Source-linked AI summary
Knowledge-Based Trust: Estimating the Trustworthiness of Web Sources
Xin Luna Dong, Evgeniy Gabrilovich, Kevin Murphy, Van Dang, Wilko Horn, Camillo Lugaresi, Shaohua Sun, Wei Zhang
TL;DR
Web-source quality is difficult to assess from popularity signals because factual reliability and extraction accuracy are different properties. The paper introduces KBT, a multi-layer probabilistic method that jointly estimates them from extracted facts. On web-scale data, it estimates trustworthiness for 119 million webpages and reports improvement over existing knowledge-fusion techniques.
Problem
Popularity-based signals such as hyperlinks do not directly measure whether a source provides correct facts, while extraction errors can be mistaken for source errors.
Method
KBT jointly infers fact correctness, source accuracy, extraction correctness, and extractor quality in a multi-layer probabilistic model.
Results
119 million webpages and 5.6 million websites received trustworthiness estimates from 2.8 billion extracted triples, with reported improvement over existing knowledge-fusion techniques.
Takeaways & Limitations
KBT provides an additional signal for web-source quality that can complement traditional signals such as PageRank.
Takeaways & Limitations
The method is limited by extractors’ ability to extract information comprehensively across websites.
Abstract
from arXiv · showhide
The quality of web sources has been traditionally evaluated using exogenous signals such as the hyperlink structure of the graph. We propose a new approach that relies on endogenous signals, namely, the correctness of factual information provided by the source. A source that has few false facts is considered to be trustworthy. The facts are automatically extracted from each source by information extraction methods commonly used to construct knowledge bases. We propose a way to distinguish errors made in the extraction process from factual errors in the web source per se, by using joint inference in a novel multi-layer probabilistic model. We call the trustworthiness score we computed Knowledge-Based Trust (KBT). On synthetic data, we show that our method can reliably compute the true trustworthiness levels of the sources. We then apply it to a database of 2.8B facts extracted from the web, and thereby estimate the trustworthiness of 119M webpages. Manual evaluation of a subset of the results confirms the effectiveness of the method.
1. INTRODUCTION
The paper introduces Knowledge-Based Trust (KBT), which estimates web-source reliability from factual correctness rather than popularity signals. Its multi-layer probabilistic model separates source errors from extraction errors and scales to web-scale evaluation.
- Motivation: Web-source quality matters for search, but hyperlinks and browsing history primarily capture popularity rather than factual reliability.Less popular websites may nevertheless provide accurate information.
- Approach: KBT estimates source trustworthiness by jointly inferring fact correctness and source accuracy from redundantly extracted web facts.The iterative inference process uses source accuracy to assess facts and fact correctness to assess sources.
- Approach: The probabilistic model distinguishes incorrect facts on webpages from incorrect extractions, improving estimates of source reliability.This distinction addresses the observation that extraction errors can be more prevalent than source errors.
- Evaluation and scale: 2.8 billion extracted triples supported trustworthiness estimates for 119 million webpages and 5.6 million websites.The method also dynamically aggregates or splits sources to manage sparse data and computational bottlenecks.
- Implications: The approach provides an additional web-quality signal that can complement signals such as PageRank and extend beyond knowledge extraction.The paper identifies combining KBT with existing signals as a research opportunity.
2. PROBLEMDEFINITIONANDOVERVIEW
The paper formalizes KBT estimation from extractor observations and explains why a single-layer fusion model cannot separate source reliability from extraction noise. The proposed multi-layer model adds latent extraction correctness and separate source and extractor parameters.
- Problem definition: The observation matrix records whether extractor e extracted value v for data item d on web source w.These observations can be represented as a sparse data cube.
- Problem definition: KBT estimation seeks web-source accuracies from an observation matrix of extracted triples.Each source accuracy is the probability that a value provided for a fact is correct.
- Single-layer limitations: A single-layer model treats each extractor–webpage pair as a source, so it cannot estimate webpage accuracy independently of extractor quality.An extraction error could therefore make a reliable webpage appear untrustworthy.
- Single-layer limitations: Equal counts of extracted values do not establish equal truthfulness when reliable extractors agree and unreliable extractors produce errors.The Obama-nationality example illustrates how extractor agreement can explain away conflicting values.
- Multi-layer overview: The multi-layer model introduces latent variables for extraction correctness and separate accuracy parameters for webpages and extractors.This structure factors extraction noise out of web-source accuracy estimation.
- Overview: The paper reports that the model and dynamic source granularity improve performance over the single-layer model.Granularity selection addresses both sparse webpages and pages with very many extracted triples.
3. MULTI-LAYER MODEL
The multi-layer model computes web-source accuracy parameters from the observation matrix of extractor outputs.
- 3. MULTI-LAYER MODEL: The model computes A = {Aw}, the set of web-source accuracies, from X = {Xewdv}, the observation matrix.This is the central estimation task described for the multi-layer model.
3.1 The multi-layer model
The multi-layer model separates whether a webpage provided a triple from whether an extractor correctly observed it. It combines source accuracies, extractor noise parameters, priors, and conditional-independence assumptions in one joint model.
- Source layer: Latent variables Cwdv represent whether web source w actually provides triple (d, v).These variables depend on the true value Vd and the web-source accuracy Aw.
- Extractor layer: The observation model uses extractor recall Re and false-extraction probability Qe to represent noise in observed data.Re is the probability of extracting a truly provided triple, while Qe is the probability of extracting an unprovided triple.
- Extractor layer: Qe is related to extractor precision Pe and recall Re through the model’s noise parameterization.The paper gives an explicit relationship involving γ, Pe, and Re.
- Parameterization: The model uses uniform parameter priors and default initial values Aw = 0.8, Re = 0.8, and Qe = 0.2.An alternative initialization for Aw uses an external estimate of triple correctness based on Freebase.
- Joint model: The joint distribution includes latent true values V and source-provision variables C, with dependencies represented by a graphical model.Observed data are shaded, hidden variables or parameters are unshaded, and plates represent repeated variables.
3.2 Inference
The paper uses an iterative EM-like procedure to estimate latent extraction correctness, data-item values, source accuracies, and extractor reliabilities at web scale. It alternates latent-variable and parameter updates until convergence.
- Exact posterior inference is computationally intractable because of latent variables, while Gibbs sampling can be slow and difficult to implement in Map-Reduce.The paper therefore uses a faster EM alternative and an iterative EM-like procedure.
- The MULTILAYER algorithm alternates between estimating latent variables and model parameters, stopping when they converge or the iteration limit is reached.Each iteration estimates C, V, source accuracies, and extractor reliabilities.
- The latent-variable updates first estimate extraction correctness for each source-triple pair, then infer the most probable true value for each data item.These computations can be parallelized across relevant data items, sources, and values.
- After latent variables are estimated, the procedure updates source accuracies and extractor precision and recall parameters in parallel.The updated parameters are used in the next iteration.
3.3 Estimating the latent variables
The latent-variable inference estimates whether each extracted fact is correct and which value is true, while iteratively refining their dependence through source accuracy and prior updates. An improved procedure marginalizes uncertainty in extraction correctness rather than relying only on MAP assignments.
- 3.3.1 Estimating extraction correctness: The model estimates extraction correctness from extractor observations using Bayes rule, extractor likelihoods, and presence and absence votes.For each source-triple pair, extractor outputs are combined into a posterior probability that the source provides the triple.
- 3.3.1 Estimating extraction correctness: Source-specific vote counts can produce near-certain posteriors: W1 receives vote count 11.7 and probability 1, whereas W6 receives -9.4 and probability 0 for the Obama-nationality-USA triple.These values use the extractor votes described for the example.
- 3.3.2 Estimating data-item values: The value-inference step aggregates estimated source evidence to compute a posterior distribution over possible values, assigning remaining mass uniformly to unobserved domain values.In the example, USA receives probability 0.995 and Kenya 0.004.
- 3.3.3 An improved estimation procedure: The improved procedure marginalizes over extraction-correctness uncertainty instead of using only MAP estimates, and experiments show that this improves estimation.A weighted vote-counting heuristic is proposed as a simple approximation.
- 3.3.4 Updating the prior: The prior probability that a source provides a triple is re-estimated after each iteration using inferred value correctness and source accuracy.The refined prior is then used in the following iteration.
3.4 Estimating the quality parameters
The quality-parameter stage estimates source accuracy from the probabilities that its provided facts are true, while also estimating extractor precision and recall. The source-accuracy equation weights fact correctness by the probability that each fact was actually provided by the source.
- 3.4.1 Source quality: The method estimates source accuracy as the average probability that the values provided by that source are true.It incorporates uncertainty about whether an extraction corresponds to a fact contained in the source.
- 3.4.1 Source quality: Knowledge-Based Trust is the weighted average of fact-truth probabilities, with weights given by the probabilities that the facts are contained in the source.This equation is the basis of the paper’s source-quality estimate.
- 3.4.2 Extractor quality: Extractor precision and recall are estimated from the inferred extraction outcomes, with extractor quality Qe computed from Pe and Re rather than estimated directly.The paper cites prior work as the reason for preferring this parameterization.
3.5 Handling confidence-weighted extractions
The method extends binary extraction observations to confidence-weighted outputs because thresholding can discard useful evidence. It modifies inference and parameter estimation to retain extractor confidence information.
- Real extractors return confidence scores, but binarizing those scores by thresholding loses information.The paper treats a confidence score as the extractor’s probability that the triple is present.
- With collective evidence, partially confident extractions can support confidence that W3 and W4 provide the Obama-nationality-USA triple.The example uses extractor confidence values of .85 and .5.
- A threshold of .7 would ignore E3’s extractions from W3 and W4, leading the example to assign equal probability to USA and Kenya.The threshold discards the lower-confidence evidence before inference.
- The confidence-weighted extension modifies both the extraction-correctness inference and the precision and recall estimates.These changes follow the weighted approach introduced earlier.
4. DYNAMICALLY SELECTING GRANULARITY
The paper dynamically selects source granularity to balance reliable accuracy estimates against computational efficiency. It backs off small sources to shared parents and splits large sources into independently evaluated sub-sources.
- Motivation: Source granularity trades off estimation reliability against computational cost: finer sources may lack data, while large sources can create bottlenecks.The finest granularity could distinguish webpage- or predicate-specific accuracy, but insufficient data can make estimates unreliable.
- Merge: Small sources are merged with sibling sources sharing a parent, allowing related pages or predicates to contribute statistical strength.Merging removes a feature and combines child sources under their common parent; the merge threshold m is chosen to preserve enough statistical strength.
- Source hierarchy: Sources are organized hierarchically as ⟨website, predicate, webpage⟩, from general website-level sources to specific webpage-and-predicate sources.For example, a website is the parent of its predicate-specific source, which is the parent of a webpage-specific source.
- Split: Large sources are randomly split into similarly sized sub-sources, with the number of buckets determined by the desired maximum size M.Triples are uniformly distributed across ⌈|W|/M⌉ buckets, and M is selected to avoid unnecessary splitting and computational bottlenecks.
- Algorithm: SPLITANDMERGE processes finest-granularity sources by splitting oversized sources, backing off undersized sources to parents, and retaining sources already within bounds.Merging or splitting may need to repeat because a merged parent can remain too small or become too large.
- Extractor granularity: The same granularity strategy extends to extractors, whose quality is represented across extractor, pattern, predicate, and website levels.Different extractor patterns may have different quality for the same predicate, motivating multiple extractor resolutions.
5. EXPERIMENTAL RESULTS
Synthetic experiments show that MULTILAYER estimates extraction correctness, triple truthfulness, and source accuracy more effectively than SINGLELAYER. Real-world evaluations further examine calibration, prediction quality, ablations, and computational efficiency.
- Synthetic data: MULTILAYER always outperforms SINGLELAYER as the number of extractors increases, with lower square loss across evaluated quantities.SqV decreases quickly and SqC more gradually for MULTILAYER, while SqA remains stable despite additional noise extractions.
- Synthetic data: Higher extractor and source quality generally lowers loss, although increased recall or precision can introduce noise that slightly worsens selected metrics.The overall synthetic experiments nevertheless show that the algorithm approximates true parameter values in controlled settings.
- Real-world data: The multi-layer model has a better PR curve than the compared methods, while all methods are fairly well calibrated.SINGLELAYER often assigns low probabilities to true triples, producing many false negatives.
- Real-world data: MULTILAYERSM performs better than MULTILAYER, whereas MULTILAYERSM+ performs worse than MULTILAYER+, revealing an interaction between source granularity and initialization.The comparison shows that initialization and granularity cannot be evaluated independently.
- Real-world data: 80% of type-error triples receive predicted probabilities below 0.1, while 54% of correct Freebase triples receive probabilities above 0.7.For correct triples, 26% receive probabilities below 0.1 and 54% above 0.7; for type errors, only 8% exceed 0.7.
- Computational efficiency: 8.8× speedup results from splitting large extractors, source-accuracy computation falls 20%, average iteration speedup reaches 3×, and total execution time drops by half.Adding merging increases overall execution time over splitting by only 8.6%, while parameter choices trade preparation time against inference time.
5.4 Experiments related to KBT
The experiments assess KBT as a trustworthiness signal at web scale and compare it with PageRank, while identifying limitations involving relevance, trivial facts, extraction coverage, and copied content.
- Web-scale evaluation: 119 million webpages and 5.6 million websites had at least five triples that the multi-layer model believed were correctly extracted.The dataset contained more than 2 billion webpages from 26 million websites.
- Web-scale evaluation: 52% of websites had KBT above 0.8, with the score distribution peaking at 0.8.
- KBT vs PageRank: KBT and PageRank were almost orthogonal across 2,000 randomly selected websites, indicating that they provide different signals.
- Discussion and limitations: Further improvement requires filtering topic-irrelevant and trivial triples, extending extraction capabilities, and identifying websites that scrape other sources.The authors note that broader extraction methods may increase coverage while also introducing more noise.
6. RELATED WORK
The paper distinguishes KBT from link-, behavior-, and spam-oriented web-quality signals by using factual correctness, and from prior data-fusion models by explicitly modeling extraction errors.
- Web-source quality signals: PageRank, Authority-hub analysis, EigenTrust, TrustMe, Web topology, TrustRank, and AntiTrust rely on link structure, source behavior, or spam detection rather than factual correctness.
- Data fusion: Data-fusion research resolves conflicts among multiple sources using link-based, IR-based, accuracy-based, and graphical-model measures of source trustworthiness.
- Graphical models: Earlier graphical models address variants such as single truths, numerical values, multiple truths, and source correlations, but do not model extractors as distinct error sources.
- Prior knowledge-fusion work: Compared with the authors’ previous knowledge-fusion work, MULTILAYER improves over SINGLELAYER and enables evaluation of web-source quality with KBT.
7. CONCLUSIONS
The conclusion presents KBT as a metric for web-source quality based on jointly estimating extraction correctness, source data correctness, and source trustworthiness, with dynamic source granularity. Experiments show promise for web-quality evaluation and improvement over existing knowledge-fusion techniques.
- KBT is proposed as a new metric for evaluating web-source quality.
- A probabilistic model jointly estimates extraction correctness, source-data correctness, and source trustworthiness.
- An algorithm dynamically decides the appropriate granularity for each source.
- Experimental results show promise for web-source-quality evaluation and improvement over existing knowledge-fusion techniques.