Edmund S. Yu

Text categorization for multiple users based on semantic features from a machine-readable dictionary

The text categorization module described here provides a front-end filtering function for the larger DR-LINK text retrieval system [Liddy and Myaeing 1993]. The model evaluates a large incoming stream of documents to determine which documents are sufficiently similar to a profile at the broad subject level to warrant more refined representation and matching. To accomplish this task, each substantive word in a text is first categorized using a feature set based on the semantic Subject Field Codes (SFCs) assigned to individual word senses in a machine-readable dictionary. When tested on 50 user profiles and 550 megabytes of documents, results indicate that the feature set that is the basis of the text categorization module and the algorithm that establishes the boundary of categories of potentially relevant documents accomplish their tasks with a high level of performance. This means that the category of potentially relevant documents for most profiles would contain at least 80% of all documents later determined to be relevant to the profile. The number of documents in this set would be uniquely determined by the systems category-boundary predictor, and this set is likely to contain less than 5% of the incoming stream of documents.

Development, implementation and testing of a discourse model for newspaper texts

Texts of a particular type evidence a discernible, predictable schema. These schemata can be delineated, and as such provide models of their respective text-types which are of use in automatically structuring texts. We have developed a Text Structurer module which recognizes text-level structure for use within a larger information retrieval system to delineate the discourse-level organization of each documents contents. This allows those document components which are more likely to contain the type of information suggested by the users query to be selected for higher weighting. We chose newspaper text as the first text type to implement. Several iterations of manually coding a randomly chosen sample of newspaper articles enabled us to develop a newspaper text model. This process suggested that our intellectual decomposing of texts relied on six types of linguistic information, which were incorporated into the Text Structurer module. Evaluation of the results of the module led to a revision of the underlying text model and of the Text Structurer itself.

Interpretation of proper nouns for information retrieval

Most of the unknown words in texts which degrade the performance of natural language processing systems are proper nouns. On the other hand, proper nouns are recognized as a crucial source of information for identifying a topic in a text, extracting contents from a text, or detecting relevant documents in information retrieval (Rau, 1991).

Evolving intelligent text-based agents

In this paper we describe our neuro-genetic approach to developing a multi-agent system (MAS) which forages as well as meta-searches for multi-media information in online information sources on the ever-changing World Wide Web. We present EVA, an intelligent agent system that supports 1) multiple Web agents working together concurrently and collaboratively to achieve their common goal, 2) the evolution of these Web agents and the user profiles to achieve a better filtering, classification, and categorization performance, and 3) longer-term adaptation by using our unique neuro-genetic algorithm. Individual Web agents use neural networks for local searching and learning. Genetic algorithms are used to facilitate the evolution of agents on a global scale. NLP technology allows users to write sophisticated queries, and allows the system to extract important information from the user queries and the retrieved documents. The new text categorization technology used by EVA, which is also based on the neuro-genetic algorithm, can learn to automatically categorize and classify Web pages with high accuracy, using as few terms as possible. Additionally, we have developed a technique for integrating meta-searching and Web-crawling to produce intelligent agents that can retrieve documents more efficiently, and a self-feedback or automatic relevance feedback mechanism to automatically train the Web agents, without human intervention. This algorithm, together with the neuro-genetic algorithm, has greatly enhanced the autonomy of the Web agents.

A sublanguage approach to natural language processing for an expert system

Abstract A sublanguage grammar approach with strong reliance on semantic word classes was used to develop an NLP component for processing the free-text comments on life insurance applications for evaluation by an underwriting expert system. The NLP component is implemented using a logic-programming formalism. Lexicon entries contain semantic word class tags such as BODYPART, SYMPTOM, or MEDICATION. Adjacency Rules and an Ambiguity Filter are used to interpret the input data using these semantic word classes. Across two experiments, the system produced appropriate readings for 96.8% of a sample of 2069 utterances, and the number of parses produced per utterance was 1.079 for the same sample. Misspellings caused the system its only serious problem. The sublanguage approach to processing text was shown to be very promising for expert systems and suggests itself as a useful paradigm for a range of other text-based systems which must deal with naturally occurring and frequently ungrammatical texts.

Evolving and messaging decision-making agents

In this paper we describe our neurogenetic approach to developing a multi- agent decision support system which assists users in gathering, merging, analyzing, and using information to assess risks and make recommendations in situations that may require tremendous amounts of time and attention of the users. In Phase I of this project, called the EMMA project, we demonstrated the feasibility of a set of solutions to various problems by building an intelligent agent application that makes recommendations in the credit assessment domain using a constrained, static, well- understood collection of training and testing data. More specifically, this application demonstrated: 1) The effectiveness of a hybrid learning scheme that uses neural networks for local learning by the autonomous domain agents, and a genetic algorithm for evolving the sets of features available to these agents, and the agents themselves. 2) The use of a welldefined agent communication language (IBMs Java- based Knowledge Query and Manipulation Language, or JKQML) to coordinate the training and fusing of multiple decision- making domain agents. 3) The effectiveness of a trainable decisionfusion agent for merging multiple decision- making domain agents results into coherent recommendations for the user. 4) The use of a constrained natural language interface for accepting directives from the user, and for conveying recommendations. Furthermore, the benchmark results show that our EMMA Phase I prototype is comparable to the first class machine learning algorithm in the domain of loan applications or credit- worthiness, as reflected in the published results. We have also shown that our neurogenetic learning algorithm has the potential to perform far better than others, while using just about one half of the input features.

Feature selection in text categorization using the Baldwin effect

Text categorization is the problem of automatically assigning predefined categories to natural language texts. A major difficulty of this problem stems from the high dimensionality of its feature space. Reducing the dimensionality, or selecting a good subset of features, without sacrificing accuracy, is of great importance for neural networks to be successfully applied to the area. In this paper, we propose a neuro-genetic approach to feature selection in text categorization. Candidate feature subsets are evaluated by using three-layer feedforward neural networks. The Baldwin effect concerns the tradeoffs between learning and evolution. It is used in our research to guide and improve the GA-based evolution of the feature subsets. Experimental results show that our neuro-genetic algorithm is able to perform as well as, if not better than, the best results of neural networks to date, while using fewer input features.

Controversial topic discovery on members of congress with Twitter

Abstract This paper addresses how Twitter can be used for identifying conflict between communities of users. We aggregate documents by topic and by community and perform sentiment analysis, which allows us to analyze the overall opinion of each community about each topic. We rank the topics with opposing views (negative for one community and positive for the other). For illustration of the proposed methodology we chose a problem whose results can be evaluated using news articles. We look at tweets for republican and democrat congress members for the 112th House of Representatives from September to December 2013 and demonstrate that our approach is successful by comparing against articles in the news media.