A Self-Aggregated Hierarchical Topic Model for Short Texts

With the growth of the internet, short texts such as tweets from Twitter, news titles from the RSS, or comments from Amazon have become very prevalent. Many tasks need to retrieve information hidden from the content of short texts. So ontology learning methods are proposed for retrieving structured information. Topic hierarchy is a typical ontology that consists of concepts and taxonomy relations between concepts. Current hierarchical topic models are not specially designed for short texts. These methods use word co-occurrence to construct concepts and general-special word relations to construct taxonomy topics. But in short texts, word cooccurrence is sparse and lacking general-special word relations. To overcome this two problems and provide an interpretable result, we designed a hierarchical topic model which aggregates short texts into long documents and constructing topics and relations. Because long documents add additional semantic information, our model can avoid the sparsity of word cooccurrence. In experiments, we measured the quality of concepts by topic coherence metric on four real-world short texts corpus. The result showed that our topic hierarchy is more interpretable than other methods.

A Knowledge Discovery Method for Landslide Monitoring Based on K-core Decomposition and the Louvain Algorithm

Landslide monitoring plays an important role in predicting, forecasting and preventing landslides. Quantitative explorations at the subject level and fine-scale knowledge in landslide monitoring research can be used to provide information and references for landslide monitoring status analysis and disaster management. In the context of the large amount of keyword co-occurrence network information, it is difficult to clearly determine and display the domain topic hierarchy and knowledge structure. This paper proposes a landslide monitoring knowledge discovery method that combines the K-core decomposition and Louvain algorithms. In this method, author keywords from the literature are used as nodes to construct a weighted co-occurrence network, and a pruning standard value is defined for K. The K-core approach is used to decompose the network into subgraphs. Combined with the unsupervised Louvain algorithm, subgraphs are divided into different topic communities by setting a modularity change threshold, which is used to establish a topic hierarchy and identify fine-scale knowledge related to landslide monitoring. Based on the Web of Science, a comparative experiment involving the above method and a high-frequency keyword subgraph method for landslide monitoring knowledge discovery is performed. In the resulting 5-core network subgraph of landslide monitoring keyword co-occurrence, 17 community structures can be identified, and the degree value and density of subcommunities are analysed by taking the community with the largest proportion of nodes as an example. The results show that the retention time of the proposed method is significantly lower than that of the traditional method.

DeepTileBars: Visualizing Term Distribution for Neural Information Retrieval

Most neural Information Retrieval (Neu-IR) models derive query-to-document ranking scores based on term-level matching. Inspired by TileBars, a classical term distribution visualization method, in this paper, we propose a novel Neu-IR model that handles query-to-document matching at the subtopic and higher levels. Our system first splits the documents into topical segments, “visualizes” the matchings between the query and the segments, and then feeds an interaction matrix into a Neu-IR model, DeepTileBars, to obtain the final ranking scores. DeepTileBars models the relevance signals occurring at different granularities in a document’s topic hierarchy. It better captures the discourse structure of a document and thus the matching patterns. Although its design and implementation are light-weight, DeepTileBars outperforms other state-of-the-art Neu-IR models on benchmark datasets including the Text REtrieval Conference (TREC) 2010-2012 Web Tracks and LETOR 4.0.