A data representation code for text processing systems

We demonstrate that digital signal processing of biomolecular sequences provides powerful approaches for solving highly relevant problems in bioinformatics by properly mapping character strings into numerical sequences. As examples, we show that color spectrograms visually provide, in the form of local texture, significant information about biomolecular sequences, thus facilitating understanding of local nature, structure and function; we provide an optimization procedure predicting protein-coding regions in DNA sequences including reading frame and coding direction, using both the magnitude and the phase of properly defined Fourier transforms; and we present a digital filtering approach to the process of translating nucleic acids into proteins. These approaches result in alternative mathematical formulations and often provide improved computational techniques for the solution of useful problems in genomic information science and technology.

Computerized text editing and processing with built-in indexing

A direct character-string retrieval method based on human search characteristics is proposed for Japanese textual image processing. In an experiment in which human subjects were to identify characters shown on a CRT screen, character-string search performance was better than single-character search performance in Japanese sentence contexts. This result suggests that character strings contain more information than characters alone. The analysis of reaction times shows that there are two stages in visual search. A character-string image retrieval system which, human-like, has two stages is effective in searching target images in ninety Japanese textual images. Moreover, human-like performance is obtained from the system: for example, it more easily identifies character-string images than single-character images. >

This talk will describe the use of weighted transducers in speech and language processing. Weighted transducers are automata where each transition has an input label, an output label, and a weight. They have key applications in speech recognition and synthesis, machine translation, optical character recognition, pattern matching, string processing, machine learning, information extraction and retrieval among others. We give a brief history of the field, an overview of the theory and algorithms of weighted finite-state and pushdown transducers, a discussion of available software libraries, a description of applications to speech recognition, speech synthesis, machine translation and NLP, and a short overview of current research problems.

Our objective is to enhance the effectiveness and efficiency of ad hoc and routing retrieval for large scale textbases. Effective retrieval means ranking relevant answer documents of a user's information need high on the output list. Our text processing and retrieval system PIRCS (Probabilisitc Indexing and Retrieval -Components- System) handles English text in a domain independent fashion, and is based on a probabilistic model but extended with the concept of document components as discussed in our last year's site report. Our focus for enhancing effectiveness remains on three areas: 1) improvements on document representation; 2) combination of retrieval algorithms; and 3) network implementation with learning capabilities. Using representation with more restricted contexts such as phrases or subdocument units help to decrease ambiguity in both retrieval and learning. Combining evidences from different retrieval algorithms is known to improve results. Viewing retrieval in a network helps to implement query-focused and document-focused retrieval and feedback, as well as query expansion.

Icon is a high-level programming language designed for nonnumerical applications with an emphasis on string processing. A goal-directed evaluation mechanism, combined with string analysis operations that are capable of producing alternative values, allows string processing to be integrated with the rest of the features of the language.

Processing Dream Texts

The minimum requirements in building a multilingual corpus of parallel texts are tags in the texts that establish cross-references between corresponding parts. These tags can be used to merge several independent text files into one, displaying aligned units of text, in order to do research with the help of the TUSTEP word processor. Extensive use is made of the complex possibilities of pattern matching: words, word forms, prefixes, suffixes, and one or more explicit or abstract character strings can be searched, while at the same time excluding others, by entering instructions into the word processor interactively. By doing so, the user does not depend upon predefined tags which contain semantic or grammatical information. Examples taken from the Finnish-German parallel corpus show how such instructions are written and what the results look like. Apart from interactive work, pattern matching and text processing tasks can be done by executing parameter driven program files. Examples show how four text files are aligned to produce a POSTSCRIPT as well as an HTML file.

At Indiana University in the late 1970's, the demand for word processing, the burden on the University's text processors, and a variety of commercial equipment demanded coordination. In 1978 a task force studied and made recommendations on the future of word processing throughout the corporation. As a result of these efforts, the Office of Information and Computer Services (OICS) assumed leadership and developed a small consulting team to assist users in choosing and installing word processing systems. By 180 the consultants' workload and the existence of perceived needs for $1.8 million in word processing products demanded a more systematic and encompassing strategy. Therefore, OICS formed a temporary group, the Office Systems Group, to devise objectives, policies, and implementation strategies for the 180's. This team needed to elarn what questions to ask and how to find answers. Furthermore, because modern office technologies will not work unless people accept them, the group had to deal with the social dynamics of office units. The group did a technical study of available products and developed a self-study method to determine office needs. Today the University's policies and strategy foster compatible systems, enable adequate maintenance, and promote staff mobility. The products chosen meet the needs of all users ᾢ teachers, students, clerical staff, professional staff, and administrators.

3 Knowledge from Text and Pictures: A Neuropsychological Perspective

: This report describes research done by the PROTEUS Project at New York University during the period January 15, 1985 to September 15, 1987. All of the activities described below were supported in part by the Strategic Computing Program of the Defense Advanced Research Projects Agency under Contract N00014-K-0163 from the Office of Naval Research. The PROTEUS Syntactic Analyzer is intended to provide an efficient, easy-to-use base for the various experiments in computational linguistics. The basic, long-term objective, as part of the Strategic Computing Program in Natural Language Processing, is to develop the technology necessary for the robust automated processing of messages containing natural language narrative. One aspect of the development of such language processing systems is the incorporation of detailed domain knowledge and the effective use of such knowledge in language analysis. The research has focussed on one type of message, CASREPs (equipment casualty reports), on developing detailed domain knowledge (a model of the equipment), and on using this knowledge for language understanding. Keywords: Text processing, Parallel parsing, Semantics.

Programming from metaphorisms

As a kind of important structured data, tree is widely used in domains like biology, molecular chemical, and text processing etc. However, many traditional machine learning methods cannot directly deal with tree-structured data. Currently, the commonly adopted approach is based on the subtree. It is supposed that the more common structure between two trees the more similarity between them. To effectively and efficiently learn tree-structured data, this paper proposes to learn tree-structured data using the echo state network based model space. The key idea is to learn models for the complex tree-structured data and convert the original tree data into points in the model space. The similarity between the tree-structured data is then measured by the distance of the models. Kernel method is combined to improve the discrimination performance of the classifier. Experiments are carried out on benchmark data sets to evaluate our proposed approach. Compared with existing methods, experimental results show that the proposed approach achieves better performance and wider application range.

BackgroundUse of routinely collected patient data for research and service planning is an explicit policy of the UK National Health Service and UK government. Much clinical information is recorded in...

The aim of this study was to build electronic algorithms using a combination of structured data and natural language processing (NLP) of text notes for potential safety surveillance of 9 postoperative complications. Postoperative complications from 6 medical centers in the Southeastern United States were obtained from the Veterans Affairs Surgical Quality Improvement Program (VASQIP) registry. Development and test datasets were constructed using stratification by facility and date of procedure for patients with and without complications. Algorithms were developed from VASQIP outcome definitions using NLP-coded concepts, regular expressions, and structured data. The VASQIP nurse reviewer served as the reference standard for evaluating sensitivity and specificity. The algorithms were designed in the development and evaluated in the test dataset. Sensitivity and specificity in the test set were 85% and 92% for acute renal failure, 80% and 93% for sepsis, 56% and 94% for deep vein thrombosis, 80% and 97% for pulmonary embolism, 88% and 89% for acute myocardial infarction, 88% and 92% for cardiac arrest, 80% and 90% for pneumonia, 95% and 80% for urinary tract infection, and 77% and 63% for wound infection, respectively. A third of the complications occurred outside of the hospital setting. Computer algorithms on data extracted from the electronic health record produced respectable sensitivity and specificity across a large sample of patients seen in 6 different medical centers. This study demonstrates the utility of combining NLP with structured data for mining the information contained within the electronic health record.