Systems such as Alexa, Cortana, and Siri appear rather smart.
However, they only react to predefined wordings and do not actually
grasp the user’s intent. To overcome this limitation, a system must understand the topics the user is talking about. Therefore, we apply unsupervised multi-topic labeling to spoken utterances. Although topic labeling is a well-studied task on textual documents, its potential for spoken
input is almost unexplored. Our approach for topic labeling is tailored
to spoken utterances; it copes with short and ungrammatical input.
The approach is two-tiered. First, we disambiguate word senses. We utilize Wikipedia as pre-labeled corpus to train a naïve-bayes classifier.
Second, we build topic graphs based on DBpedia relations. We use two
strategies to determine central terms in the graphs, i.e. the shared topics. One focuses on the dominant senses in the utterance and the other
covers as many distinct senses as possible. Our approach creates multiple
distinct topics per utterance and ranks results.
The evaluation shows that the approach is feasible; the word sense disambiguation achieves a recall of 0.799. Concerning topic labeling, in a user
study subjects assessed that in 90.9% of the cases at least one proposed
topic label among the first four is a good fit. With regard to precision,
the subjects judged that 77.2% of the top ranked labels are a good fit or
good but somewhat too broad (Fleiss’ kappa κ = 0.27).
We illustrate areas of application of topic labeling in the field of programming in spoken language. With topic labeling applied to the spoken
input as well as ontologies that model the situational context we are able
to select the most appropriate ontologies with an F1-score of 0.907.
Large-Scale Dependency Knowledge Acquisition and its Extrinsic Evaluation Through Word Sense Disambiguation
