We address the task of unsupervised Semantic Textual Similarity (STS) by ensembling diverse pre-trained sentence encoders into sentence meta-embeddings.We apply, extend and evaluate different meta-embedding methods from the word embedding literature at the sentence level, including dimensionality reduction (Yin and Schütze, 2016), generalized Canonical Correlation Analysis (Rastogi et al., 2015) and cross-view auto-encoders (Bollegala and Bao, 2018).Our sentence metaembeddings set a new unsupervised State of The Art (SoTA) on the STS Benchmark and on the STS12-STS16 datasets, with gains of between 3.7% and 6.4% Pearson's r over singlesource systems.

This paper describes USI Answers a natural language question answering system for semi-structured industry data. The paper reports on the progress towards the goal of offering easy access to enterprise data to a large number of business users, most of whom are not familiar with the specific syntax or semantics of the underlying data sources. Additional complications come from the nature of the data, which comes both as structured and unstructured. The proposed solution allows users to express questions in natural language, makes apparent the system’s interpretation of the query, and allows easy query adjustment and reformulation. The application is in use by more than 1500 users from Siemens Energy. We evaluate our approach on a data set consisting of fleet data.

This paper is concerned with the use of conversational agents as an interaction paradigm for accessing open domain encyclopedic knowledge by means of Wikipedia. More precisely, we describe a dialog-based question answering system for German which utilizes Wikipedia-based topic models as a reference point for context detection and answer prediction. We investigate two different perspectives to the task of interfacing virtual agents with collaborative knowledge. First, we exploit the use of Wikipedia categories as a basis for identifying the broader topic of a spoken utterance. Second, we describe how to enhance the conversational behavior of the virtual agent by means of a Wikipedia-based question answering component which incorporates the question topic. At large, our approach identifies topic-related focus terms of a user’s question, which are subsequently mapped onto a category taxonomy. Thus, we utilize the taxonomy as a reference point to derive topic labels for a user’s question. The employed topic model is thereby based on explicitly given concepts as represented by the document and category structure of the Wikipedia knowledge base. Identified topic categories are subsequently combined with different linguistic filtering methods to improve answer candidate retrieval and reranking. Results show that the topic model approach contributes to an enhancement of the conversational behavior of virtual agents.

This paper introduces a model harvesting the crowd-sourced encyclopedic knowledge provided by Wikipedia to improve the conversational abilities of an artificial agent. More precisely, we present a model for automatic topic identification in ongoing natural language dialogs. On the basis of a graph-based representation of the Wikipedia category system, our model implements six tasks essential for detecting the topical overlap of coherent dialog contributions. Thereby the identification process operates online to handle dialog streams of constantly changing topical threads in real-time. The realization of the model and its application to our conversational agent aims to improve human-agent conversations by transferring human-like topic awareness to the artificial interlocutor.

This paper presents an approach using social semantics for the task of topic labelling by means of Open Topic Models. Our approach utilizes a social ontology to create an alignment of documents within a social network. Comprised category information is used to compute a topic generalization. We propose a feature-frequency-based method for measuring semantic relatedness which is needed in order to reduce the number of document features for the task of topic labelling. This method is evaluated against multiple human judgement experiments comprising two languages and three different resources. Overall the results show that social ontologies provide a rich source of terminological knowledge. The performance of the semantic relatedness measure with correlation values of up to .77 are quite promising. Results on the topic labelling experiment show, with an accuracy of up to .79, that our approach can be a valuable method for various NLP applications.

This paper presents an approach for predicting context sensitive entities exemplified in the domain of person names. Our approach is based on building a weighted context but also a weighted people graph and predicting the context entity by extracting the best fitting sub graph using a spreading activation technique. The results of the experiments show a quite promising F-Measure of 0.99.