Neural Wikipedian: Generating Textual Summaries from Knowledge Base Triples

TL;DR

This work addresses generating natural language summaries from Semantic Web knowledge-base triples using an end-to-end neural encoder-decoder framework. It introduces a feed-forward triple encoder and an RNN-based decoder that generates text conditioned on the encoded triples, trained on large, automatically aligned Wikipedia-DBpedia/Wikidata corpora. The authors propose data preparation techniques, including a multi-placeholder approach to verbalize rare entities and two decoding variants (URIs vs surface forms), and demonstrate improvements in perplexity, BLEU, and ROUGE, complemented by human judgments. The approach scales to large vocabularies and offers a template-free alternative for data-to-text generation in open-domain semantic-web contexts, with potential impact on accessibility and downstream NLP tasks.

Abstract

Most people do not interact with Semantic Web data directly. Unless they have the expertise to understand the underlying technology, they need textual or visual interfaces to help them make sense of it. We explore the problem of generating natural language summaries for Semantic Web data. This is non-trivial, especially in an open-domain context. To address this problem, we explore the use of neural networks. Our system encodes the information from a set of triples into a vector of fixed dimensionality and generates a textual summary by conditioning the output on the encoded vector. We train and evaluate our models on two corpora of loosely aligned Wikipedia snippets and DBpedia and Wikidata triples with promising results.

Figures (3)

• Figure 1: The triple encoder computes a vector representation for each one of the two input triples $h_{f_1}$ and $h_{f_2}$. Subsequently, the decoder uses the concatenation of the two vectors, $[h_{f_1};h_{f_2}]$ to initialise the decoding process that generates the summary word per word. Each textual summary starts and ends with the respective start-of-sequence <start> and end-of-sequence <end> tokens.
• Figure 2: An idealised example of a beam-search decoder with a beam $B$ of size $2$ and target vocabulary size $|X|$ equal to $9$. The scores at the right-hand side of the words in the vocabulary is the probability of the summary when it is extended by that particular word.
• Figure 3: Performance of our models with the BLEU $4$ metric across the difference number of input triples on DBpedia (a) and Wikidata (b).