The 4th COE Postdoctoral and Doctoral Researchers
Technical Presentation

Program (20 mins each: 15 mins presentation and 5 mins discussion)

1. "Comparison of Multiple Model Based Controller Coordination"
   Yoshiyasu Sakakura ( Systems & Control Lab. : D2 )
   坂倉　義康 (システム制御・管理講座 : D2 )
   

   [Abstract]
   In this presentation, I present a controller coordination method based on multiple local models for nonlinear chemical processes. In the proposed method, a bank of local controllers is designed based on multiple models and run in parallel. Manipulated variables are calculated by the affine combination of the weighted outputs of local controllers. The soft-max function with a tuning parameter is used to determine the weight coefficients of local controller outputs. The method is simulated to control the level control of tank and compared with Dougherty's method. The results of the proposed method indicated better control capability than the comparative method.
   
   
   Top

2. "Morphological Analysis with Robust Unknown Word Processing"
   Ai Azuma ( Computational Linguistics Lab. : D2 )
   東　藍 ( 自然言語処理学講座 : D2 )
   

   [Abstract]
   Morphological analysis is important elemental technology and has very wide application. In the pase, many researches put a great deal of effort towards morphological analysis, and achieve success in analysis of "formal" texts, i.e. texts in newspapers. On the other hand, there is an enormous amount of "informal" language data on the Internet, including e-mails, BBSs and blogs. Methods used in the past are no longer so effective to analyze these data. Detecting previously unseen words (so-called unknown words) are particularly difficult, especially in Japanese or Chinese because there are no spaces between words. In this presentation, I propose a morphological analysis framework based on machine learning, designed for powerful and robust unknown word processing. I use more expressive features which are useful to capture unknown words. I report the current progress of my research and discuss future directions.
   
   
   Top

3. "Recognizing paraphrase and entailment relationships between complex sentences based on logic relations"
   Shuya Abe ( Computational Linguistics Lab. : D2 )
   阿部　修也 ( 自然言語処理学講座 : D2 )
   

   [Abstract]
   Consider sentences 1 and 2:
   1. Put a marbles away, before a child swallows them.
   2. Put a marbles away, so that a child doesn't swallow them.
   Although, it is clear that there is a paraphrase relation between Sentence 1 and 2, they represent different rhetorical relations. Sentence 1 is a temporal rhetorical relation, whereas Sentence 2 is a rhetorical relation expressing an objective. As these sentences show, two sentences can be in a paraphrase relationship even when they represent different rhetorical relations.
   We will develop a model to recognize paraphrases between complex sentences, even those representing different rhetorical relations.
   
   
   Top

---

----- 10 minutes break -----

---




4. "Evaluation of a Spoken Dialogue Interface Based on Sound Field Control and Source Separation"
   Shigeki Miyabe ( Acoustics and Speech Processing Lab. : D2 )
   宮部　滋樹 ( 音情報処理学講座 : D2 )
   

   [Abstract]
   We are researching a novel sound interface of man-machine communication, which realizes high-quality sound reproduction and noise-free sound capturing simultaneously. Our proposed method is a combination of sound field reproduction technique and source separation. Since the sound field reproduction can reproduce any expected signal accurately at control points, virtual reality of sound can be presented to a user. In addition, by representing silent signal at the positions of microphones, the response sound is prevented to be recorded by the microphones. After the observed signals of microphones are processed to cancel the environmental noise, an output signal includes only the user's speech and its speech recognition performance is improved. In this presentation we evaluate the speech recognition performance of the proposed system in a real environment, and show that the performance is superior to that of the conventional method in an ideal condition.
   
   
   Top

5. "The Speech Guidance System "Kita-chan": Portability Challenges in Automatic Speech Recognition"
   Tobias Cincarek ( Acoustics and Speech Processing Lab. : D2 )
   Tobias Cincarek ( 音情報処理学講座 : D2 )
   

   [Abstract]
   With the emergence and spread of Internet technologies such as the World Wide Web, a still growing large pool of text, audio and video data accrued. Accessing certain pieces of information available from this data pool has become possible to a certain degree via text-based search engines. However, especially in the ubiquitous world, where everybody would like to have access to any data from everywhere (e.g. via cellular phone or via a public terminal) in many different situations it would be far more convenient to use speech (instead of keyboard and/or mouse) as a more natural user interface.
   For example, at the Acoustics and Speech Processing Lab of NAIST, the speech-oriented multimodal guidance system "Takemaru-kun" has been developed and been in operation at the Ikoma City North Community Center since November 2002. The system provides information on the weather, news, the facility, the surrounding environment, tourist information, a.s.o. Recently, a second speech-oriented system called "Kita-chan" has been installed at the newly opened subway station "Gakken-Kita-Ikoma", providing additional services such as information on the location of surrounding places, facilities, temples, parks, and the possibility of a speech-oriented search of train connections.
   Reduction of development costs of the Kitachan system being similar to Takemaru is an important issue. The talk is about portability challenges regarding the acoustic model of the speech recognition component of the Kitachan system. Speech data collected with "Takemaru-kun" over two years can be employed effectively for building the acoustic model of Kitachan. It is shown experimentally, that the data collected during the first month since "Kita-chan" is in operation is not enough to build an acoustic model which can outperform the Takemaru system's model.
   
   
   Top

6. "Hose bandwidth allocation method using feedback control and class-based queueing for VPNs"
   Masayoshi Shimamura ( Internet Engineering Lab. : D2 )
   嶋村 昌義 ( インターネット工学講座 : D2 )
   

   [Abstract]
   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.
   
   
   Top

NAIST Graduate School of Information Science