NAIST Graduate School of Information Science
| Date: | Thursday, Nov. 24, 2005 |
|---|---|
| Time: | 13:30 - 16:40 |
| Place: | L1 Lecture Room |
| Language: | English (Oral Presentation), English/Japanese (Question) |
| Chairperson: | Hideki Shimada (Internet Architecture and Systems Lab.: PD) Ryuhei Tenmoku (Vision and Media Computing Lab.: D3) |
In order to support the design and evaluation of user interfaces in plant operations, a human model incorporating physical and mental workload estimation and attention resource allocation is proposed. Knowledge bases for variable information, alarm management, user interface information, and failure-symptom relation and procedures for monitoring and fault diagnosis are built in the model. We can analyze human operators' behavior to detect and identify a failure cause in an emergency based on this model. As a case study, a user interface (UI) system, including an overview panel and alarm settings for a boiler plant simulator, was evaluated and modified from the viewpoints of decreasing workloads and shortening the fault detection and identification (FDI) track in an emergency. The evaluation results showed that the UI system was improved by the proposed model.
We propose a machine learning-based approach to zero-anaphora resolution using structural information from a sentence. Our model detects the most likely candidate antecedent of a given zero-pronoun and determines the anaphoricity of zero-pronouns using the boosting algorithm, a method for classification of semi-structured text proposed by Kudo and Matsumoto (2004). By comparing our model with previous work on anaphora resolution, we will show the effectiveness of introducing structural information into the task of zero-anaphora resolution.
We have argued that a family of kernels on graph nodes rovides a new
perspective on traditional link analysis measures of relatedness and
importance.
In this presentation, we discuss an issue (computational complexity)
encounterded in the practical application of the kernel-based link
analysis developed in our previous work.
We propose two types of apploximation methods for graph kernels, and
compare their performances using citation network data.
Firstly we report experimental results on applying information extraction (IE) methodology to the task of summarizing clinical trial design information in focus on gCompared Treatmenth, gEndpointh and gPatient Populationh from clinical trial MEDLINE abstracts. From these results, we have come to see this problem as one that can be decomposed into a sentence classification subtask and an IE subtask. By classifying sentences from clinical trial abstracts and only performing IE on sentences that are most likely to contain relevant information, we hypothesize that the accuracy of information extracted from the abstracts can be increased. As preparation for testing this theory in the next stage, we conducted an experiment applying state-of-the-art sentence classification techniques to the clinical trial abstracts and evaluated its potential in the original task of the summarization of clinical trial design information.
We have so far considered the task of extracting triplets from a document collection assuming that opinions are often expressed in the form of attribute-value pairs. However, we have observed the cases which are not captured in the present framework. For example, the hierarchical relations between attributes are currently untreatable. To investigate how many opinions have such relations, we annotated documents taken from Japanese weblogs with relation tags. In this presentation, we explain how to build an opinion-tagged corpus and discuss future directions of this work.
Blind separation of mixed binaural sounds using Single-Input Multiple-Output (SIMO)-model-based Independent Component Analysis (SIMO-ICA) is now being studied by the authors. The SIMO-ICA consists of multiple ICA parts and a fidelity controller, and each ICA runs in parallel under the fidelity control of the entire separation system. In the SIMO-ICA scenario, unknown multiple source signals which are mixed through unknown acoustical transmission channels are detected at the microphones, and these signals can be separated, not into monaural source signals but into SIMO-model-based signals from the independent sources. In this paper, we introduce an audio tele-presendce system which can reduce the interference sound and reproduce the target sound using SIMO-ICA for rescue robots. To evaluate it, we carried out the sound decomposition experiments under the real acoustic conditions by multiple users. The objective experimental results reveal that the decomposition performance of the proposed method is superior to that conventional method, and we can confirm that the output signals maintain information about the directivity of each sound source through the objective and subjective evaluation experiments.
In previous work, we propose a design-for-testability (DFT) method based on the partially strong testability that guarantees complete fault efficiency (FE) for register transfer level (RTL) data paths separated the controllers. In this presentation, I will first show the problems in the test which separated a data path and a controller. Then, I will introduce the partially strong testability for the whole RTL circuits without separation of a data path and a controller to reduce the problems.
VLSI technology has been growing and the speed of VLSI circuits has increased in recent years. Delay testing is an important technology to verify that a given logic operates correctly at the desired clock speed. Path delay fault model is one of the delay fault models and can model the delay between two flip-flops. Generally there are many path delay faults which do not have bad influence for normal operation of the circuit. We should not test such path delay faults. However the number of path delay faults in the circuit is extremely large, hence it is not practical to extract the path delay faults with no influence for normal operation at gate level. Our proposed method can find these path delay faults within practical time at register transfer level.
A strong demand for quality of service (QoS) to subscribers of VPNs has emerged. Weighted proportional fair rate allocation (WPFRA) provides a QoS method for one VPN model, the customer-pipe model. However, constructing large-scale VPNs with the customer-pipe model is difficult because of an exponential increase in required bandwidth. In future VPNs, a new model, the hose model, will be useful to improve the scalability. In this research, we clarify the condition and the characteristic for the QoS mechanism of the hose model by the performance evaluation.