Ambient Search: A Document Retrieval System for Speech Streams

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 2082–2091, Osaka, Japan, December 11-17 2016. Ambient Search: A Document Retrieval System for Speech Streams Benjamin Milde1,2, Jonas Wacker1, Stefan Radomski2, Max M¨uhlh¨auser2, and Chris Biemann1 1 Language Technology Group / 2 Telecooperation Group Computer Science Department Technische Universit¨at Darmstadt, Germany Abstract We present Ambient Search, an open source system for displaying and retrieving relevant doc- uments in real time for speech input. The system works ambiently, that is, it unobstructively listens to speech streams in the background, identifies keywords and keyphrases for query con- struction and continuously serves relevant documents from its index. Query terms are ranked with Word2Vec and TF-IDF and are continuously updated to allow for ongoing querying of a document collection. The retrieved documents, in our case Wikipedia articles, are visualized in real time in a browser interface. Our evaluation shows that Ambient Search compares favorably to another implicit information retrieval system on speech streams. Furthermore, we extrinsically evaluate multiword keyphrase generation, showing positive impact for manual transcriptions. 1 Introduction Recent advancements in Automated Speech Recognition (ASR) and Natural Language Understanding (NLU) have proliferated the use of personal assistants like Siri1 or Google Now2, with which people interact naturally with their voice. However, the activation of such systems has to be specifically triggered and they are targeted to an (ever–growing) set of anticipated question types and commands. When taking part in a conversation or listening to a lecture, people may want to look up helpful in- formation or check facts. Manually checking this information or interacting with a personal assistant would hamper the flow of the discussion, respectively distract from the lecture. In the following, we present Ambient Search, a system that ambiently researches relevant information, in the form of propos- ing relevant documents to users in conversations or users who passively listen to spoken language. In contrast to other personal assistants, our system is not specifically triggered, it unobtrusively listens to speech streams in the background and implicitly queries an index of documents. We see the following utility in our approach: The assistant stays in the background and does not disturb the user. Access to the displayed snippets is on demand and the user can access the information in context without the need to formulate a specific query. On the other hand, these advantages are fundamentally based on how well the system is able to retrieve relevant documents, as the system’s utility diminishes when proposing a lot of irrelevant documents. In this paper, we also evaluate how well the system is able to retrieve relevant Wikipedia articles in spite of average speech recognition word error rates (WER) of 15.6% on TED talks and show that it finds more relevant articles compared to another implicit information retrieval system on speech streams. The next section discusses related research, while we give an overview and technical details of our ap- proach in Section 3. We evaluate keyword recognition and retrieval relevance in Section 4, …

Content truncated to 500 words