Discovering Hypernymy in Text-Rich Heterogeneous Information Network by Exploiting Context Granularity

Abstract

Text-rich heterogeneous information networks (text-rich HINs) are ubiquitous in real-world applications. Hypernymy, also known as is-a relation or subclass-of relation, lays in the core of many knowledge graphs and benefits many downstream applications. Existing methods of hypernymy discovery either leverage textual patterns to extract explicitly mentioned hypernym-hyponym pairs, or learn a distributional representation for each term of interest based its context. These approaches rely on statistical signals from the textual corpus, and their effectiveness would therefore be hindered when the signals from the corpus are not sufficient for all terms of interest. In this work, we propose to discover hypernymy in text-rich HINs, which can introduce additional high-quality signals. We develop a new framework, named HyperMine, that exploits multi-granular contexts and combines signals from both text and network without human labeled data. HyperMine extends the definition of "context" to the scenario of text-rich HIN. For example, we can define typed nodes and communities as contexts. These contexts encode signals of different granularities and we feed them into a hypernymy inference model. HyperMine learns this model using weak supervision acquired based on high-precision textual patterns. Extensive experiments on two large real-world datasets demonstrate the effectiveness of HyperMine and the utility of modeling context granularity. We further show a case study that a high-quality taxonomy can be generated solely based on the hypernymy discovered by HyperMine.

MATCH: Metadata-Aware Text Classification in A Large Hierarchy

“…Note that one can interpret DIH in various ways depending on how "contexts" are defined. For example, if "contexts" are defined as documents [45], then DIH states that: if a word (e.g., "Crawling") appears in a document, then its parent (e.g., "World Wide Web") is also expected to be in that document. In contrast, if "contexts" are defined based on the local context window (i.e., the previous and the latter words in a sequence) [46], then DIH becomes: if a context word 𝑐 occurs 𝑛 times in the context window of a child 𝑤 1 , then it is expected to occur no less than 𝑛 times in the context window of its parent 𝑤 2 .…”

“…In contrast, if "contexts" are defined based on the local context window (i.e., the previous and the latter words in a sequence) [46], then DIH becomes: if a context word 𝑐 occurs 𝑛 times in the context window of a child 𝑤 1 , then it is expected to occur no less than 𝑛 times in the context window of its parent 𝑤 2 . DIH is a classic tool in constructing topic taxonomies [44,45], which motivates us to propose the following DIH-based regularization.…”

STEAM: Self-Supervised Taxonomy Expansion with Mini-Paths

“…word embeddings [9,27,31]) of terms. For a term pair ⟨x, y⟩, their embeddings are used for learning a binary classifier to predict whether it has the hypernymy relation [4,7,12,37]. As embeddings are directly learned from the corpora, distributed methods eliminate the needs of designing hand-crafted patterns and have shown strong performance.…”

Fully-Unsupervised Embeddings-Based Hypernym Discovery

“…CRIM [32] combines projection learning, trying to achieve a linear transformation which goes from hyponyms to hypernyms, with pattern-based extraction. It is also worth mentioning that recognising hypernymy represents a fundamental feature in taxonomy induction and enrichment [18,[33][34][35]. When evaluating our approach we compare ourselves with all the systems who took part in SemEval 2018 Hypernym Discovery task.…”