コロキアムB発表

日時: 6月9日(木)3限(13:30-15:00)


会場: L1

司会: 嶋利 一真
海木 延佳 D, 中間発表 知能コミュニケーション 中村 哲, 渡辺 太郎, Sakriani Sakti
title: Neural speech synthesis using local phrase dependency structure information
abstract: In order to synthesize Japanese speech with natural prosody, we introduce an end-to-end TTS with new prosodic symbol representing phrase components based on local phrase dependency structures to end-to-end text-to-speech synthesis (TTS). In this paper, we propose two TTS models: 1) a model with prosodic symbols that represent the depth at phrase boundaries, and 2) a model with prosodic symbols that reflects a folded model of phrase and accent components based on a prosodic generation control mechanism. In synthesized speech at left-branching boundary using these two models, 1) pause indicating the phrase boundary is generated. 2) the re-rebuilding phrase component of F0 may occur. To verify the effect of these two proposed models on a conventional model using prosodic symbols using only accent components, we conducted a subjective evaluation on the AB test. As a result, it was confirmed that by using local phrase boundary information of sentences and prosodic generation model in Japanese end-to-end text-to-speech synthesis, synthetic speech with more natural prosody that reflects the intention of the utterance could be generated.
language of the presentation: Japanese
発表題目: 局所的な句構造の情報を用いたニューラル音声合成
発表概要: 自然な韻律をもつ日本語音声を合成するため、局所的な句構造に基づくフレーズ成分を表す韻律記号をend-to-end音声合成に新たに導入すること提案する.本発表では、フレーズ成分を表現するために、1)句境界に係り受けの深さを表す韻律記号を追加するモデルと、2)韻律生成制御機構に基づき、フレーズ成分とアクセント成分の重畳型モデルを反映させた韻律記号を採用するの2つのモデルを提案する.この2つのモデルを用いた音声合成により,右枝分かれ境界音声の係るり受けの深い句境界位置においてで、1)フレーズ境界を示すポーズが生成されることがあること .2)F0のフレーズ成分の立て直し現象が生じることがあることが観察された.アクセント成分のみの韻律記号を用いた従来モデルに対し、これら2つの提案モデルの効果を検証するため対比較の聴取実験を行った.この結果、日本語end-to-end音声合成に文の局所的な句境界の情報や、韻律の生成モデルを取り入れることにより、発話者の意図をより正しく反映した自然な韻律を持つ合成音声が生成できることが確認された.
 
NOVITASARI SASHI D, 中間発表 知能コミュニケーション 中村 哲, 渡辺 太郎, Sakriani Sakti
Title: A machine speech chain approach for self-adaptive Lombard TTS in noisy environments
Abstract: Recent end-to-end text-to-speech synthesis (TTS) systems have successfully synthesized high-quality speech. However, TTS speech intelligibility degrades in noisy environments because most of these systems were not designed to handle noisy environments. Several works attempted to address this problem by using offline fine-tuning to adapt their TTS to noisy conditions. Unlike machines, humans never perform offline fine-tuning. Instead, they speak with the Lombard effect in noisy places, where they dynamically adjust their vocal effort to improve the audibility of their speech. This ability is supported by the speech chain mechanism, which involves auditory feedback passing from speech perception to speech production. We propose an alternative approach to TTS in noisy environments that is closer to the human speech chain and the Lombard effect. Specifically, we implement Lombard TTS in a machine speech chain framework to synthesize speech with dynamic adaptation. Our TTS performs adaptation by generating speech utterances based on the auditory feedback that consists of the automatic speech recognition (ASR) loss as the speech intelligibility measure and the speech-to-noise ratio (SNR) prediction as power measurement. Two versions of TTS are investigated: non-incremental TTS with utterancelevel feedback and incremental TTS (ITTS) with short-term feedback to reduce the delay without significant performance loss. Furthermore, we evaluate the TTS systems in both static and dynamic noise conditions. Our experimental results show that auditory feedback enhanced the TTS speech intelligibility in noisy environments.
Language of the presentation: English
 
MUTINDA FAITH WAVINYA D, 中間発表 ソーシャル・コンピューティング 荒牧 英治, 渡辺 太郎, 若宮 翔子, 矢田 竣太郎
Title: Automatic Data Extraction to Support Meta-Analysis Statistical Analysis: A Case Study on Breast Cancer
Background: Meta-analyses aggregate results of different clinical studies to assess the effectiveness of a treatment. Despite their importance, meta-analyses are time-consuming and labor-intensive as they involve reading hundreds of research articles and extracting data. The number of research articles is increasing rapidly and most meta-analyses are outdated shortly after publication as new evidence has not been included. Automatic extraction of data from research articles can expedite the meta-analysis process and allow for automatic updates when new results become available. In this study, we propose a system for automatically extracting data from research abstracts and performing statistical analysis.
Materials and Methods: Our corpus consists of 1011 PubMed abstracts of breast cancer randomized controlled trials annotated with the core elements of clinical trials: Participants, Intervention, Control, and Outcomes (PICO). We proposed a BERT-based named entity recognition (NER) model to identify PICO information from research abstracts. After extracting the PICO information, we parse numeric outcomes to identify the number of patients having certain outcomes for statistical analysis.
Results: The NER model extracted PICO elements with relatively high accuracy, achieving F1-scores greater than 0.80 in most entities. We assessed the performance of the proposed system by reproducing the results of an existing meta-analysis. The data extraction step achieved high accuracy, however the statistical analysis step achieved low performance because abstracts sometimes lack all the required information.
Conclusion: We proposed a system for automatically extracting data from research abstracts and performing statistical analysis. We evaluated the performance of the system by reproducing an existing meta-analysis and the system achieved a relatively good performance, though more substantiation is required.
Language of the presentation: English
 

会場: L2

司会: Md.Delwar HOSSAIN
SITTHITHANASAKUL SUPAVAS D, 中間発表 ソフトウェア工学 松本 健一, 飯田 元, 石尾 隆, 畑 秀明, Raula Gaikovina Kula
Title: On the Usage of Non-textual README Elements in OSS Projects in GitHub
Abstract: README files play an essential role, serving as the face of a software repository, and the initial point of contact for potential users and contributors for an Open Source Software (OSS) repositories. With the recent evolution of documentation, non-textual elements has been utilized to enhance and attract visitors to the repositories. In this paper, we conduct an empirical study to analyze how non-textual elements have been adopted in README files. First, we conducted an preliminary study on 4.4 million GitHub README files to identify five non-textual elements, i.e., image, table, emoji, code and link, revealing the code snippets and links being most prevalent. Then, taking a qualitative sample of 768 non-texual elements(i.e., code snippets and links), we then characterize their usage in README files. We find that script (55.21%) and source code (29.69%) examples are common code snippet usages, while links tend to target software homepage (42.19%). The study highlights how the README file could be a useful resource not only for documentation, but can be exploited for other non-textual elements that it contains.
language of the presentation: English
 
MAEPRASART VITTUNYUTA M, 2回目発表 ソフトウェア工学 松本 健一, 飯田 元, 石尾 隆, 畑 秀明, Raula Gaikovina Kula
title: How Do External Pull Requests contribute to a NPM library?
abstract: Third-party libraries are now commonplace when building contemporary software applications. Despite their popular use, most libraries are open source software that often rely on volunteer (usually unpaid and overworked) contributions from outside the core team (external). To understand the role that these contributions play to sustain third-party libraries, we analyze Pull Requests (PRs) submitted from outside the core team of contributors (i.e., External PRs). We analyze 945,291 PRs from the NPM ecosystem to empirically investigate External PR prevalence, acceptance, need for attention, and content. Our findings indicate that External PR are indeed prevalent (75.02% of all received PRs are External PR and 88.87% submitted on average). Interestingly, External PRs are just as likely to be accepted compared to Internal PRs (50% per package). For a package, we find that on average, 26.75% of External PRs submitted are linked to an already existing issue, and require the attention of the core team, with labels such as breaking changes, urgent, and on-hold. In a qualitative analysis, we find that External PRs focus on documentation (44 out of 384 PRs), which is different to adding new features (380 out of 384 PRs) for Bot PRs, and refactoring (34 out of 384 PRs) for Internal PRs. The study confirms that external contributions play a key role in the ecosystem, and reinforces the call for help, especially from single maintainer libraries that serve a massive client user-base.
language of the presentation: English
 
SAKULNIWAT TATTIYA M, 2回目発表 ソフトウェア工学 松本 健一, 飯田 元, 石尾 隆, 畑 秀明, Raula Gaikovina Kula
Title: Recommending Python Code Snippets Based on Competency: A Case Study of Stack Overflow
Abstract: Stackoverflow is one of the biggest question-and-answer cite used by developers with millions of visits per month and 13.6 seconds average time between new questions. Sometimes there are more than one answers per question which leads to the problem stated that what kind of answer is likely to be the most recommended. The goal is to find recommended answers focusing on competency of Python code snippets containing in the question and asker competency, using pycefr tool that classifies Python code based on the competency. We conduct qualitative analysis by grouping the answer based on score and manual classify the related questions based on the type of the question asked. In addition, we perform quantitative analysis by sorting response time of the answers and find relationship between the time and competency. From qualitative analysis, we find that each type of questions has slight difference with all of them fall to basic competency while from quantitative analysis, we find that the more competency the answer,the shorther timespan from the longest to shortest response time. Overalls, the study shows the sign of how the code snippets competency is one of the factor that can encourage the asker and answer which type of Stackoverflow answers they are looking for.
Language of the presentation: English
 
HOVHANNISYAN ANI M, 1回目発表 ソフトウェア工学 松本 健一, 飯田 元, 石尾 隆, Raula Gaikovina Kula
Title: Partial static analysis of modified source code
Abstract: At the moment, large scale source code becomes massive, and its maintenance becomes difficult. To reduce the amount of time on review of source code it is recommended to use pre-checking tools, for instance, Static Analysing Tools (SAT), which can easily detect bug-full code. However, SAT mostly scan full source code, and mostly do not support caching mechanisms for reanalyzed code, also they give too extensive list of warnings not related to developers work. Therefore, our proposal is to reduce the amount of source code by extracting only modified methods, after that, partially analyze only modified method-level files. The research proposes the extraction from full source code to method-level files mechanism. And the initial experiments show, modified code analysis reports similar warnings as full analysis does, but the time is not too different. In the next year of course, research assumes to have the datasets with extracted method-level files, then to make a comparison between full source code analysis and partial analysis by considering essential factors.
Language of the presentation: English