An analysis of textual inference in German customer emails

Proceedings of the Third Joint Conference on Lexical and Computational Semantics (*SEM 2014), pages 69–74, Dublin, Ireland, August 23-24 2014. An analysis of textual inference in German customer emails Kathrin Eichler∗, Aleksandra Gabryszak∗, G¨unter Neumann† ∗German Research Center for Artificial Intelligence (DFKI), Berlin (kathrin.eichler|aleksandra.gabryszak@dfki.de) †German Research Center for Artificial Intelligence (DFKI), Saarbr¨ucken (neumann@dfki.de) Abstract Human language allows us to express the same meaning in various ways. Recogniz- ing that the meaning of one text can be in- ferred from the meaning of another can be of help in many natural language process- ing applications. One such application is the categorization of emails. In this paper, we describe the analysis of a real-world dataset of manually categorized customer emails written in the German language. We investigate the nature of textual infer- ence in this data, laying the ground for de- veloping an inference-based email catego- rization system. This is the first analysis of this kind on German data. We compare our results to previous analyses on English data and present major differences. 1 Introduction A typical situation in customer support is that many customers send requests describing the same issue. Recognizing that two different customer emails refer to the same problem can help save resources, but can turn out to be a difficult task. Customer requests are usually written in the form of unstructured natural language text, i.e., when automatically processing them, we are faced with the issue of variability: Different speakers of a lan- guage express the same meanings using different linguistic forms. There are, in fact, cases where two sentences expressing the same meaning do not share a single word: 1. “Bild und Ton sind asynchron.” [Picture and sound are asynchronous.] 2. “Die Tonspur stimmt nicht mit dem Film ¨uberein.” [The audio track does not match the video.] Detecting the semantic equivalence of sentences 1 and 2 requires several textual inference steps: At the lexical level, it requires mapping the word pic- ture to video and sound to audio track. At the level of compositional semantics, it requires de- tecting the equivalence of the expressions A and B are asynchronous and A does not match B. In this paper, we describe our analysis of a large set of manually categorized customer emails, lay- ing the ground for developing an email catego- rization system based on textual inference. In our analysis, we compared each email text to the description of its associated category in order to investigate the nature of the inference steps in- volved. In particular, our analysis aims to give an- swers to the following questions: What text repre- sentation is appropriate for the email categoriza- tion task? What kind of inference steps are in- volved and how are they distributed in real-world data? Answering these questions will not only help us decide, which existing tools and resources to integrate in an inference-based email catego- rization system, but also, which non-existing tools may be needed in addition. 2 Related Work The task of email categorization has been ad- …

Content truncated to 500 words