NAIST Graduate School of Information Science
| Date: | Thursday, Jan. 26, 2006 |
|---|---|
| Time: | 13:30 - 16:20 |
| Place: | L3 Lecture Room |
| Language: | English (Oral Presentation), English/Japanese (Question) |
| Chairperson: | Osamu Tobe (Internet Architecture and Systems Lab. : D3) Ismail Arai (Internet Architecture and Systems Lab. : D2) |
Temporal information is an useful information for many NLP tasks, such as machine translations, information retrieval and Q-A system. In the extraction and recognization of temproal information in Chinese, syntactic analysis is necessary. Because the temporal relation between verb and temporal phrase are ambiguious. We are constructing a head-modify based temproal processor for Chinese. This framework includes extracting the temporal properties in articles and a head-modify based temporal recognizer. We will adopt a Chinese word dependency analyzer for reasoning and extracting the temporal relation. I had presented the research about the dependency analyzer in last technical presentation. In this presentation, I will introduce the framework of our Chinese temporal information processor and report the current progress of our research.
There is an enormous amount of language data, including news, e-mails, BBSs and blogs. In these data, language tends to be somewhat informal. To process these data, morphological analysers are important fundamental tools. Morphological analysers have to be able to robustly handle previously unseen words, so-called unknown words. This is by no means an easy problem to solve. In Japanese or Chinese, it is particularly difficult to detect unknown words because there are no spaces between words. I propose a morphological analyser based on Conditional Random Fields (CRF), couples with enumeration of all possible unknown words, of all possible lengths. In this system, correct word segmentations and parts of speech are determined using a probability distribution, which is trained on a large-scale language resource.
For a robust automatic speech recognition of a spoken dialogue system, we have proposed an interface to control the sound field at the microphones where the response sound of the system is not observed. In this presentation, to give the method a faculty to eliminate the redundant response sound following fluctuation of the room transfer function, we introduce a semi-blind source separation technique estimating independence of the observed signal and the known response sound. Experimental results revealed that the proposed method improves robustness against fluctuation of the room transfer function.
To build a high-performance acoustic model for speech recognition, a large amount of real speech data has to be collected and transcribed by humans, which is a very time-consuming and costly process. Since the performance of speech recognition depends on the user (speaker characteristics) and the environment (acoustic conditions), the acoustic model has to be tuned separately for each recognition task in order to realize speech recognition which works anytime, everywhere and for anybody (=ubiquitous). However, it is infeasable to collect and transcribe enough speech data for every new recognition task in practise. Therefore, we propose a method for acoustic model construction, which is able to reduce the costs of speech data preparation. By applying a recently developed method for utterance-based selective training of acoustic models, the speech data for training can be selected automatically either from existing speech data resources or from newly collected data in supervised (requiring a word level transcription) or unsupervised (not requiring a transcription) manner. We will explain the selection algorithm and show experimental results for unsupervised construction of one acoustic model each for adults and children for a speech-oriented public guidance system.
As processor speed continues to rise beyond the giga-hertz range, at-speed testing is becoming an absolute necessity for these processors. Software-based self test (SBST) of processors attracts attention as a testing strategy achieving at-speed testing. In this strategy, a processor enables self-test by running a sequence of instructions called a test program. Moreover, hardware overhead, any performance or excessive power are not required. We propose the efficient test generation for SBST of pipeline processors and evaluated the proposed method for typical modules of a pipeline version of DLX processor. The proposed method can achieve the high fault efficiency without hardware overhead, any performance and excessive power.
Many systems around us can be considered as hybrid dynamical systems, which has continuous and discrete variables. Control methods for hybrid dynamical systems have been proposed in the recent decade. Some of them have been applied to simple chemical processes and showed good performance capabilities. While hybrid control approaches can handle discrete event, application of them have been limited to linear hybrid systems. In this presentation, I'll show you an application example of a hybrid control approach for a nonlinear hybrid chemical process. Mixed Logical Dynamical System formulation is used to express the discrete events in the process. The simulation result shows a good control capability in the control.
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 drastically 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.
In the presentation, I would like to give you an example of the fusion of social networking and ubiquitous computing technologies, which have attracted a considerable amount of attentions in the last few years. I introduce SCACS, a Social Context-Aware Communication System that facilitates face-to-face communications between old-timers and newcomers in a research community, by visualizing social networks though using a wearable interface. SCACS provides users with information on coauthor relationships in order to help users understand conversational partners' research background and relationships among them. In contrast to the systems that exploit social interactions in real worlds to enhance experiences of social networking services in virtual worlds, SCACS collects information on social networks (e.g., coauthor relationships) from virtual spaces (that is, databases), and then visualizes them to facilitate face-to-face communications among people in physical environments through using a wearable interface.