<p>We are developing a conversation system which efficiently transfers a massive amount of information like news articles by spoken dialogue. Here, "efficient" means that only the necessary information is transferred except unnecessary information for the user from target articles. In our system, feedbacks from the user are indispensable in order to realize high EoIT (Efficiency of Information Transfer). Therefore, we propose a utterance intention recognition method combining language information and prosodic information for the purpose of understanding diverse feedbacks from users. The feature of the proposed method is that it automatically extracts prosodic features with a high contribution ratio by using deep learning. We confirmed the effectiveness of the proposed method using a corpus with utterance intention tags designed based on dialogue data collected using our conversation system.</p>

<p>In a human-human conversation, a listener frequently conveys various intentions explicitly/implicitly to the speaker with reflexive short responses. The speaker recognizes the intentions of these feedbacks and changes the utterance plans to make the communication more smooth and efficient. These functions are expected to be useful for human-system conversation also, but when human face the system they do not always give the same feedback as they did with human. We investigated the feedback phenomena of human users against our system which is designed to transfer a massive amount of information like news articles by spoken dialogue. First, user's intentions of feedbacks that can affect the behavior of the system were classified from the viewpoint of roles in the progress of conversation: presence/absence of a user's interest, comprehension state, release/keep turn, and so on. Then, using our news delivery conversation system, we gathered user's short responses, and labeled the intention through listening test by the subjects. We also attempted automatic identification of those intentions. As a result, in the current system, it is found that there are many feedbacks to convey doubt and automatic identification of them is possible stably, but there are not so many other feedbacks.</p>

<p>想起は，ある概念が別の概念を思い浮かばせる程度と定義される．伝統的な語彙意味論に基づく意味関係を補完するものとして，多義解消や意味的類似度などの言語処理の応用への適用が想定されている．本研究では，想起を概念連鎖(連想)のショートカットであると仮定し，背景にある概念連鎖を意味ネットワーク上の経路として分析する．系列パターンマイニングの手法を適用することにより，想起の背景にある概念連鎖の類型を探る．</p>

<p>ニュース記事の伝達をタスクとして，聞き手が挟む相槌や聞き返しに応じて適宜情報を補完しながら会話を進める音声対話システムを提案する．要点を伝える主発話計画と，それに対し情報を補完する副発話計画を記事から自動生成する機能，これらの発話計画を聞き手の反応に暗に表れる理解状態に応じて切り替える機能を実装した．これらにより，リズムある対話による効率的な情報伝達が実現できた．</p>

This paper proposes a method to compute cross-lingual semantic textual similarities between sentences in different languages. The method is based on a machine-learning approach, which employs several monolingual similarity features chiefly proposed by projects participated in the *SEM STS (Semantic Textual Similarity) task. To utilize the monolingual features in a cross-lingual setting, we tested two methods to cross language borders: one utilizes bilingual dictionaries and the other employs off-the-shelf machine-translation engines. While our experimental results demonstrated such a translate then monolingual similarity method was reasonably effective in cross-lingual STS for English, Japanese, and Chinese, it has also been insisted that syntactic/structural differences have to be incorporated into the machine-learning based framework.

<p>文章中の語句に関連するWikipedia記事を対応付けるタスクはWikificationと呼ばれる．本発表は，日本語・英語の対訳文書に含まれる固有表現に対して並行的に対応付けを行う方式を提案する．提案方式は，他方の言語の情報を利用することにより，双方の言語における固有表現抽出処理の結果を補完し，言語横断的なWikificationを達成する．</p>

Wikipedia has attracted many researchers trying to extract structured information from its resources. In this research, we particularly draw attention on the multilingualism of Wikipedia and the semi-structured table-like information (Infobox) provided in each of the Wikipedia articles. More precisely, this research aims to establish a useful computational mechanism to extract multilingual attribute information from the Infobox templates and the actual Wikipedia articles in Japanese, Chinese, and English Wikipedias. The experimental results suggest that a clue, what we call attribute variables, plays a significant role in achieving high precision, and the overall performance can be further improved by incorporating attribute names and values.

This paper proposes a method to discover conceptual mates in the lexical-semantic resources of other languages. The proposed method integrates two types of cross-lingual semantic relatedness by balancing them using a weighted sum : One, definitional relatedness, estimates the probabilistic correspondence between sets of synonyms, while the other, explana-tory relatedness, measures the textual similarity between gloss texts. Our experimental results show that (1) the weighted-sum method was simple yet relatively robust; (2) the extended gloss texts gathered from the conceptual neighbors of objective concepts played a certain role in improving matching performances (mean average precision); and especially, (3) the nega-tive impact of machine translation, inevitably introduced to measure monolingual textual similarities, could be substantially reduced, and performances even slightly improved, thanks to the translation redundancy yielded by the use of multiple trans-lation engines.

This paper proposes a method for finding possibly corresponding synonym sets in a different lexical resource in a different language. By making use of Princeton Annotated Gloss Corpus as the source of target language statistics, the proposed method exhibited promising results in the experiments: More than 73% of the corresponding Princeton WordNet synsets were successfully recovered within the top-5 candidates, given a query set of Japanese WordNet synsets. As the proposed method minimally requires that the input to be seen as an apparently synonymous word set, the method could be extended and the performance would be further improved by incorporating richer information such as textual gloss overlaps and/or structural constraints posed by the lexical resources at hand.

In a Wab-based language service infrastructure, Web-servicized language resources can be combined and a new language resource is virtually realized as a composite Web service. Furthermore, if the infrastructure, as in the Language Grid, provides a collaborative environment, a new type of language resource that evolves through utilizations is feasible. This paper discusses possible architecture for a combined lexicon access service, which utilizes monolingual lexical semantic lexicons and bilingual dictionaries. This service could virtually realize a multilingual semantic lexicon by opportunistically associating semantically corresponding entries in the relevant lexical resources across languages. Given this scenario, this paper particularly proposes a framework for representing cross-lingual/interlingual lexical semantic correspondences that are expected to be incrementally recovered through a series of service invocations. This paper further discusses that recovered correspondences can be organized as a dynamic secondary language resource, while keeping the existing primary language resources intact.

We have been investigating characteristics of empathic states expressed in spontaneous utterances by analyzing speech data annotated with attitude and emotion categories. Basing on these investigations, this paper examines a method to estimate empathic states from speech utterances. The idea is to use the annotated data as training data for a machine learning algorithm (SVM) to construct relevant classifiers. These classifiers are arranged in such a way to reflect our definition of empathic states. The experimental results suggest: (1) there exist dependencies between attitude and emotion; (2) the estimation can be improved if augmented with the categories for the preceding utterance. This paper further discusses details of the initial experimental resutls and proposes a research agenda for the future study.

The purpose of this study was to examine how a transition of the conversation would occur among strangers, when they were to engage in multiple conversations and also to investigate how the transition would be associated with the atmosphere, and nonverbal behavior, such as the number of gesture, .head turning toward partner, the amount of utterance. In the experiment, 66 university students an unacquainted with one another participated as one of three-person group participated in dyadic and triadic conversations. Then, the influence of dyadic conversation on following triadic conversation was analyzed particularly in regard with the perceived activation level of atmosphere and nonverbal behavior.

The purpose of this study was to examine how a transition of the conversation would occur among strangers, when they were to engage in multiple conversations and also to investigate how the transition would be associated with the atmosphere, and nonverbal behavior, such as the number of gesture, .head turning toward partner, the amount of utterance. In the experiment, 66 university students an unacquainted with one another participated as one of three-person group participated in dyadic and triadic conversations. Then, the influence of dyadic conversation on following triadic conversation was analyzed particularly in regard with the perceived activation level of atmosphere and nonverbal behavior.

Aiming at computational estimation of participants' affective states in spontaneous dialogues, this paper analyzes the statistical relationships between affective states and speech features, such as F0 frequency, speech intensity, and speech speed. The results from statistical analyses applied to human-annotated speech data show that certain affective states, especially emotional states, are expressed and conveyed by particular speech features. The results further suggest that there exist possible gender differences in ways representing affective states.

The purpose of this study was to examine correlations among skill of building relationships, nonverbal behaviors (i.e. hand movements, interpersonal distance, head direction and the amount of utterance) and personality in a dyadic conversation. A total of 66 undergraduates (51 males and 15 females) participated in this experiment. The results were that skill of building relationships was positively correlated with ACT, self-monitoring and hand movements, and it was also negatively correlated with LSO-U. In addition, LSO-U was positively correlated with the amount of utterance.

From a Brunswikian lens model approach, this study examined relationships between nonverbal cues (i.e. hand movements, interpersonal distance, head direction and the amount of utterance) and skill of building relationships. A total of 66 undergraduates (51 males and 15 females) participated in this study. Participants rated their own skill of building relationships and were also rated by their conversational partners. Moreover, independent raters, who were not in this experiment, rated all participants on the skill. As a result, hand movements were perceived as a good indicator of the skill by those who actually exhibited as well as by their conversational partners and the independent raters.

This study experimentally investigated relationships between such nonverbal behaviors as degree of speech, hand gestures, head direction and body distance, and subjective evaluation of an activity of atmosphere created during a three-person conversation, in order to establish a technique for automatically estimating atmosphere in interpersonal communication. This study also investigated effect of information (topic) presentation by a display located in the conversation situation. Three participants who were the same generation and met for the first time talked for 8 minutes overall and then rated the activity of atmosphere of their own conversation presented by video every two minutes. Out results showed that the degree of speech and the hand gestures had a high correlation with subjective atmosphere while the head movement and the speech bias had a low correlation. Our results also revealed that the information presentation affected conversation and these correlations.

Language services provided by or registered to the Language Grid are classified into one of the predefined language service types, and can be employed through accordingly defined APIs. More than sixty actual Web services have been provided by the infrastructure, and the number is expected to further increase. In order to properly accommodate these services of various types, it is required to establish a principled way to organize them. Given this situation, this paper argues that a kind of ontological knowledge that backs up classification and definition of the services, including their APIs, has to be organized. In addition, this paper discusses role of the international standards associated with linguistic data and language resources. These standards should be highly appreciated not only for securing so-called interoperability, but also for characterizing some types of the language services. This paper further describes some experimental Web services whose results are represented in XML documents that are compliant with particular ISO standards : LAF/GrAF for syntactic dependency, and LMF for conceptual lexicon.

This paper discusses characteristics of empathic states expressed in spontaneous utterances. We basically define the empathic state of an utterance as a combination of attitude and emotional state that are expressed toward the preceding utterance to which the target utterance is directed. Among the possible combinations, we define &quot;empathy&quot; as a special empathic state, where the attitude and the emotional state are identical. Through a series of statistical analyses of speech data collected from three-party spontaneous dialogues, we show that some empathic states are identifiable by using a range of speech parameters. These results serve as solid foundation on which a computational mechanism for estimating a speaker&#039;s empathic states can be developed.

We propose an application independent Sinhala character input method called sri shell with a principled key assignment that is based on phonetic transcription of Sinhala characters. We argue that a good character input method should fulfill two criteria, which are efficiency and user-friendliness. We have introduced several quantification methods to quantify the efficiency and user-friendliness of Sinhala character input methods. Experimental results are shown to prove the efficiency and user-friendliness of the proposed method.

This paper proposes a method for generating Kanji reading alternatives for autonomous Kanji learning. The proposed method generates reading alternatives by applying rules based on pronunciation, shape, and semantic similarity of Kanjis. To reduce the number of alternatives properly, we applied an edit distance-based filter to the candidates generated through the pronunciation-based rules. This paper shows the preliminary evaluation results using the test questions formerly issued in the Japanese language proficiency tests, and discusses the effectiveness of the method.

To increase the accessibility and usability of online language services, this paper proposes the language grid to create composite language services for various communities. The language grid is called "horizontal," when the grid connects the standard languages of nations, or "vertical," when the grid combines the language services generated by communities. Semantic Web service technologies are applied in a human-centered fashion, to create composite language services through the collaboration of users and agents. Three example scenarios are given to illustrate how the language grid will organize standard and community language services for intercultural collaboration activities.

音声認識では認識辞書に含まれない単語は認識できないという、いわゆる未登録語の問題があるが、認識結果の内容に関連する語彙を獲得して認識辞書に追加することにより入力音声に適応した未登録語の少ない辞書を作成することができ、その辞書を用いて再度認識を行うことにより未登録語の影響を抑えて認識精度を改善できると考えられる。そこで本稿では、音声認識結果の内容に関連する語彙をコーパスから獲得し、獲得した語彙を辞書に追加することによって辞書を入力音声に適応させる手法を提案する。提案手法は、テキストコーパス中の語彙に対して語彙の分野を表す語彙分野ベクトルを算出しておき、入力の認識結果に対して発声内容の分野を推定し、その分野に近い語彙分野ベクトルを持つ語彙を入力に対する関連語彙として獲得し辞書に追加するという処理をオフラインで行う教師なし適応手法である。毎日新聞コーパスから各語彙の語彙分野ベクトルを求め、TVニュース音声を用いて提案手法の評価を行った。One of the most common problems in speech recognition system is the out-of-vocabulary (OOV) problem. Although we cannot avoid that input data includes OOV words for a vocabulary, we can reduce the number of them by adapting the vocabulary to inputs. Extracting relevant words to the content of an input speech based on a speech recognition result obtained using a reference vocabulary and adding them to the vocabulary enable to build an expanded vocabulary that includes less OOV words. The second recognition process using the new vocabulary is supposed to be performed better than the first process. In this paper, we propose vocabulary adaptation that acquires relevant words to an input from corpus and builds an expanded vocabulary by adding them to the reference vocabulary.

One of the most common problems in speech recognition system is the out-of-vocabulary (OOV) problem. Although we cannot avoid that input data includes OOV words for a vocabulary, we can reduce the number of them by adapting the vocabulary to inputs. Extracting relevant words to the content of an input speech based on a speech recognition result obtained using a reference vocabulary and adding them to the vocabulary enable to build an expanded vocabulary that includes less OOV words. The second recognition process using the new vocabulary is supposed to be performed better than the first process. In this paper, we propose vocabulary adaptation that acquires relevant words to an input from corpus and builds an expanded vocabulary by adding them to the reference vocabulary.

In this study, we analyzed user adaptation to machine translation(MT) systems through a controlled experiment. In the experiment, 67 subjects translated sentences using a Japanese-English-Japanese turnback translation, then we found out following results : 1) The more knowledge user has about source language, the better user can adapt to system, 2) The rule instruction beforehand affects for all user's adaptation, especially, it has a great effect for subjects who have medium level of source language, then following those who have low level, and those who have higher level.

In this study, we analyzed user adaptation to machine translation (MT) systems through a controlled experiment. In the experiment, 67 subjects translated sentences using a Japanese-English-Japanese turnback translation, then we found out following results: 1) The more knowledge user has about source language, the better user can adapt to system, 2) The rule instruction beforehand affects for all user's adaptation, especially, it has a great effect for subjects who have medium level of source language, then following those who have low level, and those who have higher level.

One of the most common problems in speech recognition system is the out-of-vocabulary (OOV) problem. Although we cannot avoid that input data includes OOV words for a vocabulary, we can reduce the number of them by adapting the vocabulary to inputs. Extracting relevant words to the content of an input speech based on a speech recognition result obtained using a reference vocabulary and adding them to the vocabulary enable to build an expanded vocabulary that includes less OOV words. The second recognition process using the new vocabulary is supposed to be performed better than the first process. In this paper, we propose vocabulary adaptation that acquires relevant words to an input from corpus and builds an expanded vocabulary by adding them to the reference vocabulary.

This paper describes an automatic multimedia content indexing system, which includes acoustic segmentation, automatic speech recognition, natural language processing, and video indexing features. The system mainly focuses on multimedia news programs. Speech segments, which are extracted from news content, are delivered to the speech recognition module. The speech recognition result sequence is segmented into topics, and those audio indexing results can be integrated with video indexing information to extract news story structure. Experiments show that the integrated results provide better news story structures than any single type of information

Recently XML has been considered as an emerging standard for describing resources on the Web. It is indispensable for next generation Web search systems to rank XML documents based on their relevance to the query and structural constraints.This paper describes requirements for earch systems, which handle XML documents describing resources on the Web. It then describes a new search system, which makes index files according to a format file specifying 1)the sub-structure of XML documents, 2)the type of index files and 3)the schema of extracting terms to satisfy the requirements.

Recently XML has been considered as an emerging standard for describing resources on the Web. It is indispensable for next generation Web search systems to rank XML documents based on their relevance to the query and stuructural constraints. This paper describes requirements for search systems, which handle XML documents describing resources on the Web. It then describes a new search system, which makes index files according to a format file specifying 1)the sub-structure of XML documents, 2)the type of index files and 3)the schema of extracting terms to satisfy the requirements.

The noun part of the WordNet 1.6 English lexical database was translated into Japanese by human translators. In the course of the translation work, elemental English words/collocations in a synset were translated into Japanese counterparts, as well as the "gloss", which defines the English lexical concept. This translation work turns out to give us an aligned and conceptually indexed pararell corpus. This paper shows the basic characteristics of the "Japanized" noun part of the WordNet. We argue that the resulted data might be a good resource for comparative linguistics. In addition to these, we examine how the derived bilingual lexical data can be applied to cross-linguistic natural language applications, such as cross-1anguage information retrieval.

This paper proposes an architecture for distrubuted cross-language informationo retrieval and a protocol used in it. The architecture includes metasearch engines and search engines, which communicate each other by using the protocol. The protocol is designed to communicate not only a search request and its result, but also inis also introduced, and it is concerned with formation of the functionalities and the contents provided by search engines The architecture allows a user to automatically choose and exploit various search engines probiding a large amount of documents in all sorts of languages in the Internet/WWW, according to his/her information need. The paper also introduced the joint project on a distributed cross-language information retrieval using the proposed architecture, involving KRDL in Singapore, KAIST in Korea and us NTT.

This paper proposes an architecture for distributed cross-language information retrieval and a protocol used in it. The architecture includes metasearch engines and search engines, which communicate each other by using the protocol. The protocol is designed to communicate not only a search request and its result, but also inis also introduced, and it is concerned with formation of the functionalities and the contents provided by search engines The architecture allows a user to automatically choose and exploit various search engines providing a large amount of documents in all sorts of languages in the Internet/WWW, according to his/her information need. The paper also introduces the joint project on a distributed cross-language information retrieval using the proposed architecture, involving KRDL in Singapore, KAIST in Korea and us NTT.

近年，電子化テキストの急激な増加，および，インターネットによる一般利用者の電子媒体への日常的なアクセスに伴って，言語処理研究と言語に関する実用技術の間のギャップが徐々に狭まってきており，実用的な自然言語処理研究という言葉が真に現実的な意味を持ち出してきた．本報告では，そのような実用的言語処理技術の事例のいくつかを「ここまでできるぞ言語処理技術」というタイトルで紹介する．In recent years, the amount of online text data rapidly increased, and ordinary users have come to access online texts through internet in their daily lives. This change around the situation of language data and language processing technologies gradually made the gap between research on natural language processing and practical technologies for processing natural language narrower, and has given the term &quot;reserach on practical natural language processing&quot; much more reality than before. This article reports case studies of those practical natural language processing technologies with the title &quot;Recent Advances in Language Processing Technologies&quot;.

In recent years, the amount of online text data rapidly increased, and ordinary users have come to access online texts through internet in their daily lives.This change around the situation of language data and language processing technologies gradually made the gap between research on natural language processing and practical technologies for processing natural language narrower, and has given the term "reserach on practical natural language processing" much more reality than before.This article reports case studies of those practical natural language processing technologies with the title "Recent Advances in Language Processing Technologies".

In recent years, the amount of online text data rapidly increased, and ordinary users have come to access online texts through internet in their daily lives. This change around the situation of language data and language processing technologies gradually made the gap between research on natural language processing and practical technologies for processing natural language narrower, and has given the term "reserach on practical natural language processing" much more reality than before. This article reports case studies of those practical natural language processing technologies with the title "Recent Advances in Language Processing Technologies".

近年, WWW上の検索エンジンが, 盛んに開発されている。このような検索エンジンは, 通常, 検索結果を検索要求と結果の類似度あるいは関連度に基づいて順位付けし出力する。本稿では, この検索結果を様々な観点 (ビュー)から再構成することにより, ナビゲーションを支援する機能について報告する。

World Wide Web(WWW)上の情報空間の拡大に伴い、いわゆる情報資源の探査の問題が顕著化し、WWW空間からの効率的な情報探索手法に関して議論が進められている。情報量が増大し、更に内容が多様化している情報空間の効率的なアクセスのためには、従来の情報探索手法はそのまま適用できない.本稿では、単なる情報空間の探索だけではなく情報空間の知的アクセス手法の実現を目指したインターネット情報探索システムTITAN (Total Information Traverse AgeNt)に関して述べる。

Recent progresses of the World-Wide Web (WWW) are remarkable. Already huge amount of information is available on the WWW, and it is expected to grow further. However, in general, it is not that easy for users to retrieve relevant information from the distributed information resources on the WWW. A user navigation processes on the WWW should consist of two sub processes; a process for locating relevant information resources based on their information needs, and that for browsing resources around the ones located initially. This paper first argues the resource discovery problem on the WWW, and proposes an incremental approach to it. A currently being developed prototype system, named TTTAN Ver.1, for the WWW navigation is then described. Finally, future research directions are discussed in terms of intelligent information processing including natural language processing.

Natural language processing is one of the most important technologies in the computerlized society. Special Interest Group of Natural Language Processing of IPSJ has been continuing its high activity since its establishment. In this paper we present the topics of the papers presented at its periodical conferences in this one decade, focusing on these three years. We also list a set of key words for the further elaboration in natural language research.

1992年7月23日から28日にかけて、第14回計算言語学国際会議 (e 14 th International Conferenceon Computational Linguistics，COLING?) がフランスのナント市会議場で開催された。COLINGは2年に1回開催されている計算言語学、自然言語処理関係の最大の国際会議である。今回は、500人以上の参加者を集め、約420件の投稿から採択された約200件の論文発表が行なわれた。本報告では、会議の概要および、いくつかの興味ある論文を紹介する。The 14th International Conference on Computational Linguistics (COLING-92) was held at City Congress Hall of Nantes, France from July 23rd to 28th. COLING is the largest international conference in the field of computational linguistics and natural language processing. Over 500 people from all over the world attended the conference and about 200 papers were presented. This report gives the outline of the conference and summarizes several papers, which the authors are interested in.

談話やテキストでは、2文間に何らかの関係を持ちながら展開していることが多く、その連鎖の仕方には対象世界の特徴が反映されていることが多い。これに基づき、日本語テキストの構造を、隣接する2文の並びのパターン（連接パターン）を基に解析する手法を提案する。ここでは、単付文のタイプとそれを接続する接続属性の組み合わせにより、文をタイプ分類し、そのタイプの並びのパターンから単位文をノードとするネットワーク構造を導出する手法を述べる。また、クローズドデータ、オープンデータによる評価実験の結果から、本手法の有効性、解析規則の収束性について考察する。This paper proposes a new method to analyze cohesive relations between adjacent Japanese sentences based on constraints naturally contained in objective texts. We especialy focused on explanative texts such as technical manuals. In these texts, the world to be explained constrains the sentence type and sentence adjacent patterns. For example, text for explaining a software command will describe some user operations and the associated system responses in some cohesive ways. We have developed a sentence type system suitable for software manuals and constructed a set of analysis rules. An experimental result toward the closed data shows the effectiveness of our method. We also discuss the coverage of the rule set through an experiment toward open data.

マニュアル文書を対象とした日本文推敲支援システムの研究開発を進めている。これは文章中の推敲すべき表現(不適性表現)を検出するとともに、それに対する書換え候補を提示するシステムである。マニュアルに頻出する不適性表現の1つに長文があり、上記システムにおいては、形態素解析情報に基づく文分割機能の検討を進めている。従来の文分割方式では、基本的に(1)分割する文節(以下、分割点と呼ぶ)の接続形式と分割時に挿入する接続詞を一対一に対応させる、(2)分割点を終止形に変換する、という方式が多い。このとき連用中止表現や接続助詞「て」による接続(以下、本稿では両者を併せて連用中止表現と呼ぶ)は、通常並列接続として扱われるため、上記のような方式では「並列の接続詞の挿入(または接続詞なし)、分割点文節の終止形への書換え」という分割方法を採ると考えられる。しかし、連用中止表現はその前後に特定の意味関係(以下、連用中止表現の意味用法という)を担うことがあることが指摘されている。また実際にマニュアル文中の連用中止表現を収集した結果でも、単なる並列接続とみなし得ないものが比較的多く存在する[figure]1(1)。さらに連用中止表現は、述語に付加される付属表現(時制表現や様相表現)が省略された表現形式であるため、分割点文節を終止形に変換するという方法では有効な分割文候補を生成することができない場合がある[figure]1(2)(3)。本稿では、文分割の観点から連用中止表現の意味用法を整理するとともに、表現の形態的特徴と意味用法、分割時の接続詞、連用中止文節の形態素操作との関係に着目した連用中止表現を分割点とする文の分割方式について述べる。

計算機による文書処理の研究開発が盛んに行われている。その中の一分野に、訂正支援あるいは推敲支援といわれる分野がある。成熟しつつある形態素解析技術を基盤として、単語レベルの訂正/推敲を支援するシステムがいくつか発表されている。しかし、文/文章レベルの推敲を支援するシステムは、あまり発表されていない。文/文章レベルの推敲支援を実現するためには、構文解析、意味解析などの言語解析の結果を用いてその中に含まれる推敲を心要とする表現(=不適性表現)を検出する技術をまず確立する必要がある。その際、不適性表現検出のための規則及びそれに付随する手続きをどのような枠組みで記述するか、また、その解釈をどのような機構によって行うかは、大きな問題である。 本稿では、特に文レベルの推敲支援を実現する際にその基本となる文レベル不適性表現の検出機構を提案する。

本稿では、日本語用言の持つ結合価構造を文解析の色々な局面に適用する手法について述べる。結合価構造は、結合価パターンと呼ぶ構造パターンとして辞書中に記述される。この構造パターンは、単位文の構造決定を行なうためのパターンマッチングに適用される他、埋め込み名詞句の構造決定、複合用言におけるゼロ代名詞化された格要素の補完、複合名詞・複合用言の解析などにも適用される。すなわち、結合価構造に基づくことにより、これらの言語現象に対しても統一的な解析を行なうことが可能となる。本稿では、さらに日英機械翻訳への適用における問題点についても論じる。In this paper, a method to apply the valency structure of Japanese predicates to Japanese sentences analysis is presented. In the dictionary, each of the valency structures is formulated as a sturucture pattern named &quot;Valency-pattern&quot;. This strutural pattern is applied to the analysis of unit-sentences, suppliment of zero-pronoun case-elements. analysis of embedded noun-phrases, analysis of complex words. These analysis processes are discussed. Taking this structure to be basis of the Japanese sentences analyzing process allows the uniform treatment of such language phenomena. The problems in applying this method to Japanese-to-English machine-translation are also discussed.

現在、開発中である日英翻訳システムALT-J/Eにおける日英変換技術について報告する。本システムの変換処理は、今回新たに提案する多段変換方式の考え方に基づいて設計されている。本稿では、多段変換方式の考え方、本システムにおける変換処理の位置付け・特徴、変換処理の概要及び、これに用いるパターン変換規則辞書の構成法について述べる。

現在、開発中の日英翻訳システムALT.-J/Eにおける日本語解析技術について報告する。本システムの日本語解析処理は、解析の各段階で生じる多義を効率的に絞り込むこと、多様な語いと表現をもつ文を取り扱えることに重点をおいて設計されている。以下、本稿では、上記のような点をいかにして達成するかということに的を絞り、本システムの日本語解析手法について述べる。

The recent research direction toward multimodal semantic representation would be further advanced, if we could have a machinery to collect adequate images from the Web, given a target concept. With this motivation, this research particularly investigated into the Web imageabilities of the behavioral features of a basic-level concept. The primary contributions made in this research are twofold: (1) "beaver building dams"-type queries can better yield relevant Web images, suggesting that the present participle form ("-ing" form) of a verb, as a query component, is more effective than the base form; (2) the behaviors taken by animate beings are likely to be more depicted on the Web, particularly if the behaviors are, in a sense, inherent to animate beings (e.g., motion, consumption). Web further analyzed linguistic annotations that were independently given to some of the images, and discusses an aspect of the semantic gap between image and language.

There are two methods for multilingualcommunication: using machine translation and using parallel texts. We studied the high-reliability of Back translation and parallel textsfor high-reliability multilingual communication. We also developed multilingual support software for medicalfield and education field.

Focusing on complex verbs, language resources, dictionaries, used in natural language processing were analyzed. Verbs in a dictionary used for morphological analysis were annotated by their morpho-semantic relations and structural relations, such as composition and deliberation. Meaning descriptions of complex verbs in a machine readable dictionary were analyzed from the perspective of their relations to the meanings of their components. These analyses provide some clues to organizing multiple language resources and providing those with a background of lexical conceptual structure theory.

It is crucially important to understand potential differences in concept lexicalization among various languages. To tackle this linguistic issue and to develop a computational mechanism for constructing a useful multilingual lexical ontology based on existing lexical resources, the reported research developed a method to associate possibly corresponding lexicalized concepts across languages by using a sense-tagged corpus, and proposed a framework for representing a dynamically acquired cross-lingual/ interlingual lexical-conceptual relations.