The main idea in On-The-Fly Computing is to automatically compose existing software services according to the wishes of end-users. However, since user requirements are often ambiguous, vague and incomplete, the selection and composition of suitable software services is a challenging task. In this report paper, we present our current approach to improve requirement descriptions before they are used for software composition. This procedure is fully automated, but also has limitations, for example, if necessary information is missing. In addition, and in response to the limitations, we provide insights into our above-mentioned current work that combines the existing optimization approach with a chatbot solution.
Copyright c 2018 by the paper's authors. Copying permitted for private and academic purposes.
1See https://sfb901.upb.de for more information about OTF Computing.
Requirements
Software End user
Service market Composition +
Software services
Service 1
...
Service 2
Service 3 not always expedient and, above all, not always what the user originally wanted. For this reason, we think that minimal communication between the compensation system and the user is required, even if this means slowing down the creation of software compositions. We also believe that this can be realized through chatbot technology. In this case, the challenge is to keep the conversation with the user so slim and meaningful that it will not be annoying for end-users. Furthermore, untrained users must have all the information needed for a decision (e.g., description of the problem and examples). This is part of our current and future research activity (see Section 4). 2
The development of techniques and processes for the automatic on-the- y con guration and provision of individual IT services is the goal of the Collaborative Research Centre 901 \On-The-Fly Computing" at the University of Paderborn, Germany. In particular, subproject B1 \Parameterized Service Speci cations" deals with the e cient processing of di erent types of requirements speci cations, which enable the successful search, composition and analysis of services. In the very rst phase (until June, 2015), the focus of the project was on developing a speci cation language for services in order to automatically and e ciently process them. In the second phase (until June, 2019), user-friendly speci cations are introduced. These are intuitively understandable speci cations and are used primarily by end-users and domain experts. However, these requirements descriptions have errors that need to be corrected using Natural Language Processing (NLP). This is the task of the NLP team in subproject B1, which consists of Prof. Dr. Michaela Geierhos and Dr. Frederik Baumer. Prof. Geierhos is a full professor for Digital Humanities at the University of Paderborn, Germany and Dr. Baumer works as a postdoctoral researcher at the same chair.
Michaela Geierhos. Before Prof. Dr. Geierhos became a professor for Digital Humanities, she was an assistant professor for Semantic Information Processing at the University of Paderborn and worked as a postdoc at the Ludwig-Maximilians-University in Munich. After completing her studies in computational linguistics, computer science and phonetics, she worked as a research associate at the Center for Information and Speech Processing from 2006 to 2012. In 2010 she obtained a doctoral degree in computer linguistics on \BiographIE Classi cation and Extraction of Career-Speci c Information".
Frederik Baumer. Dr. Frederik Baumer has been working as a research associate at the Semantic Information Processing group in Paderborn since completing his master's degree in \Management Information Systems" with a focus on Semantic Information Processing. Since July 2013, he has been involved in various research projects at the Heinz Nixdorf Institute { initially as a master's student and then as doctoral student { and is an expert for search engine technology and information extraction. While participating in the Collaborative Research Center 901 \On-the-Fly Computing", he has been working on his doctorate since October 2014. Dr. Baumer successfully defended his PhD thesis in July 2017 on \Indicator-based Detection and Compensation of Inaccurate and Incompletely Described Software Requirements".
At the beginning of our work, we dealt in particular with the question of the characteristics of NL requirements descriptions [GSB15]. It quickly became clear that a major challenge is o -topic information, incompleteness and ambiguity { problems that have already been subject of research in this area for a long time. For the automatic composition of software services, it is particularly crucial because in OTF Computing, it is better assumed that high-quality, formal speci cations are provided as input, which are qualitatively better than NL requirements. There was, at this point, no automated procedure for checking and compensating requirements.
As a rst challenge, we understood how to separate relevant information from irrelevant and how to identify the core components of requirements [BDG17, DG16]. After some attempts with rule-based approaches, we decided to go for a machine-learning approach because the text quality of user-generated requirements is very poor and we achieved only a bad recall through using rule-based approaches [DG16]. We then focused on the language de cits mentioned before. They are di cult to recognize and to compensate for end-users and sometimes cause great damage in the composition process [Bau17]. However, each language de cit has di erent characteristics and must be recognized and compensated di erently. In the case of incompleteness, we focused on the predicate-argument structure in requirements descriptions and tried to identify missing arguments and thus incompleteness [BG16]. For this procedure, extensive resources are necessary, which did not exist so far and were created by ourselves. Since there are hardly any real-life requirements to nd, we have instead used software descriptions and feature requests from app stores (e.g., download.com) and developer forums (e.g., sourceforge.com) [BDG17, DG16].
There is a lot of work done on ambiguity in requirement descriptions. We distinguish between lexical, syntactic and referential ambiguity. All forms require di erent detection and compensation methods [Bau17, GB17, GSB15]. Even more di cult to detect is vagueness. Here we have developed a rule-based approach that can recognize some forms of vagueness, but not all [GB17]. Since the compensation of vagueness is more di cult because of their di erent characteristics compared with ambiguities, we introduce a dialogue with the user (cf. Section 4). Based on this research and the realization that linguistic de cits are so di erent, we focused on a strategic approach for the optimization of NL: The goal was to develop a parameterized model (so-called strategy con guration) that automatically selects the best tting strategy to compensate de ciencies in software speci cations [Bau17]. In order to enable this, suitable methods for the recognition and resolution of lexical, syntactic, and referential ambiguity and for the completion of NL requirements were rst identi ed in literature and then implemented. This is the starting point of the indicator-based strategies (cf. Figure 2).
Requirements Description
Indicators
Strategy index Indicator-based
Con guration Preprocessing and
compensation
These strategies enable CORDULA to execute and monitor individual methods, to compare results and to correct them if necessary as well as to use synergy e ects. For this purpose, it was necessary to de ne contextsensitive indicators to determine the compensation need, which can trigger individual strategies. Employing these strategies, CORDULA is able to apply existing, highly heterogeneous processing and compensation techniques on requirements descriptions. It is only possible by means of the linguistic indicators to capture and optimize the individual text quality in a data-driven manner, as required, because CORDULA's compensation pipeline is con gured on the basis of the detected shortcomings. In addition, the individual results of the components are merged and structured as a consistent compensation result. 4
As mentioned before, language de cits in NL requirements can only be compensated automatically in parts. A complete compensation without any questions to the users is in many cases not possible, especially if vagueness and incompleteness are present. For this reason, we are currently working to extend CORDULA with a chat functionality. The users can submit their requirements to CORDULA via chat (left part of Figure 3) and receive an overview of the processed requirements as return (right part of Figure 3). If CORDULA is able to detect but not to compensate a problem, the system automatically asks the user to provide some helpful information or to make a decision (e.g., annotate (a part of) a sentence as an o -topic, as shown in Figure 3).
First attempts with existing chatbot frameworks such as Google's api.ai or IBM Watson Conversations worked well and brought up interesting ndings. They pointed out that we need more exibility and control in the conversation and less free communication (which the framework aims for). The dialogue should not be conducted completely by the chatbot, but should address a speci c problem and provide targeted information for CORDULA. For this reason, in our current vision we use our own chat solution, which can ask prede ned questions to the user about prede ned problems. In this way, we expect to get further relevant user input for a speci c problem just-in-time. One criticism of this bidirectional chat approach remains, however: chatting slows down the entire OTF service composition process. Nevertheless, an improvement in the overall result can be expected, since the requirements will be more concrete and complete. This trade o needs to be discussed in future work. Acknowledgements This work was partially supported by the German Research Foundation (DFG) within the Collaborative Research Center \On-The-Fly Computing" (SFB 901).