convoy commander. Messages injected over an enterprise service bus (ESB) provide the changing conditions that affect the battlespace. We then were able to show that a dynamic event, such as an unexpected sandstorm, causes the appropriate set of rules of engagement grounded in the ontologies to be applied to the service to guide the convoy to safety. This paper describes the overall approach and the challenges we encountered. We outline the architectural options for constructing dynamic services and describe the semantic-based approach selected. We conclude with our findings and recommendations, including a set of requirements for a standard rule language needed to support agile services

Randomized rule selection in transformation-based learning: a comparative study

Transformation-Based Learning (TBL) is a relatively new machine learning method that has achieved notable success on language problems. This paper presents a variant of TBL, called Randomized TBL, that overcomes the training time problems of standard TBL without sacrificing accuracy. It includes a set of experiments on part-of-speech tagging in which the size of the corpus and template set are varied. The results show that Randomized TBL can address problems that are intractable in terms of training time for standard TBL. In addition, for language problems such as dialogue act tagging where the most effective features have not been identified through linguistic studies, Randomized TBL allows the researcher to experiment with a large set of templates capturing many potentially useful features and feature interactions.

Ontologies and Rules for Rapid Enterprise Integration and Event Aggregation

Ontologies enable explicit expression of collective concepts and support Machine-to-Machine (M2M) interactions at the semantic level. Ontologies expressed in a standard language, such as the Web Ontology Language (OWL) and exposed on a network offer the potential for unprecedented interoperability solutions since they are semantically rich, computer interpretable and inherently extensible. Rules that operate over ontologies allow for fusion of events in a semantically rich way. In this paper, we describe how we applied ontologies and rules for rapid enterprise integration of heterogeneous data sources and aggregation of events in the battlefield. We found that once a robust foundational domain ontology is established, it is easy and quick to integrate new data sources and therefore rapidly provide new system capabilities. We also found that rules that operate over ontological concepts are useful in agile aggregation of events for enhanced situational awareness.

Dynamic service oriented architectures through semantic technology

The behavior of Department of Defense (DoD) Command and Control (C2) services is typically embedded in executable code, providing static functionality that is difficult to change. As the complexity and tempo of world events increase, C2 systems must move to a new paradigm that supports the ability to dynamically modify service behavior in complex, changing environments. Separation of service behavior from executable code provides the foundation for dynamic system behavior and agile response to real-time events. In this paper we show how semantic rule technology can be applied to express service behavior in data, thus enabling a dynamic service oriented architecture.