Source-linked AI summary
Improving Wikipedia Verifiability with AI
Fabio Petroni, Samuel Broscheit, Aleksandra Piktus, Patrick Lewis, Gautier Izacard, Lucas Hosseini, Jane Dwivedi-Yu, Maria Lomeli, Timo Schick, Pierre-Emmanuel Mazaré, Armand Joulin, Edouard Grave, Sebastian Riedel
TL;DR
Wikipedia citations can fail to support claims, making verification difficult. Side learns from existing Wikipedia references to flag weak citations and retrieve alternatives; in user evaluation, people preferred Side’s suggestions over Wikipedia citations more than 60% of the time.
Problem
Wikipedia citations may fail to support claims or rely on questionable sources, leaving readers unable to verify claims from the cited material.
Source: Introduction · p. 1Method
Side trains a retriever and a verification model on Wikipedia claims and citations to find candidate sources and rank them for claim support.
Source: Introduction · p. 2Results
More than 60% of the time, users preferred Side citations over Wikipedia’s, rising above 80% when Side assigned the Wikipedia citation a very low verification score.
Source: Introduction · p. 2Takeaways & Limitations
The authors suggest the work could help people perform fact-checking beyond Wikipedia.
Source: Discussion · p. 8Side assumes Wikipedia claims are verifiable and improves references only for existing claims.
Source: Discussion · p. 8
Abstract
from arXiv · showhide
Verifiability is a core content policy of Wikipedia: claims that are likely to be challenged need to be backed by citations. There are millions of articles available online and thousands of new articles are released each month. For this reason, finding relevant sources is a difficult task: many claims do not have any references that support them. Furthermore, even existing citations might not support a given claim or become obsolete once the original source is updated or deleted. Hence, maintaining and improving the quality of Wikipedia references is an important challenge and there is a pressing need for better tools to assist humans in this effort. Here, we show that the process of improving references can be tackled with the help of artificial intelligence (AI). We develop a neural network based system, called Side, to identify Wikipedia citations that are unlikely to support their claims, and subsequently recommend better ones from the web. We train this model on existing Wikipedia references, therefore learning from the contributions and combined wisdom of thousands of Wikipedia editors. Using crowd-sourcing, we observe that for the top 10% most likely citations to be tagged as unverifiable by our system, humans prefer our system's suggested alternatives compared to the originally cited reference 70% of the time. To validate the applicability of our system, we built a demo to engage with the English-speaking Wikipedia community and find that Side's first citation recommendation collects over 60% more preferences than existing Wikipedia citations for the same top 10% most likely unverifiable claims according to Side. Our results indicate that an AI-based system could be used, in tandem with humans, to improve the verifiability of Wikipedia. More generally, we hope that our work can be used to assist fact checking efforts and increase the general trustworthiness of information online.
Introduction
Wikipedia citations may fail to support claims, making verification difficult for readers. Side learns from existing Wikipedia references to flag potentially weak citations and suggest alternatives, which users often prefer.
Introduction
- Wikipedia verification requires readers to check whether a citation supports a claim and, if it does not, seek new evidence; sources may also be questionable.Claims may still be true even when readers cannot verify them using the cited source.Source: Introduction · p. 1
- Side suggests a retrieved citation when its verifier ranks the original citation below the candidate documents.The system first retrieves candidates from Sphere, then ranks them alongside the original citation.Source: Figure 1 · p. 2
- Automated verification matters because assessing language support and finding evidence online require substantial search effort and careful reading.A machine verifier could help editors identify citations to check and replacements to consider.Source: Introduction · p. 2
- Side trains a retriever to find candidate citations and a verifier to rank existing and retrieved citations for how likely they are to support a claim.The training data comes from a curated corpus of Wikipedia claims and their current citations.Source: Introduction · p. 2
- Nearly 50% of the time, Side returned the exact source used in a high-quality Wikipedia article as its top solution.The other half were not necessarily wrong; they simply did not match Wikipedia’s current source.Source: Introduction · p. 2
- More than 60% of the time, users preferred Side citations over Wikipedia’s, rising above 80% when Side gave the Wikipedia citation a very low verification score.The user study compared existing citations with Side’s suggestions for the same claim.Source: Introduction · p. 2
System Architecture
Side retrieves candidate evidence for Wikipedia claims using sparse and dense search, then ranks original and retrieved sources by how well they support the claim. Its ranking is designed to prioritize claims for human checking and surface likely stronger evidence.
System Architecture
- Side represents a claim as the sentence or clause before a citation, with meaning that may depend on context; cited documents are divided into fixed-word passages.Claim context also includes preceding sentences, the section title, and the enclosing article title.Source: System Architecture · p. 3, The Retrieval Engine · p. 3
The Retrieval Engine
- The retrieval engine combines sparse and dense search over Sphere to find documents that may support a claim.Sparse search generates query text from citation context; dense search encodes context and passages as vectors.Source: The Retrieval Engine · p. 3
The Verification Engine
- The verification engine scores claim-document support using concatenated claim and passage text, then assigns each document its maximum passage score.A fine-tuned BERT transformer computes the per-passage scores.Source: The Verification Engine · p. 3
- The verifier ranks claim-document pairs by verifiability, prioritizing weakly supported existing claims for review and stronger candidate evidence for display.Its training objective rewards ranking existing Wikipedia evidence above retrieved evidence, despite noise in that signal.Source: The Verification Engine · p. 3
Evaluation and results
The evaluation tests retrieval and failed-verification detection before assessing Side’s citation recommendations with crowd and Wikipedia-user annotations.
Experimental Data and Setting
- WAFER contains approximately 3.8M English Wikipedia citations aligned to source text, while approximately 250k citations were retrievable from the August 2019 Sphere snapshot.Sphere contains approximately 900M passages from more than 100M documents and only approximates open-web evidence seeking.Source: Experimental Data and Setting · p. 4, Retrieval evaluation · p. 5
- Evaluation separates featured and non-featured articles and holds failed-verification citations aside, with article-level cross-validation used to avoid test leakage.Retrieval metrics include P@1 and SR@k, which measure the rank of the originally cited document; the highest-ranked passage determines its rank.Source: Experimental Data and Setting · p. 4
Retrieval evaluation
- Combining 100 sparse and 100 dense retrieval results gave the best overall SR@200, and the verification engine ranked the original citation document first nearly 50% of the time.The combined retrieval results feed the verification engine; the sparse retriever outperformed the dense approach on web retrieval.Source: Retrieval evaluation · p. 4, Figure 2 · p. 5
- Featured-article retrieval had a -7.0%/-10.4% P@1 difference from non-featured articles for dense/sparse retrieval, while verification narrowed the gap.The authors attribute the verification engine’s gains to fine-grained text comparison using cross-attention.Source: Retrieval evaluation · p. 5
Detecting Failed Verification
- At 15% recall, passage-level verification achieved approximately 90% precision for failed citations, versus approximately 80% for document-level verification; URL depth was a strong baseline.The verification engine was trained only on positive examples, without explicit failed-verification supervision.Source: Detecting Failed Verification · p. 6
Evaluation of the final system
- More than 40% of the time, crowd annotators found no evidence in the cited reference, and lower verifier scores aligned with more preferences for Side and less evidence in Wikipedia citations.The authors suggest the ranker score may help identify citations needing editor attention.Source: Evaluation of the final system · pp. 6–7, Figure 3 · p. 6
- Wikipedia users selected the existing citation in 16% of demo cases, while Side’s recommendation or neither was preferred in 65%, and Side’s top-1 recommendation was judged appropriate in 26%.The study included 43 authenticated users and 106 annotations.Source: Evaluation of the final system · p. 7, Figure 4 · p. 7
Related Work
Related work situates Side among tools and research aimed at improving Wikipedia citations and verifying claims. It presents Side as a complement to human review and existing article-generation tools.
Related Work
- Side could complement Citation Hunt by surfacing potentially unverified claims and suggesting alternative citations for human validation.Citation Hunt lets people review claims flagged as lacking a reliable source and propose better citations.Source: Related Work · pp. 7–8
- Citation-retrieval research spans approaches from linking articles to detecting citation spans, while broader source recommendation often draws on smaller, less diverse corpora and focuses less on verifiability.The passage names Jana et al. (2018), Fetahu et al. (2017), and Färber & Jatowt (2020).Source: Related Work · p. 8
- AI article-generation tools could be complemented by Side's suggestions of web evidence to support generated articles.The passage discusses Scribe (2022) for bootstrapping articles in underrepresented languages.Source: Related Work · p. 8
- Wikipedia fact-checking research often evaluates systems on synthetic resources, while the authors propose real-world Wikipedia citations as supervision for broader applicability.The passage cites Thorne & Vlachos (2018), Thorne et al. (2018, 2019), Schuster et al. (2021), and Trokhymovych & Saez-Trumper (2022).Source: Related Work · p. 8
Discussion
Side applies neural methods to improving Wikipedia citations in a noisy, loosely defined real-world setting. The authors identify limits in its source and language coverage, and describe possible broader uses.
Discussion
- Side demonstrates neural methods for citation improvement in a noisier, less specified real-world task than the synthetic, well-specified settings common in prior work.The cited prior studies include Bowman et al. (2015), Wang et al. (2018), Camburu et al. (2018), Nie et al. (2019), Pérez-Rosas et al. (2017), Thorne et al. (2018), and Thorne & Vlachos (2018).Source: Discussion · p. 8
- The current system covers web pages and English, leaving books, scientific articles, other document types, non-text modalities, and other languages outside its scope.The authors identify multimodal and multilingual support as future research directions.Source: Discussion · p. 8
- Side assumes Wikipedia claims are verifiable and improves references for existing claims; detecting unverifiable claims is a distinct extension without straightforward training supervision.The authors note that demonstrating unverifiability requires contradictory evidence, which Wikipedia currently does not contain.Source: Discussion · p. 8
- The authors release the data, code, and models and suggest the work could support human fact-checking beyond Wikipedia.They also state that the work could lead to more trustworthy information online.Source: Discussion · p. 8
Sphere retrieval
Side combines sparse and dense retrieval approaches and trains its neural components using existing Wikipedia citations. The training setup addresses noisy document-level data and passage-level evidence through a latent-variable approach.
Sphere retrieval
- BM25 benefits from lexical overlap, but claims may be fragmented or depend on context; a small manual review found co-reference crucial in roughly one-third of 30 sentences.Using the sentence before the claim and the article title yielded the best BM25 results.Source: Sphere retrieval · p. 12
- DPR encodes queries and passages as dense vectors and learns to score correct query-document pairs above incorrect ones, helping retrieve rephrased claims.Its BERT-like query and passage encoders learn which text elements matter.Source: Sphere retrieval · p. 13
- Wikipedia's existing citations provide training examples, but they are noisy and pair claims with documents rather than passages; expectation-maximization models the evidence passage as latent.The data contains positive examples only, and the authors note uncertainty about how well that supervision trains the verification engine.Source: Sphere retrieval · p. 13
Evaluation details
The supplied evaluation details identify the considered web snapshots and describe several dataset and evaluation visuals. The available table and figure descriptions provide limited information about results.
CommonCrawl snapshots considered
- The listed Common Crawl snapshot codes range from 2017-26 to 2020-10, with multiple captures in between.The passage enumerates the considered snapshot codes.Source: CommonCrawl snapshots considered · p. 13
- Figure 5 presents an example citation from WAFER, whose instances include claim context and cited-source metadata and text.The WAFER description identifies the article context around the citation and the cited source's title and full text.Source: Figure 5 · p. 13
- Figure 6 compares majority-vote crowd preferences for Side versus Wikipedia citations on 2016 claims where Side has a higher evidence score, grouped by ranker-score buckets.Five annotations are collected per claim, and bucket sizes are reported above the buckets.Source: Figure 6 · p. 14
- Figure 7 is identified as a URL-depth analysis, but its description does not specify the encoding or findings.Source: Figure 7 · p. 14
- Table 2 describes an example where both citations receive relatively low verification scores, suggesting Side found insufficient evidence for the claim.Source: Table 2 · p. 15
- Tables 3 and 4 are identified as WAFER dataset statistics and WAFER test results, respectively; the supplied descriptions do not give their values.Source: Table 3 · p. 16, Table 4 · p. 16