| 平尾 俊貴 | D, 中間発表 | ソフトウェア工学 | 松本 健一, 飯田 元, 石尾 隆, Raula G. Kula |
|
title: Understanding Factors that Impact Code Review Efficiency For the Automation
abstract: Software quality assurance plays an important role in ensuring high quality and performance of softwares. Code review is a well known practice where developers manually assess source codes and find a variety of potential problems; However, code review process is expensive in terms of human resource. Since code review is made up of multiple activities (e.g., patch submission and remote discussion), little is know about which aspect can affect code review efficiency to automate code review process. To address this concern, it is essential to understand what affects code review efficiency. This research comprises of three main studies which are (i) Process aspect; (ii) Human aspect; and (iii) Product aspect. language of the presentation: English | |||
| 井ノ口 輝 | 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 |
|
title: Detection of Academic Paper Referencing in Source Code Comments
abstract: During implementing code, academic papers can be knowledge source of programming. Previous study reported that several papers are referenced in source code comments of C and Java projects. Since the previous study analyzed those referencing manually, the scale of the analysis was limited. In this study we try build machine learning classifier to identify academic paper referencing code comments in large scale dataset. language of the presentation: Japanese | |||
| 田内 遥夏 | 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 |
|
title: Identifying "Waiting" Self-Admitted Technical Debts
abstract: Technical debt is a metaphor for non-optimal solutions in software development. Although it can accelerate development in the short term, it can cause issues in software maintenance in the long term. Self-admitted technical debts (SATD) are one of technical debts that are intentionally introduced in source code along with comments, such as “FIXME” and “Hacky.” In this study, we focus on a specif type of SATD, waiting SATD, which are non-optimal solutions because of other factors and waiting for them resolved. We try build machine learning classifier to identify waiting SATD automatically. language of the presentation: Japanese | |||
| 入江 琴子 | M, 2回目発表 | ソフトウェア設計学 | 飯田 元☆, 松本 健一, 片平 真史（客員）, 石濱 直樹（客員） |
|
title: *** Comparing safety analysis methods from the perspectives on change impact analysis during system development ***
abstract: *** When a change is applied to system artifacts, an impact corresponding to the applied change also affects safety analysis. Developers frequently use FMEA(Failure Mode and Effect Analysis), a safety analysis method, to determine the existence of specific safety risks after a new change was applied. However, FMEA cannot identify all the impacts from complex systems because of its limitations, for instance, it does not consider the effect of multiple failures. Therefore, this study set up hypothesizes for differences regarding the change impact analysis between FMEA and FTA (Fault Tree Analysis), a frequently used safety analysis method along with FMEA, and then show the effects of uses these methods by examining each hypothesis. This presentation will mainly explain the experiment for evaluating these hypothesizes. *** language of the presentation: *** Japanese *** 発表題目: *** システム開発過程での変更の影響度から見る安全解析手法の比較検討 *** 発表概要: *** システムの開発工程の成果物に再び変更が適用される場合，そのシステム開発と並行して行われた安全解析にもその変更に対応する影響が及びうる． 開発の成果物が変更され，特定の安全性に関するリスクが存在するか判断する際，開発者は安全解析手法であるFMEA(Failure Mode and Effect Analysis)の解析結果を頻繁に用いる．しかし，FMEAは多重故障の影響を考慮しないなどの短所があるため，近年の複雑なシステムにおいてFMEAのみを用いる場合は識別しきれない安全解析への影響箇所が存在する可能性がある．そこで，本研究ではFMEAおよびFMEAとともによく用いられる安全解析手法であるFTA(Fault Tree Analysis)との間に，変更の影響分析において評価の差異が出そうな箇所を挙げて仮説を立て，それぞれの仮説を検証することで両手法の結果を影響分析に用いた際の効果を示す．本発表では，上記の仮説を評価するための実験に向けた計画について重点的に述べる．*** | |||
| 勝見 久央 | M, 2回目発表 | 知能コミュニケーション | 中村 哲, 松本 裕治, 吉野 幸一郎, 須藤 克仁 |
|
title: Optimization of Information-Seeking Dialogue Strategy for Argumentation-Based Dialogue System
abstract: Argumentation-based dialogue systems, which can exchange arguments with their dialogue partners, have been widely researched. It is required that the system has sufficient supportive information to argue its claim rationally. However, in realistic situations, the system does not often have enough supportive information; thus the system needs to acquire the missing information from the dialogue partner. This information acquisition process is known as ``information-seeking dialogue''. Existing works on information-seeking dialogues are based on the handcrafted dialogue strategy that exhaustively examines the missing information. When the system's dialogue partners are real peoples, the dialogue strategy to collect necessary information effectively is also essential. However, it is not trivial to manually design such a dialogue strategy, because the number of question candidates grows more significant as that of the information candidates increases. Thus, we formalize the process of information-seeking dialogue as Markov decision processes (MDPs) and apply Deep Reinforcement Learning (DRL) for automatically optimizing the dialogue strategy. By utilizing DRL, our method enables to optimize the dialogue strategy for two objective functions: the success in information collection. We also proposed an incorporation method with conventional heuristic dialogue strategies for exploring the strong candidates at the early stage of learning to avoid the cold start of DRL. We conduct experimental dialogues using the arguments from ``Twelve Angry Men dataset", and evaluate how quickly the proposed system collects enough information to argue the main claim. Experimental results show that the strategy trained by the proposed DRL-based optimization realizes a supporting information collection with fewer questions. language of the presentation: Japanese | |||
| 竹内 瞭 | M, 2回目発表 | 知能コミュニケーション | 中村 哲☆, 松本 裕治, 荒牧 英治, 諏訪 博彦 |
|
title: Crowd sound estimation by Social media
abstract: The atmosphere of the city is a combination of various information such as landscape, environmental sound, odor, etc. When searching for real estate or traveling, if you want to know the atmosphere of the city, it is most likely to use street view, aerial photography, etc. However, they can obtain only landscape information. Under this condition, this research aims to estimate all environmental sounds using social media. This present was focused on crowd sounds in environmental sounds and tried to estimate.As a result, it was not possible to obtain a significant correlation between the estimated value and the correct answer data. Finally, this presentation will discuss future tasks and methods. language of the presentation: Japanese | |||
| 中村 良 | M, 2回目発表 | 知能コミュニケーション | 中村 哲, 松本 裕治, 須藤 克仁, 吉野 幸一郎 |
|
title: Ground Truth-level Diversity in Neural Dialogue Generation
abstract: Neural network-based dialogue systems are widely researched; however, response generations by most existing systems have very low diversities. The main reason for this problem is the Maximum Likelihood Estimation (MLE) with softmax cross-entropy loss, because a generated response is not uniquely determined in real dialogue, although MLE trains the model to generate the most frequent response from enormous generation candidates. In this paper, we propose a new objective function called Token-Frequency-Dependent (TFD) loss, that scales the loss smaller for frequent token classes and larger for rare token classes, individually. It encourages the model to generate rare tokens aggressively even if the model fails to generate frequent tokens. On the OpenSubtitles dialogue dataset, our loss model establishes a state-of-the-art DIST-1 of 7.56, the unigrams diversity score, while maintaining a good BLEU-1 score. On the Japanese Twitter replies dataset, our loss model achieves a DIST-1 score comparable to ground truth. language of the presentation: Japanese | |||