Point at the Triple: Generation of Text Summaries from Knowledge Base Triples

Abstract

We investigate the problem of generating natural language summaries from knowledge base triples. Our approach is based on a pointer-generator network, which, in addition to generating regular words from a fixed target vocabulary, is able to verbalise triples in several ways. We undertake an automatic and a human evaluation on single and open-domain summaries generation tasks. Both show that our approach significantly outperforms other data-driven baselines.

Using natural language generation to bootstrap missing Wikipedia articles: A human-centric perspective

“…In all these examples, linguistic information from the knowledge base is used to build a parallel corpus containing triples and equivalent text sentences from Wikipedia, which is then used to train the NLG algorithm. Directly relevant to the model we propose are the proposals by Lebret et al [51], Chisholm et al [11], Liu et al [55], Yeh et al [91] and Vougiouklis et al [84,85], which extend the general encoder-decoder neural network framework from [12,82] to generate short summaries in English. The original of English biographies generation was introduced by Lebret et al [51] who used feed-forward language model with slot-value templates to generate the first sentence of a Wikipedia summary from its corresponding infobox.…”

“…The original of English biographies generation was introduced by Lebret et al [51] who used feed-forward language model with slot-value templates to generate the first sentence of a Wikipedia summary from its corresponding infobox. Incremental upgrades of the original architecture on the same task include the introduction of an auto-encoding pipeline based on an attentive encoder-decoder architecture using GRUs [11], a novel double-attention mechanism over the input infoboxes' fields and their values [55], and adaptations of pointergenerator mechanisms [85,91] over the input triples.…”

“…All these approaches use structured data from Freebase, Wikidata and DBpedia as input and generate summaries consisting either by one or two sentences that match the style of the English Wikipedia in a single domain [11,51,55,84,91] or more recently in more open-domain scenarios [85]. While this is a rather narrow task compared to other generative tasks such as translation, Chisholm et al [11] discuss its challenges in detail and show that it is far from being solved.…”

WDV: A Broad Data Verbalisation Dataset Built from Wikidata

“…More recent datasets attempt to remedy some of these limitations. Pavlos et al [38,39] propose two large corpora that align Wikidata and DBpedia claims to Wikipedia text. However, they focus on verbalisations of multiple claims at a time, which limits its usefulness for important tasks e.g.…”

“…Wikidata, the web's largest collaborative KG, has very few such datasets [39,7], and existing ones rely on distant supervision to prioritise the sheer number of couplings in exchange for coupling tightness. In addition, they disproportionately represent specific entity types from Wikidata, such as people and locations, when Wikidata covers a much wider variety of information.…”