Source-linked AI summary
Learning to Generate Wikipedia Summaries for Underserved Languages from Wikidata
Lucie-Aimée Kaffee, Hady Elsahar, Pavlos Vougiouklis, Christophe Gravier, Frédérique Laforest, Jonathon Hare, Elena Simperl
TL;DR
Wikipedia content is unevenly distributed, and editors in under-resourced communities cannot curate the article volume available in larger language communities. An end-to-end neural model generates summaries from Wikidata triples, using property placeholders to copy entity labels from the input. At least 5.25 BLEU 4 in Arabic and 1.31 BLEU 4 in Esperanto over IRext, the strongest baseline; the model outperforms all baselines on all metrics in both languages.
Problem
Wikipedia content is unevenly distributed, and editors in under-resourced communities cannot curate the article volume available in larger language communities.
Source: §1 · p. 1Method
An end-to-end neural model generates summaries from Wikidata triples, using property placeholders to copy entity labels from the input.
Source: §1 · p. 1, §2.3 · p. 3Results
Takeaways & Limitations
Structured knowledge-base data and careful multilingual preparation offer a practical route to underserved-language summaries that otherwise would require substantially more data.
Source: §6 · p. 5Around ≤1% of Arabic and 2% of Esperanto input validation and test triple sets led to generated summaries without the <end> token.
Source: §5 · p. 4
Abstract
from arXiv · showhide
While Wikipedia exists in 287 languages, its content is unevenly distributed among them. In this work, we investigate the generation of open domain Wikipedia summaries in underserved languages using structured data from Wikidata. To this end, we propose a neural network architecture equipped with copy actions that learns to generate single-sentence and comprehensible textual summaries from Wikidata triples. We demonstrate the effectiveness of the proposed approach by evaluating it against a set of baselines on two languages of different natures: Arabic, a morphological rich language with a larger vocabulary than English, and Esperanto, a constructed language known for its easy acquisition.
1 School of Electronics and Computer Science, University of Southampton, UK
The passage lists contact addresses for ecs.soton.ac.uk and univ-st-etienne.fr, alongside French institutional names.
1 School of Electronics and Computer Science, University of Southampton, UK
- The listed French institutions are Laboratoire Hubert Curien, CNRS, UJM-Saint-Étienne, and Université de Lyon.The passage also lists contact addresses for ecs.soton.ac.uk and univ-st-etienne.fr.Source: §1 · p. 1
1 Introduction
Wikipedia content is unevenly distributed, and editors in under-resourced communities cannot curate the article volume available in larger language communities. The paper proposes generating summaries from Wikidata triples and evaluates an end-to-end model with copy actions for underserved languages.
1 Introduction
- Underserved Wikipedia communities have limited editors and cannot curate the volume of articles maintained by larger language communities.The paper identifies Esperanto and Arabic as languages with a severe lack of content and active editors relative to English Wikipedia.Source: §1 · p. 1
- Prior encoder-decoder summary work focused on English biographies, whereas this study targets open-domain multilingual generation (Lebret et al., 2016; Chisholm et al., 2017; Vougiouklis et al., 2017).The approach uses copy actions over input entity labels and does not assume unseen triples follow training patterns of properties and entity types.Source: §1 · pp. 1–2
- The model substantially outperforms language-model and information-retrieval template baselines across evaluation metrics in Esperanto and Arabic.The introduction reports the comparison for both languages.Source: §1 · p. 2
- The paper investigates generating summaries from Wikidata triples across domains and adapts an end-to-end model with copy actions to the task.Its summaries are intended as starting points for editors of underserved Wikipedias.Source: §1 · pp. 1–2
2 Model
The model encodes unordered Wikidata triples into a fixed-size representation, then decodes a summary token by token and uses property placeholders to copy entity labels.
2.1 Encoding the Triples
- The feed-forward encoder combines each triple’s subject, property, and object into a fixed-length set representation that is invariant to triple order.That representation initializes the decoder.Source: §2.1 · p. 2
2.2 Decoding the Summary
- A multilayer GRU decoder generates the summary one token at a time, and surface form tuples allow entities to have multiple lexicalisations.The first-layer hidden state is initialized with the encoded triple set.Source: §2.2 · p. 2
2.3 Copy Actions
- Property placeholders replace rare entities with relation-property tokens, which are later mapped to a matching triple’s entity label.When multiple triples share that property, the system randomly chooses an entity.Source: §2.3 · pp. 2–3
2.4 Implementation and Training Details
- The target dictionaries include the 15,000 most frequent Esperanto tokens and 25,000 Arabic tokens, then add 80 and 113 property placeholders, respectively.Rare entities unmatched to a triple subject or object are replaced by the special <resource> token.Source: §2.4 · p. 3
- Training minimizes predicted-summary negative log-likelihood with Adam; evaluation uses beam search of size 20 and retains the highest-probability summary.The reported learning rate is 10^-5.Source: §2.4 · p. 3
3 Dataset
The dataset aligns introductory Wikipedia sentences with Wikidata triples and identifies entity mentions for property-placeholder training.
3 Dataset
- For each Wikipedia article, the dataset aligns its tokenized first introductory sentence with Wikidata triples where the article item appears as subject or object.Entity mentions and their verbalisations are identified to create surface form tuples, using Wikidata labels expanded through Wikimedia’s language fallback chain.Source: §3 · p. 4
- Property placeholders use distant supervision: entities in relations with the article’s main entity are replaced by the corresponding property-placeholder tag.The paper reports corpus statistics for Arabic and Esperanto in Table 3.Source: §3 · p. 4
4 Baselines
The section compares the proposed approach with language-model and information-retrieval baselines. It describes two Kneser-Ney variants and two information-retrieval variants.
4 Baselines
- KN is a 5-gram Kneser-Ney language-model baseline.Source: §4 · p. 4
- KNext is a second KN model trained on summaries containing copy-action tokens.Source: §4 · p. 4
- IR uses TF-IDF and LSA to represent input triples, then retrieves the nearest training item’s summary; the authors provide IR and IRext versions.Source: §4 · p. 4
5 Results and Discussion
The model outperforms the baselines across most evaluation metrics in both languages, with further BLEU-4 gains from copy actions. It also performs and generalizes well across domains.
5 Results and Discussion
- The model improves on baselines across most evaluation metrics in Arabic and Esperanto, using BLEU, METEOR, and ROUGE-L and excluding summaries without an end token.Source: §5 · pp. 4–5
- Around ≤1% of Arabic and 2% of Esperanto validation and test triple sets produced summaries without the end token; the authors attribute the difference to Esperanto’s limited dataset size.Source: §5 · p. 4
- At least 5.25 BLEU-4 in Arabic and 1.31 BLEU-4 in Esperanto over IRext; copy actions add 0.61–1.10 BLEU-4.Source: Table 4 · p. 5, §5 · p. 5
- The model scores better than any baseline across 50 entity-type domains, without its high performance being skewed toward a few domains.Source: §5 · p. 5
6 Conclusions
The conclusion presents an adapted encoder-decoder architecture for generating summaries in underserved languages. It emphasizes structured data and multilingual data preparation, and reports better performance and domain generalization than strong baselines.
6 Conclusions
- An adapted encoder-decoder architecture is presented as a way to address challenges in generating summaries for underserved languages.Source: §6 · p. 5
- The approach leverages structured knowledge-base data and careful multilingual data preparation, which the authors say would otherwise require substantially more data.Source: §6 · p. 5
- The model performs and generalizes across domains better than a set of strong baselines.Source: §6 · p. 5