Enhancements to the DARPA Communicator Architecture Theban Stanley

Enhancements to the DARPA Communicator Architecture Theban Stanley, Julie Baca, Matt Elliott and Joseph Picone Human and Systems Engineering Center for Advanced Vehicular Systems Mississippi State University

Abstract Monolithic Systems Speech Recognition Natural Language Understanding Natural lan. . Distributed Systems Informati. . Information Retrieval Speech Rec. . Problem Statement: The distributed framework has enabled the development of client-server applications with complex inter-process communication which has resulted in the reduction of the overall system robustness to failure. Some of the notable projects: ØAdvanced Language Engineering Platform (ALEP) ØGeneral Architecture for TEXT Engineering (GATE) ØTIPSTER project ØDARPA Communicator program: An open source architecture for spoken language applications. Human Language Technology Page 1 of 20

Introduction Advantages of the DARPA architecture: • The Communicator has a flexible “hub and spoke” architecture with a programmable hub which allows flexible control of interaction among servers. • Open source architecture. • Reduced prototype development time. • Standard platform for evaluation of systems developed by different laboratories. Human Language Technology Page 2 of 20

HLT System Initial Prototype system: • Dialog System application ØAudio Server ØSpeech Recognizer ØNLP Parser ØDialog Manager ØDatabase Server • Speaker Verification application ØAudio Server ØSpeaker Verification Server Human Language Technology Page 3 of 20

Original Architecture - Disadvantages Deadlocks in the communication between servers: • The two main reasons for these deadlocks were inter-process communication error and misfiring of user interface events. • Need for more organized mechanism for logging all the communication between servers. • These issues were anticipated to grow in number and complexity as we moved to a multi user/application platform. Need for automated monitoring of servers: • When a process fails, the user has to manually track this development and restart the process. • Need for modules with the intelligence to start servers, check on their status and terminate these processes when the application is closed. Human Language Technology Page 4 of 20

Original Architecture - Disadvantages Inability to automatically handle multiple users: • DARPA communicator provided no mechanism to handle multiple users. • The servers needs to be manually started for each user. • In a multi user environment, port allocation needs to be handled with care. • Need for modules that can keep track of the client-server association and manage port allocation. Need for a common user interface for all the applications: • Supporting multiple applications required a common interface that allows the user to choose from a list of applications. • The module should be able to start the user interface for the requested application and coordinate the creation of the appropriate servers. Human Language Technology Page 5 of 20

Architectural Enhancements – Common Application Interface • The Demo Selector module: Ø Provides desired interface Ø Coordinates with the Process Manager module to start the required servers. • The user selects one of the application • The Demo Selector loads and displays the appropriate user interface (UI). • Directs the Process Manager to start the appropriate servers. Human Language Technology Page 6 of 20

Architectural Enhancements – Automated Server Management • Automated server management becomes critical when the system has to handle multiple users. • Process Manager module was designed to automatically manage and control all server processes in the prototype system architecture. • The Process Manager exploits the powerful functionality of Java’s Process Object. • Process Manager the capability to create a process, wait on a process, perform input/output on a process and also check the exit status of a process. Human Language Technology Page 7 of 20

Architectural Enhancements – Process Manager Client Side Server Side Process Manager Speech Analysis Signal Detector Speech Recognition Hub Speech Recognition Human Language Technology Hub Data Recorder Signal Detector Page 8 of 20

Architectural Enhancements – Robustness Improvements Improving system robustness was the primary focus of these enhancements. State Machine architecture: • Redesign of the servers as state machines. • The server anticipates a particular message. • Any inter-process communication error leading to a server failure or a deadlock can be trapped by this architecture. Speech Analysis Client Signal Detector Initialization state Audio Ready Audio_Ready state _ eady R _ o i Aud ck A Audio_Ready_Ack Human Language Technology Ready state Data_Transfer State Data_Ack State End_Points End_Of_Utterance Handshaking: Handshaking was implemented by using the implicit capabilities of the Communicator frame. Wait_for_Audio_ State End_Of_Utterance State Ack Page 9 of 20

Results and Analysis – Quantitative Analysis Pilot Corpus Experiment: • Pilot corpus – 403 spontaneous utterances collected from the user • Conducted to test against an established base line. • About 4% of the system failures were caused due to server errors and deadlocks. • The enhanced architecture significantly reduced the system failures • Restrictions: Ø One user Ø Conducted in text mode Human Language Technology Page 10 of 20

Results and Analysis – Quantitative Analysis General Usage Scenarios Experiment: Procedure: ØUsers: Five ØComposition: 3 Males and 2 Females ØScenarios: 24 ØPractice Session: 10 minutes ØExperiment: 1 hour 30 minutes The user was asked to cease testing if there was a system failure or he/she exceeded the allotted time of 30 minutes. Scenario sample: (Dialog system Application) Imagine you are in a big city to attend a conference. Once the conference proceedings are over for the day, you want to visit some sites of interest. You don’t have a map with you and have no idea about the layout of the city. Use the system to plan your trip. Human Language Technology Page 11 of 20

Results and Analysis – Quantitative Analysis Results: • 129 interactions passed the enhanced architecture. • Analysis of data showed that the failures where more frequent in Dialog system running in speech mode. • More experimentation is needed. Human Language Technology Page 12 of 20

Results and Analysis – Quantitative Analysis Dialog System – Speech mode Experiment: Procedure: ØUsers: Five ØComposition: 4 males and 1 female ØScenarios: 9 ØPractice Session: 10 minutes ØExperiment: 1 hour The user was directed to ask for assistance in restarting the system in case of a system failure. Scenario sample: Imagine you are working in a big city for quite a few years. You plan to make a visit to Starkville. So you start on a road trip from your city. You are almost near Starkville when you find that you are really low on gas. Use the system to make a decision on whether you can make it without filling gas. Human Language Technology Page 13 of 20

Results and Analysis – Quantitative Analysis Results: • The results have shown a 7. 2% robustness improvement to failure. • Further experimentation with: Ø At least 20 additional users who have no familiarity with the system. Ø The system should be allowed to respond to user queries continuously for prolonged time periods. Ø The prolonged experiments must be carefully controlled using scenarios that properly exercise system functionality, so that meaningful data are collected. Human Language Technology Page 14 of 20

Results and Analysis – Qualitative Analysis • Process Manager • Debug Window Human Language Technology Page 15 of 20

Conclusion • Vulnerabilities in the Communicator architecture have been addressed through several important enhancements. The key Contributions include: Ø Improved Robustness by incorporating state machine architecture and basic handshaking between servers. Ø Automated Server Management Ø Common Application Interface • Experiments have shown a 7. 2% improvement on the address querying task which is the most complex task in the HLT system. Human Language Technology Page 16 of 20

Future Work • Further Wizard of Oz experiments can be performed to improve the grammar and the language model which will allow the dialog system to handle a wider range of user queries. • A future enhancement would be to extend the dialog system to accommodate a statistical parser which can be trained on any data set. • The enhanced system can be tested on ERC supercomputer clusters which will significantly improve the resources available to the system. Human Language Technology Page 17 of 20

Questions Human Language Technology Page 18 of 20

Results and Analysis – Qualitative Analysis Client Side Server Side Process Manager Au dio Signal Detector _R ea d y Speech Analysis Audio_Ready_ack Initialization state Wait_Audio_Ready state Audio_Ready Hub Data Recorder Data Transfer state The Signal Detector errors out End_of_Utterance state Human Language Technology Page 19 of 20

Network Configuration window Human Language Technology Page 20 of 20