Concepts used in this library
Glossary
Working definitions for analysing product content. They provide consistent names for operations and errors across the articles.
Claim
Information a text communicates about a product, its application or the result of using it. For verification, a complex sentence can be divided into claims that can each be checked against evidence. For example, “the set contains 6 tubes of 20 ml each; the paint is suitable for fabric” contains independently verifiable information about the set and its intended use.
Scope of applicability of a claim
The products, variants, situations and conditions to which the information applies. A time of “30 minutes” may refer to a thin coat at +20 °C. Removing these qualifications extends the claim to cases that the source data do not support.
Evidence supporting a claim
A source or reproducible inference that supports a statement. A specification may be stated directly in instructions, while total volume may be calculated from tube count and volume per tube. Evidence must apply to the product being checked and retain material conditions. A link alone does not establish that a statement is supported.
No data available
A state in which available information is insufficient for a definite answer. If suitability for fabric is not described, neither permission nor prohibition follows. The gap is recorded separately and may prompt a clarification request to the data owner.
Critical error
An error that prevents publication under the agreed requirements. Examples include describing a different product variant, unsupported compatibility, or losing a condition that materially changes suitability. The list depends on the category and task. Such an error is assessed separately from stylistic improvements.
Partial answer
An accurate answer to only part of the question. “Thinning with water is allowed” answers which thinner may be used but leaves the dosage unknown. The status depends on the question: the same statement may fully answer a narrow need and only partly answer a broader one.
Testable hypothesis
An assumption for which a test method and potentially contradictory data have been specified. “The revision retains all material conditions” is checked by comparison with the source. “The new revision reduces mistaken purchases” requires behavioural observations and an appropriate comparison. These hypotheses need different tests.
Data provenance
Information about where a record came from, which operations were performed on it, and who created or changed it. In a product dossier, this might be a chain from an instruction passage to an extracted value, a calculation and a sentence in the description. This history supports evidence checks and discovery of dependent records; the truth of the source document requires a separate assessment.
Data normalisation
Converting notation to an agreed form while preserving meaning. For example, 0.02 l and 20 ml can be represented in the same unit. The associated object, conditions and source record must be retained. Calculating the total volume of several tubes is a separate operation from unit conversion.
Contradictory data
Incompatible information about the same property with matching objects, conditions and scope. First establish whether the records concern different product variants, periods or modes. If no explanation is available, retain both records and an open question. Missing information in one document is not in itself a contradiction of another.
Search query
An observed expression submitted to search. Its words describe what a person is trying to find, but may not fully reveal the situation or goal. Preserve the original query separately from its interpretation.
User need
In this library’s working distinction, the outcome a person needs in a particular situation—for example, transporting belongings under the conditions of a trip. A need is inferred from available data and refined through research; finding a query does not establish a particular person’s motive.
Search intent
The inferred immediate goal of a search: finding options, comparing them, checking a specification or obtaining instructions. One broader need can involve different search intents. The working classification allows several goals to coexist or change during a search.
Search query volume
The number of searches counted by a chosen tool under specified selection conditions. Interpret the number together with its source, period, region, reporting mode and query restrictions. It is not a count of unique buyers. Volumes from overlapping selections cannot be added as independent searches.
Token
A text element identified under a tool’s rules: a word, number, symbol or other segment. Token boundaries make passages addressable but do not establish the domain meaning of the whole phrase.
Lemma
The dictionary form to which an inflected word is mapped during lemmatisation. Matching lemmas helps identify comparison candidates but does not establish that claims are equivalent. Prepositions, negation, word roles and conditions retain their own significance.
Group of user needs
A working grouping of formulations for a chosen analytical task. To check completeness, this library separates individual questions and preserves conditions that change the answer. A shared topic and neighbouring words alone do not establish one question. The level of aggregation is set in advance.
Vector representation of text
A numerical representation produced by a chosen model for a particular task. It can be compared with other representations using a suitable measure. The score depends on the model and text preparation; proximity does not equal factual truth or guarantee a shared answer.
Entailment in text verification
A relationship in which a premise supports a claim under a consistent interpretation of objects and task conditions. Checking this relationship does not independently establish the premise’s truth. A neutral relation means that the given evidence is insufficient, not that the claim must be false.
Benefit of a feature
The relevance of a verified product capability to a particular task. Removable dividers let the user change an organiser’s compartments; the usefulness of doing so depends on the items and storage method. Benefit does not equal comparative superiority, a guaranteed experience or a measured purchasing effect.
Voice and tone
In editorial work, voice is a brand’s consistent way of communicating, while tone responds to the situation: introducing a product, giving instructions or explaining a limitation. Record requirements through examples and observable editorial decisions while preserving precise names and properties.
Precision and recall in error detection
If the positive class means a claim contains an error, precision is the proportion of justified signals among all signals, and recall is the proportion of detected errors among all reference errors. Their denominators differ. Values are ambiguous without the class, verification unit, rubric and source counts; a metric with a zero denominator is undefined.
Editorial regression
Loss of a previously satisfied requirement after editing. For example, correcting care instructions also removes a drying condition from a neighbouring sentence. Rechecking must establish both that the original error has been fixed and that other required information remains intact.
A publication’s data dependencies
A stored relationship showing which facts and transformations a particular passage, attribute, image caption or translation uses. It helps identify what needs rechecking when a source changes. The completeness and correctness of the relationships themselves need validation; identical content across channels does not establish that their shared evidence is true.
Cosine similarity
A measure of the angle between two non-zero vectors: their dot product divided by the product of their lengths. It is symmetric and does not by itself express entailment direction, a statement’s truth or the probability of an identical question. Interpretation depends on the representations and comparison task.
Search visibility
An observed appearance of a page, product or link in a particular channel and context. Describe the signal together with the system, query or report conditions, period and counting method. Eligibility, indexing, display, citation and orders are different outcomes; one observed instance does not characterise all questions and users.
Question-plan coverage
The proportion of content addressed under a predefined rubric for a fixed question set. If weights are used, numerator and denominator must account for them consistently. The metric applies to its version of the plan and does not automatically measure satisfaction, search impressions or purchases. Source-data readiness is assessed separately.
Model-based evaluator
A model in a specified configuration assigned to check against a rubric and available evidence. Its verdict and explanation must themselves be tested on independently labelled cases. Distinguish the quality of evidence retrieval, claim verification and publication approval.