Takagi KAZUYUKI (高木 一幸)

Offer Organization: （研）科学技術振興機構「ＳＩＰ」, System Name: 研究助成, Category: -, Fund Type: competitive_research_funding, Overall Grant Amount: - (direct: 1550000, indirect: 0)
模擬実験用管路で収集したデータを検聴してイベントのラベル付けを行い、パターン学習および評価用のデータを整備する。
-上記のデータに基づき、漏水箇所の音響的特徴を解析し、パターン認識処理に適した信号処理方法の設計を行う。
-適切な条件で信号処理された音響的特徴を基に統計的音声認識技術を応用した機械学習を行い、漏水箇所とそれ以外の箇所の自動検出の可能性を検討する。

Offer Organization: Japan Society for the Promotion of Science, System Name: Grants-in-Aid for Scientific Research, Category: Grant-in-Aid for Scientific Research (C), Fund Type: -, Overall Grant Amount: - (direct: 3000000, indirect: -)
1.In this work, we use a corpus in which pairs of newspaper articles and corresponding hand-made short summaries are contained. This corpus provides information about how humans make short summaries. To obtain such information effectively, phrase alignment is necessary between the original sentence and its summary. We developed a phrase aligner that makes use of conceptual distance and inter-phrase dependency.
2.Before the research period started, we were using the inter-phrase dependency strength estimated from the distribution of dependency distance in the set of original sentences. This method misses, however, the relationship between the original sentence and its summary. In this work, we estimated the inter-phrase dependency strength from the relative frequency of phrase pairs that exist in the original sentence with a certain dependency path length and remain having modifier-modified relation in the corresponding summary. The result of a subjective evaluation experiment showed significant improvement in the quality of compressed sentences.
3.In the phrase extraction type sentence compression, which is employed in this research, phrases that are not in modifier-modified relation in the original sentence sometimes appear to have modifier-modified relation in the compressed sentence. Such a phenomenon may degrade the readability of compressed sentences. We worked out a method to modify the phrase ending of the modifier-phrase for improving the readability of compressed sentences. The result of subjective evaluation experiment showed the effectiveness of the method.
4.We reformulated our sentence compression method in a probabilistic framework. In calculating the probability that a compressed sentence is generated from an original sentence, quantities similar to phrase significance and inter-phrase dependency appear, which can be estimated from a training corpus. It was shown that this probabilistic approach attains comparable performance as our former, heuristic approach.

Offer Organization: Japan Society for the Promotion of Science, System Name: Grants-in-Aid for Scientific Research, Category: Grant-in-Aid for Scientific Research (C), Fund Type: -, Overall Grant Amount: - (direct: 1100000, indirect: -)
1. Development of Sentence Compression Algorithm
The sentence compression problem was formulated as a problem of selecting an optimal subsequence of phrases from a given sentence. Then, based on our dependency analysis technique, an efficient algorithm was developed to solve the problem.
2. Estimation of inter-phrase dependency strength and phrase significance
By using about 34,000 sentences in Kyoto University Corpus, inter-phrase dependency strength was estimated. It is based on the statistical frequency of inter-phrase dependency distance, and was estimated for each modifying phrase class and modified phrase class. Also, a sentence compression experiment was conducted in which human subjects compressed 200 sentences. The result was analyzed statistically and the remaining rate for each phrase class was calculated. Based on the result, phrase significance value for each phrase class was estimated.
3. Subjective Evaluation of Compressed Sentences
A subjective evaluation experiment was performed for sentences automatically compressed by using the above algorithm together with the estimated inter-phrase dependency and phrase significance. In this experiment, 200 test sentences, which are different from the sentences in 2, were used. 5 subjects were employed for evaluating the quality of compressed sentences. Subjective evaluation was performed from the following points of view : (a) total impression, (b) retention of information, and c grammatical correctness. For comparison, the same kind of evaluation experiment was done for sentences compressed by humans, and also by a random method. It was found that the quality of sentences compressed by the proposed method lies just between those of human compression and random compression.
4. Segmentation of Long Sentences
Because long sentences are difficult to analyze syntactically, it is desirable to segment long sentences into shorter ones. In this work, a support vector machine (SVM) technique was applied to the problem. Vectors consisting of surface attribute values of relevant phrases were input to the SVM, and segmentation points were automatically estimated. As a result, 77% of precision and 84% of recall were obtained. Correct sentence segmentation rate was 72%.

Offer Organization: 日本学術振興会, System Name: 科学研究費助成事業, Category: 特定領域研究, Fund Type: -, Overall Grant Amount: - (direct: 63400000, indirect: -)
1.音声知覚における韻律の役割解明と音声認識への応用
(1)句頭アクセント核の検出とそれに基づく仮説探索制御を実装した.単語アクセントは前後の環境により変化するが,句頭に核が存在した場合は,その単語は必ず一型となる.この規則の基づき,句頭のF0情報よりその語が一型となる事後確率を求め,韻律スコアを導入した.連続音声認識システムJuliusに本モジュールを実装し,大語彙連続音声認識におけるその有効性を示した.
(2)音声の時間構造を,局所話速の分析を中心に,文内の文節継続長を決定する統計モデル,文節内のモーラ継続長制御モデル,モーラ内での子音継続時間長制御モデルの3階層でモデル化した.また,それぞれのモデルについて時間構造の知覚実験を行い,時間的制約について検討した.
2.発話の構文・意味解析における韻律情報の利用
(1)これまで利用した着目文節の直後のポーズと着目文節の直後の文節の直後のポーズに加えて,着目文節の直前のポーズを利用することにより,係り受け解析の精度が向上することを確認した.また,これらのポーズ情報にF0情報を加えることにより,さらなる解析精度の向上が得られた.
(2)多数の話者による音声データを用いて不特定話者条件の係り受け解析実験を行った結果,ポーズ長とF0特徴量のモデルは従来より簡単なものでよいこと,ポーズ長は平均音節継続長で正規化した方が良いことなどがわかった.また,大量のコーパスを用いて評価文に対する被覆率が高い係り受け規則を新たに作成した.
3.音韻情報と韻律情報を統合した音声認識・理解システム
ディクテーションシステムにおける入力補完候補の絞込みに,アクセント情報を利用する手法を開発した.また,アクセント情報の認識・ディクテーション・入力補完機能を統合した予測型音声入力システムを実装し,アクセント情報利用の有効性を検証した.
4.韻律的特徴を用いた講演音声の自動要約
重要文抽出によって講演音声の要約を自動生成するために,文単位と文重要度を韻律情報を利用して決定する手法について検討した.ポーズで区切られた発話単位境界に対し,文境界とすべきかどうかを判断する決定木を学習し94%の分類率を得た.文重要度の決定において,連続音声認識による誤りを含む言語情報奪利用する場合の方が,正しい言語情報を利用する場合よりも,韻律情報の効果が大きいことを示した.

Offer Organization: Japan Society for the Promotion of Science, System Name: Grants-in-Aid for Scientific Research, Category: Grant-in-Aid for Scientific Research (C), Fund Type: -, Overall Grant Amount: - (direct: 3000000, indirect: -)
Results of this research project can be classified into 1. theoretical basis, 2. bunsetsu segmentation and segmentation of long sentences, 3. use of prosodic information, and 4. sentence compaction, all related to the minimum total penalty method.
1. Dependency analysis was investigated from a view point of "minimum cost segmentation problem". It was shown that various dependency analysis algorithms can be derived by changing the definition of the cost. It was also made clear that the minimum total penalty method allows the use of a wide range of numerical information as linguistic knowledge.
2. It is necessary to segment a sentence into bunsetsu phrases prior to dependency analysis. In this work, a decision tree method was applied to this problem, giving higher segmentation accuracy than conventional methods data. The decision tree technique was also applied to segmentation of long sentences, which is pre-processing for dependency analysis. It was demonstrated that a set of segmentation rules was automatically acquired by this method.
3. To find out syntactic information contained in prosodic features, a statistical model was created that represents a relationship between prosodic features and inter-phrase dependency distance. The model was then incorporated into the minimum total penalty parser to measure the effectiveness of prosodic information for dependency analysis. The duration of pause was found to be very effective. Further investigation is necessary to make use of prosodic features related to pitch, power, and speaking rate.
4. An efficient sentence compaction algorithm was developed for such application as generation of on-line TV closed-captions. This algorithm selects an optimal bunsetsu subsequence from an original sentence that maximizes the sum of bunsetsu importance scores and inter-phrase dependency scores. Future work includes investigation of better definitions of bunsetsu importance score and inter-phrase dependency score. It is also necessary to evaluate the quality of shortened sentences using large amount of test data.