Search system users have to find information they need out by themselves, because most of existing search systems return a list of documents as search results. It takes a large effort to find it out from long-length documents. In addition, there is a possibility that users cannot find useful information nevertheless they spent long time on the documents. This information seeking process is much cost in information retrieval. On the other hand, users need not to find information they need out by themselves because XML element search systems return a list of elements which satisfy users' information needs. Therefore, the framework of an XML element search system can reduce the cost in information retrieval, which is the reason why XML element retrieval techniques are useful and worth working on.

There are two main streams for researches of XML element retrieval techniques, i.e., 1) attaining effective search for satisfying accurate information retrieval, and 2) attaining efficient search for fast query processing. In order to satisfy 1), we proposed a scoring method to identify informative XML elements and a reconstruction method of search results to identify the most appropriate granularity of XML elements as search results. Our experimental evaluations showed our proposed methods overwhelmed existing methods in search accuracy.

Then, we also try to handle document updates of an XML search system. This is because document updates need to be managed when we come to think of a practical use of search systems. If document updates are not handled in a search system, users cannot obtain appropriate search results, which reduces the usefulness of the search system. We propose to extend a function of incremental updates of indices to general XML element retrieval systems, with filters to reduce the update cost by eliminating unimportant elements and terms. Moreover, we apply a method for integrating path expression which estimates accurate global weights in term calculation. We confirmed the proposed methods update indices in short time without a drop in search accuracy.

As an output of these researches, we developed a practical XML element search system which achieves accurate search and fast query processing with satisfying immediate reflection of document updates. However, there still is a gap between a practical XML element search system and a practical Web search system. To fill the gap, we try to apapt (XML) element retrieval techniques to HTML documents.

The next aim of this study is to determine the location and intensity of the rainfall field from the attenuation of a mesh network. Because of the links' susceptibility to rain, it has been suggested in the past that a rainfall sensing system can be built using microwave networks. The information on rainfall attenuation can be used to detect and reconstruct the rainfall field without using other weather sensors. However, the resolution of the link attenuation data is not enough for detecting the sporadic rain. In this study, we propose an efficient use of compressed sensing algorithm for improving the resolution. The proposed method can identify intense rainfall rates like that of "guerilla rain", which are of interest because they can cause disasters like landslides and flash floods. The links' rainfall attenuation can be used as the input to a rain sensing system. This system can complement existing weather sensing systems and act as standalone rainfall sensor in areas where traditional weather sensors are not available.

Over the last decade, neural decoding technology based on machine learning has been developed and enabled us to extract fine information on visual experiences and motor commands from measured neural signals, which is becoming a powerful tool for revealing neural representations and BMI application. To extract information with high predictive performance, neural recording methods that provide high spatiotemporal resolution and signal stability are required. A promising candidate is electrocorticogram (ECoG), which measures population activity of neurons with electrodes placed on the surface of the brain.

Here, toward high-performance decoding with ECoG, we tested the utility of ECoG systems in animal and human studies, and improved techniques to extract information from ECoG data. In chapter 2, the signal stability of ECoG responses recorded via a newly developed high-density mesh electrode array was tested. Collaborators applied it to the visual cortex in rats and this thesis demonstrates above-chance, generalized decoding performance for simple visual stimulation, using six hours of continuous data. In chapter 3, by applying decoding analysis to simultaneously recorded ECoG, LFP, and MUA signals from the monkey IT cortex, extractable information on visually presented objects was compared. The resultant decoding performance with ECoG was high and comparable with LFP and MUA. In chapter 4, ECoG was used to investigate how face-selective regions and written word- selective regions are distributed on the human cortex, which is considered a challenging task with fMRI. Results reveal that there exist multiple, separate face- and written word-selective regions in the human cortex. In chapter 5, using ECoG responses from human patients when they viewed objects, efficient input signal features for decoding analysis were explored. Spectral powers, phases and temporal correlations of ECoG signals were used as input features, and the decoding performances were compared. Results show the performance using temporal correlations between ECoG electrodes is higher than using spectral powers and phases in individual electrodes.

Those results suggest that the combination of ECoG recordings and neural decoding techniques is a powerful approach for extracting neural information, and we can considerably improve a decoder’s predictive performance by using signal features that take into account fine temporal patterns in ECoG signals.

近年，コーパスを用いた日本語研究が盛んになり，国立国語研究所においては，日本語史研究のための通時コーパスを構築する準備が進められている．通時コーパスには，現代語のコーパスと同様の形態論情報を付与することが必要とされているが，従来は歴史的な日本語資料に十分な精度で形態素解析を施すことができず，形態論情報のアノテーションは困難であった．

このような中，本研究は，日本語通時コーパスのための形態論情報アノテーションを実現するために自然言語処理技術を応用して，次の貢献を行った．

1. 古文の形態素解析を実現するための言語資源として，新たに辞書と学習用のコーパスを整備し，統計的機械学習にもとづく形態素解析技術を用いて，中古和文と近代文語文について実用的な精度（見出し語認定のF値で0.96以上）が得られる形態素解析システムを実現した．
2. 上記の言語資源と通時コーパス自体の整備のために，辞書の見出し語とコーパスの出現形とを関連付けながら形態論情報の修正作業を行うことのできるデータベースシステム（国語研究所「形態論情報データベース」）を構築し，通時コーパス整備の基盤を整えた．
3. 通時コーパスに収録される多様なテキストに対して高い精度で形態論情報のアノテーションを行う方法を検討し，近世口語文，和漢混淆文，旧仮名遣いの口語文について，実際に形態論情報のアノテーションを行った．
4. 上記の形態素解析技術や形態論情報付きの通時コーパスを人文科学系の研究者に使いやすい形で提供するために，新たなツールの作成・既存のツールの適用を行った．

A study on Morphological Annotation for the Japanese Diachronic Corpus

Toshinobu OGISO (1161018)

Recently, corpus-based study of Japanese language has become popular, and a diachronic corpus of Japanese is being developed at the National Institute for Japanese Language and Linguistics (NINJAL) to study history of Japanese language. In order to construct a richly annotated diachronic corpus of Japanese, morphological analysis of historical Japanese text is required. However, morphological analysis of old Japanese texts with adequate accuracy was impossible by conventional means, and annotation of diachronic corpora with morphological information was difficult using existing technology.

Given this situation, this study applied natural language processing technology to carry out the annotation of morphological information for the diachronic corpus of Japanese, and made contributions as below.

1. Dictionaries and corpora of historical Japanese text were newly created as language resources to carry out the morphological analysis of historical Japanese. Using these resources and a morphological analyzer based on statistical machine learning, morphological analyses of historical Japanese texts in the literary style of the Meiji era and morphological analysis of literature of the Heian era were achieved with high accuracy (over 96% at lemmatization level).
2. For compilation and maintenance of the language resources mentioned above and the diachronic corpus itself, a database system (NINJAL morphological information database) was developed. The database system makes it possible to modify annotations in the diachronic corpus and in related dictionary entries, while maintaining consistency between the two.
3. For a variety of texts included in a diachronic corpus, methods for performing annotation of morphological information with high accuracy were studied, and actual annotation was conducted.
4. For researchers of Japanese language and scholars of humanities, some newly created tools and existing software were applied to the diachronic corpus in order to make it possible to use the morphological analysis system and the annotated diachronic corpora easily.

Japanese Predicate Argument Structure Analysis Based on Positional Relations between Predicates and Arguments(述語と項の位置関係に基づく日本語述語項構造解析)

述語項構造解析の目的は,述語とそれらの項を文の意味的な構成単位として, 文章から「誰が何をどうした」という意味的な関係を抽出することである.これ は,機械翻訳や自動要約などの自然言語処理の応用において重要なタスクの 1 つ である.

一般に,項は述語に近いところにあるという特性がある.そのため,従来の述 語項構造解析の研究では,候補を述語との位置関係でグループ分けし,あらかじ め求めておいたグループ間の優先順序に従って正解項を探索してきた.しかしな がら,その方法には異なるグループに属する候補同士の比較ができないという問 題がある.

そこで,異なるグループごとに最尤候補を選出し,それらの中から最終 的な出力を決めるモデルを提案する.このモデルは優先度の高いグループに属す る候補以外も参照することによって最終的な決定を行うことができ,全体的な最 適化が可能である.

実験では,提案手法は優先順序に従う解析よりも精度が向上することを確認し た.そして,述語項構造解析の精度を向上させるために必要な今後の課題につい て,述語の種類に応じて分析し議論する.