Abstract
The growth of Wikipedia, limited by the availability of knowledgeable authors, cannot keep pace with the ever increasing requirements and demands of the readers. In this work, we propose Wiki-Write, a system capable of generating content for new Wikipedia articles automatically. First, our technique obtains feature representations of entities on Wikipedia. We adapt an existing work on document embeddings to obtain vector representations of words and paragraphs. Using the representations, we identify articles that are very similar to the new entity on Wikipedia. We train machine learning classifiers using content from the similar articles to assign web retrieved content on the new entity into relevant sections in the Wikipedia article. Second, we propose a novel abstractive summarization technique that uses a two-step integer-linear programming (ILP) model to synthesize the assigned content in each section and rewrite the content to produce a well-formed informative summary. Our experiments show that our technique is able to reconstruct existing articles in Wikipedia with high accuracies. We also create several articles using our approach in the English Wikipedia, most of which have been retained in the online encyclopedia.
Introduction
The growth of Wikipedia is fundamentally constrained by the availability of human editors, leading to a proliferation of "red-linked" entities—topics that are mentioned across the site but lack dedicated pages. Previous attempts to automate article generation have faced three significant hurdles: category dependency (requiring manual classification), copyright violations (due to direct copy-pasting from web sources), and a lack of narrative coherence.
WikiWrite is an end-to-end system designed to overcome these challenges. By leveraging document embeddings to identify similar existing articles and employing a novel two-step optimization process for paraphrastic summarization, it generates coherent, informative content that adheres to Wikipedia's standards while avoiding the plagiarism issues of earlier automated systems.
Methods
WikiWrite introduces a sophisticated pipeline that transforms raw entity names into polished, sectioned articles.
Entity Representation and Similarity
Instead of relying on manual category labels, the system utilizes Paragraph Vector Distributed Memory (PV-DM) to create vector representations of entities. By calculating cosine similarity in this vector space, WikiWrite identifies "neighbor" articles to infer the ideal structure for the new topic.
Information Retrieval and Sectioning
The system reformulates search queries by combining the target entity name with frequent nouns found in the introductory sentences of its neighbors. After scraping the top 20 Google results, a Random Forest classifier assigns retrieved snippets to appropriate sections (e.g., "Early Life," "Career"). To handle variations in section titles, the system uses Repeated Bisection Clustering (RBR) to group semantically identical headers.
Paraphrastic Summarization
The core innovation of WikiWrite is a two-step Integer Linear Programming (ILP) model:
- Selection & Coherence: The model selects sentences that maximize informativeness while ensuring local coherence by modeling the transition probabilities of nouns and verbs between adjacent sentences.
- Rewriting: To avoid copyright issues, the system utilizes the Paraphrase Database (PPDB) to swap words and phrases with synonyms, optimizing for readability and meaning preservation via a trigram language model.
Datasets and Evaluation
The system was trained on a June 2015 Wikipedia dump containing 4.8 million articles. Evaluation was conducted on a set of 15,500 red-linked entities referenced at least 20 times across the site. For performance testing, the authors used 1,000 popular articles for reconstruction and 50 brand-new entities for live deployment on the English Wikipedia.
Key Results
WikiWrite demonstrated significant improvements over previous state-of-the-art systems like WikiKreator. The system achieved an F1-score of 0.622 for section classification, outperforming the baseline (0.481) while running five times faster.
In live tests, 47 out of 50 (94%) generated articles were moved into the Wikipedia mainspace and retained by human editors. Crucially, the paraphrasing engine successfully mitigated plagiarism concerns; only 3 articles had content removed for quality or plagiarism issues, a massive reduction compared to previous automated attempts. In reconstruction tasks, the system achieved a ROUGE-1 score of 0.441, indicating high efficacy in capturing the essence of a topic.
Limitations and Future Work
Despite its success, WikiWrite is not yet fully autonomous. Most articles (94% retention notwithstanding) required minor human intervention for formatting or reference adjustments. Additionally, some generated content was flagged for a "promotional" tone, likely inherited from the web sources used during retrieval. Future research will focus on refining the paraphrasing engine to better capture a neutral, encyclopedic tone and handling more complex sentence restructurings.
Summary updated