Semantic Methods for Computer-supported Writing Aids

语言技术领域的这一研究方案关注的是是否有可能开发一种语义写作辅助工具，帮助通过释义来重新组织语篇。这类似于拼写更正或语法更正：在文本处理上下文中，提供了适当的释义，从而允许更快地制定具有更多可变词汇选择的文本。通过研究一种通过使用数据提高释义质量(并以这种方式辅助写作)的机制，给出了一个特殊的特征。该建议的主要假设是，我们假设无监督和无知识的方法可以为这一应用上下文的释义产生合适的数据源。在这个写作辅助工具的原型实现的核心中可以找到一个释义组件。除了关于几个数据源的组合和关于合适的用户界面的研究的问题外，我们还将致力于将方法转移到使用数据驱动的方法的其他语言。此外，我们还探讨了利用内隐反馈改进写作辅助手段的可能性。对于单个组件的开发，以及模拟使用，我们在很大程度上依赖于众包作为一种数据收集和评估的手段。首先，我们探索如何对来自不同数据源的释义进行表征，并将这些不同来源的释义组合成一个释义组件。我们将分布语义方法与语境联系起来，用无监督和无知识的方法产生上下文相关的释义候选词。在这里，我们特别关注这种方法的数据驱动性，理想情况下，这种方法在没有任何现有词法资源的情况下应该可以工作。这是由语言和领域独立性所推动的，并将通过将方法从英语转移到德语来展示。在用户研究的支持下，我们实现了一个写作辅助工具的原型，并优化了用户交互、演示和主动性方面的解决方案。我们使用这个原型来检查，在多大程度上我们可以使用‘弱信号’，即仅仅是与原型的交互数据，来改进释义部分。这种以前未在语言技术中使用的隐式用户反馈形式引起了语言处理组件的迭代改进，以便将预处理步骤分割到应用程序所需的质量级别。

This research proposal in the field of language technology is concerned with the question whether it is possible to develop a semantic writing aid, which helps reformulating texts by paraphrasing. This works similar to a spelling correction or grammar correction: in a text processing context, suitable paraphrases are offered, which allows faster formulation of texts with a more variable vocabulary choice. A special feature is given by researching a mechanism that improves paraphrasing quality (and in this way the writing aid) by usage data. The main hypothesis in this proposal is that we assume that unsupervised and knowledge-free methods can yield suitable data sources for paraphrases for this application context. A paraphrasing component is found at the core of a prototypical implementation of this writing aid. Additionally to questions regarding the combination of several data sources and research regarding a suitable user interface, we will work on transferring the methodology to other languages using data-driven methods. Further, we explore the possibility of improving the writing aid with implicit feedback. For the development of single components, as well as for simulating usage, we massively rely on crowdsourcing as a means to data collection and for evaluation. First we explore, how paraphrases from different data sources are characterized and combine these heterogeneous sources in a paraphrasing component. We contextualize distributional semantic methods to produce context-dependent paraphrase candidates wit unsupervised and knowledge-free methods. Here, we especially focus on the data-drivenness of this approach, which should ideally work without any existing lexical resources. This is motivated by language and domain independence and will be demonstrated by transferring the methodology from English to German. Supported by user studies, we implement a prototype of the writing aid and optimize the solution regarding user interaction, presentation and pro-activity. We use this prototype to examine, in how far we can use 'weak signals', i.e. merely interaction data with the prototype, to improve the paraphrasing component. This form of implicit user feedback, that has not been utilized in language technology before, gives rise to the iterative refinement of language processing components in order to segue pre-processing steps to the quality level required by applications.