Brad Carter Don Trotter ERC Intelligent Systems Research Needs A Diverse Collection of Expertise and Experience An Intelligent Tutoring System from the Mid-80's Implementation: p The Goal Identify and provide remediation for a set of math topics at the sixth-grade level for adults preparing for the GED. p The Cognitive Model Extracted from State of Mississippi requirements for elementary teachers. p Subjects At least third grade competence in reading Some were students of a teacher instructing for GED Some were prisoners taking GED classes. p Evaluation and Testing System created pretest "on the fly" (filling slots in templates) so questions were not repeated. Students were given remediational material on subconcepts which appeared not to be mastered System created posttest. p System "successful" Statistically significant value added to student competence What did we learn? p System did what it was designed to do p No educational revolution followed p A computerized book - with an attitude INTELLIGENT SYSTEMS RESEARCH GROUP by Lois C. Boggess lboggess@cs.msstate.edu Julian E. Boggess gboggess@cs.msstate.edu Susan M. Bridges bridges@cs.msstate.edu Bradley D. Carter carter@cs.msstate.edu R. Dwight Hare rdh1@ra.msstate.edu Julia E. Hodges hodges@cs.msstate.edu Joe Picone picone@isip.msstate.edu Department of Computer Science, College of Eng. Institute for Signal and Information Processing, College of Eng. Instructional Technology Laboratory, ERC Curriculum and Instruction, College of Edu. Mississippi State University ABSTRACT We are a group with both mutual and complementary interests and strengths, in cognition, language, large bodies of data, multiple modes of communication between computer and humans, machine learning and adaptable systems. We've built systems in which the computer is seen as an aid to the human, rather than as the primary actor. Typically our goal is to achieve best possible performance when time constraints are sub-optimal, data are imperfect or incomplete, and there are multiple plausible ways for a system to proceed at any point in execution. These interests are supported by a core competency in a number of related information processing technologies including speech recognition, signal and image processing, natural language processing, machine learning, and expert database systems. A Sampling of Mid-90's Research AIMS: Automated Indexing at Mississippi State p Unrestricted vocabulary p Domain dependent p Embeds human expertise p Partners with human document analyst p Tools to tune system to the way language is used within domain KUDZU: Knowledge Under Development from Zero Understanding p Capture all the information present p Open vocabulary p Bootstrap a knowledge base Initially only metaknowledge of the domain Grows by reading the text Characterization of Waste Assay Data p Synthesis of data from multiple sources · database · multiple sensors · process history p Determining confidence of characterization · consistency checking of data from multiple sources · confidence associated with data sources p Knowledge discovery · detecting patterns in data · learning classification rules based on patterns Common Themes p Learning the lingo: Understanding language using cues in the language itself p Information extraction: from human experts from large bodies of data p Interactive systems, human-centered interfaces, multiple modes of communication p Data mining p Machine learning - supervised/unsupervised p Classification p One of the themes of "soft AI" is p Genetic algorithms, neural networks: Data driven Good performance with good data Reasonable performance in the presence of incomplete or missing data, erroneous input p Instead of systems which define to the world the boundaries within which the world must fit, we choose to build · Systems in which the domain boundaries are fuzzy · Systems which do not impose limits in some important aspects · Human-centered systems Basic TECHNOLOGY: A PATTERN RECOGNITION PARADIGM BASED ON HIDDEN MARKOV MODELS Sparcstation 1000 · 8 processors · 500 Mbytes RAM · Dedicated Server Demo Machine: · Sparcstation 5 · ATM Interface · T1 Interface Compute Server: · Sparc 20/60 · 2 processors · 192 Mbytes RAM · 1 Gbyte local disk The ISIP Computer Network: isip.msstate.edu Outside World The JEIDA Japanese Common Speech Data Corpus def:+ Automatic Generation of N-Best Proper Noun Pronunciations What Differentiates ISIP Speech Research? p Public Domain Software p Extensive Web Archive p Object-Oriented Signal Processing Software p State-of-the-Art Performance Tasks p Close Industrial Ties p Next-Generation Statistical Models Based on Chaotic Systems Applicable to acoustic and language modeling Addresses a fundamental barrier in speech understanding Gene Boggess Lois Boggess Susan Bridges Julia Hodges Computer Science Joe Picone Elect. and Comp. Eng. Dwight Hare College of Education Computer Aided Instruction Is Multidisciplinary By Nature Brad Carter Director of Education, ERC · Software Engineering · Software Metrics · Instructional Technology Gene Boggess Computer Science · Cognitive Science · Neural Networks · Genetic Algorithms Susan Bridges Computer Science · Expert Systems · Explanation-Based Learning · Hybrid Systems ISIP's Focal Projects · An Integrated Services Transactions Processor That Supports Advanced Telecommunications Interfaces such as an Asynchronous Transfer Mode (ATM) Digital Communications Link Example: Telephone-Based Natural Language Query of Entertainment Archives Customer: "Give me all movies, uh, make that only the recent movies, directed by Martin Scorsese and starring Robert DeNiro, and oh, by the way, make that movies about gangsters only." Computer: We have three titles available (the titles of the movies are shown on the television screen with real-time video of promo clips from each movie below the title). Please select a movie. Customer: "That one with the three guys looks good, I'll take that one. I want it to start at 8:00 PM tomorrow." Computer: (The promo clip for the selected movie starts playing on the television.) The movie titled GoodFellas starring Robert DeNiro and directed by Martin Scorsese will be delivered for viewing on your television on Thursday, September 25 starting at 8:00 PM. Thank you for using ISIP's Entertainment Server. Good-bye. Local Central Office ATM (160 Mbps) · Voice · Video · Data (X Windows) Unix Multiprocessor (Sparcstation 2000): · 8 Processors · 512 Mbytes of memory · videotape jukebox Search Algorithms: Pattern Matching: Signal Model: Recognized Symbols: Language Model: Sparcstation 5 Fileserver · 32 Mbytes RAM · 2 Gbyte local disk · 30 Gbytes external disk · 1 Exabyte Stacker (50 Gbyte) · 2 ethernets (Class C subnet) B&W Printer · postscript · 600 dpi · 12 ppm X/Audio Server: · Sparc SLC · 1 processor · 32 Mbytes RAM · 100 Mbyte disk · TCT DAT-Links · Network Audio NCD X Terminals · 20" Color/Audio · 15" B&W Color Scanner · 300 dpi · Adobe Photoshop · 8 processor · 500 Mbytes RAM Sony DATs · 16-bit audio · Networked 28.8K Baud Dialup · LBW X Windows · 115K TTYs URL: http://www.isip.msstate.edu FTP: ftp://ftp.isip.msstate.edu . Number of speakers 150 speakers 75 male speakers 75 female speakers Number of items per speaker monosyllables 178 isolated words 35 4-digit sequences 323 items Number of repetitions per item 4 repetitions of each item Range of speaker age 20 yrs. to 60 yrs. Amount of data 120 hours Number of Digital Audio Tapes 76 (120-minute tapes) Total number of utterances 193,800 utterances Number of channels/mic. type 2 (dynamic and condenser mics.) Anticipated size of final corpus (16-bit 16 kHz samples @ 1.0 secs per utterance) 6.5 Gbytes (13 CD-ROMs uncompressed) {s: -voiced, ...} {m: +nasal,...} {(ay: +voiced,...), (ih: +voiced,...)} ... form: acc num: pl type: bodypart HUMAN SPEECH RECOGNITION PERFORMANCE BENCHMARKS ON ARPA SLT CSR CORPORA CSR'94: SPOKE 10 Evaluation Group Vocabulary Open Closed Average 2.1 (0.7) 1.0 (0.6) Committee 1.2 (0.6) 0.5 (0.6) CSR'95: HUB-3 Evaluation Group Vocabulary Open Closed Average 2.2 2.1 Committee 0.5 0.5 · Overall human performance is at least an order of magnitude better than machine performance Semi-Parser Language Model Tagged Text Natural Language Processing Request Generator Knowledge Extractor Filled Templates Netscape Requests Netscape Knowledge Extraction Flat Parsed Structures Speech Recognition Language Model $(X) & $(Y) & $(Z) & doctor(X) & patient(Y) & knees(Z) & part-of(Y,Z) & examined(X,Z) tense: past arg1: subj [type:caregiver] arg2: obj [type:bodypart] num: sing type: patient form: gen num: sing type: patient form: nom num: sing type: caregiver def:+ Time (secs) What can you do with all of this? Text Julia Hodges Computer Science · Knowledge Bases · Database "Mining" · Machine Learning Lois Boggess Computer Science · Natural Language · Very Large Corpora · Intelligent Tutoring Dwight Hare Curriculum and Instruction · Learning and Pedagogy · Classification · Educational Policy Joe Picone Computer Engineering · Speech Recognition · Statistical Modeling · Signal Processing Natural Language Understanding "Show me all the reports from the White House on Healthcare."