Frederik Simon Bäumer

What Did You Mean

Existing approaches towards service composition demand requirements of the customers in terms of service templates, service query profiles, or partial process models. However, addressed non-expert customers may be unable to fill-in the slots of service templates as requested or to describe, for example, pre- and postconditions, or even have difficulties in formalizing their requirements. Thus, our idea is to provide nonexperts with suggestions how to complete or clarify their requirement descriptions written in natural language. Two main issues have to be tackled: (1) partial or full inability (incapacity) of non-experts to specify their requirements correctly in formal and precise ways, and (2) problems in text analysis due to fuzziness in natural language. We present ideas how to face these challenges by means of requirement disambiguation and completion. Therefore, we conduct ontology-based requirement extraction and similarity retrieval based on requirement descriptions that are gathered from App marketplaces. The innovative aspect of our work is that we support users without expert knowledge in writing their requirements by simultaneously resolving ambiguity, vagueness, and underspecification in natural language.

Studying software descriptions in SourceForge and app stores for a better understanding of real-life requirements

Users prefer natural language software requirements because of their usability and accessibility. Many approaches exist to elaborate these requirements and to support the users during the elicitation process. But there is a lack of adequate resources, which are needed to train and evaluate approaches for requirement refinement. We are trying to close this gap by using online available software descriptions from SourceForge and app stores. Thus, we present two real-life requirements collections based on online-available software descriptions. Our goal is to show the domain-specific characteristics of content words describing functional requirements. On the one hand, we created a semantic role-labeled requirements set, which we use for requirements classification. On the other hand, we enriched software descriptions with linguistic features and dependencies to provide evidence for the context-awareness of software functionalities.

How to Complete Customer Requirements

One purpose of requirement refinement is that higher-level requirements have to be translated to something usable by developers. Since customer requirements are often written in natural language by end users, they lack precision, completeness and consistency. Although user stories are often used in the requirement elicitation process in order to describe the possibilities how to interact with the software, there is always something unspoken. Here, we present techniques how to automatically refine vague software descriptions. Thus, we can bridge the gap by first revising natural language utterances from higher-level to more detailed customer requirements, before functionality matters. We therefore focus on the resolution of semantically incomplete user-generated sentences (i.e. non-instantiated arguments of predicates) and provide ontology-based gap-filling suggestions how to complete unverbalized information in the user’s demand.

A System for Uncovering Latent Connectivity of Health Care Providers in Online Reviews

The contacts a health care provider (HCP), like a physician, has to other HCPs is perceived as a quality characteristic by patients. So far, only the German physician rating website jameda.de gives information about the interconnectedness of HCPs in business networks. However, this network has to be maintained manually and is thus incomplete. We therefore developed a system for uncovering latent connectivity of HCPs in online reviews to provide users with more valuable information about their HCPs. The overall goal of this approach is to extend already existing business networks of HCPs by integrating connections that are newly discovered by our system. Our most recent evaluation results are promising: 70.8 % of the connections extracted from the reviews texts were correctly identified and in total 3,788 relations were recognized that have not been displayed in jameda.de’s network before.

Filtering Reviews by Random Individual Error

Opinion mining from physician rating websites depends on the quality of the extracted information. Sometimes reviews are user-error prone and the assigned stars or grades contradict the associated content. We therefore aim at detecting random individual error within reviews. Such errors comprise the disagreement in polarity of review texts and the respective ratings. The challenges that thereby arise are 1 the content and sentiment analysis of the review texts and 2 the removal of the random individual errors contained therein. To solve these tasks, we assign polarities to automatically recognized opinion phrases in reviews and then check for divergence in rating and text polarity. The novelty of our approach is that we improve user-generated data quality by excluding error-prone reviews on German physician websites from average ratings.

Privacy Matters: Detecting Nocuous Patient Data Exposure in Online Physician Reviews

Consulting a physician was long regarded as an intimate and private matter. The physician-patient relationship was perceived as sensitive and trustful. Nowadays, there is a change, as medical procedures and physicians consultations are reviewed like other services on the Internet. To allay user’s privacy doubts, physician review websites assure anonymity and the protection of private data. However, there are hundreds of reviews that reveal private information and hence enable physicians or the public to identify patients. Thus, we draw attention to the cases when de-anonymization is possible. We therefore introduce an approach that highlights private information in physician reviews for users to avoid an accidental disclosure. For this reason, we combine established natural-language-processing techniques such as named entity recognition as well as handcrafted patterns to achieve a high detection accuracy. That way, we can help websites to increase privacy protection by recognizing and uncovering apparently uncritical information in user-generated texts.

I grade what I get but write what I think. Inconsistency Analysis in Patients' Reviews

Received medical services are increasingly discussed and recommended on physician rating websites (PRWs). The reviews and ratings on these platforms are valuable sources of information for patient opinion mining. In this paper, we have tackled three issues that come along with inconsistency analysis on PRWs: (1) Natural language processing of user-generated reviews, (2) the disagreement in polarity of review text and its corresponding numerical ratings (individual inconsistency) and (3) the differences in patients’ rating behavior for the same service category (e.g. ‘treatment’) expressed by varying grades on the entire data set (collective inconsistency). Thus, the basic idea is first to identify relevant opinion phrases that describe service categories and to determine their polarity. Subsequently, the particular phrase has to be assigned to its corresponding numerical rating category before checking the (dis-)agreement of polarity values. For this purpose, several local grammars for the pattern-based analysis as well as domain-specific dictionaries for the recognition of entities, aspects and polarity were applied on 593,633 physician reviews from both German PRWs jameda.de and docinsider.de. Furthermore, our research contributes to content quality improvement of PRWs because we provide a technique to detect inconsistent reviews that could be ignored for the computation of average ratings.

Find a Physician by Matching Medical Needs Described in your Own Words

The individual search for information about physicians on Web 2.0 platforms can affect almost all aspects of our lives. People can directly access physician rating websites via web browsers or use any search engine to find physician reviews and ratings filtered by location resp. specialty. However, sometimes keyword search does not meet user needs because of the disagreement of users’ common terms queries for symptoms and the widespread medical terminology. In this paper, we present the prototype of a specialised search engine that overcomes this by indexing user-generated content (i.e., review texts) for physician discovery and provides automatic suggestions as well as an appropriate visualisation. On the one hand, we consider the available numeric physician ratings as sorting criterion for the ranking of query results. Furthermore, we extended existing ranking algorithms with respect to domain-specific types and physicians ratings on the other hand. We gathered more than 860,000 review texts and collected more than 213,000 physician records. A random test shows that about 19.7% of 5,100 different words in total are health- related and partly belong to consumer health vocabularies. Our evaluation results show that the query results fit users particular health issues when seeking for physicians.

Understanding the Patient 2.0

Patients 2.0 increasingly inform themselves about the quality of medical services on physician rating websites. However, little is known about whether the reviews and ratings on these websites truly reflect the quality of services or whether the ratings on these websites are rather influenced by patients’ individual rating behavior. Therefore, we investigate more than 790,000 physician reviews on Germany’s most used physician rating website jameda.de. Our results show that patients’ ratings do not only reflect treatment quality but are also influenced by treatment quality independent factors like age and complaint behavior. Hence, we provide evidence that users should be well aware of user specific rating distortions when intending to make their physician choice based on these ratings.

NLP in OTF Computing: Current Approaches and Open Challenges

On-The-Fly Computing is the vision of covering software needs of end users by fully-automatic compositions of existing software services. End users will receive so-called service compositions tailored to their very individual needs, based on natural language software descriptions. This everyday language may contain inaccuracies and incompleteness, which are well-known challenges in requirements engineering. In addition to existing approaches that try to automatically identify and correct these deficits, there are also new trends to involve users more in the elaboration and refinement process. In this paper, we present the relevant state of the art in the field of automated detection and compensation of multiple inaccuracies in natural language service descriptions and name open challenges needed to be tackled in NL-based software service composition.