Data Management
Build a data dictionary before cleaning business records
Agree what fields mean, where values come from and how exceptions should be handled.
IGS Editorial · 3 min read · Updated 17 November 2026

Data cleaning often begins with visible inconsistencies: mixed date formats, missing categories or duplicate names. Yet changing values before agreeing what they mean can make a dataset look tidier while reducing its reliability. A data dictionary establishes the rules that cleaning should follow.
The dictionary need not be elaborate at first. Start with the fields that drive decisions, transactions or reporting. Give each a definition that someone outside the original team can understand without relying on institutional memory.
Define meaning, not just format
A field called customer date might refer to registration, first purchase or most recent contact. A format rule cannot resolve that ambiguity. State the business event represented and identify the system that records it authoritatively.
Include examples and exclusions. If account status means billing status rather than service availability, say so. Distinguish an empty value from unknown, not applicable and not yet collected when those states have different consequences.
Specify valid values
Record the permitted type, range, length and controlled vocabulary where relevant. Explain whether leading zeros matter and whether a number is actually an identifier. Converting an identifier to a numeric value can silently remove meaningful characters.
For dates, define the source format and intended storage format. A value such as 04/05 can be ambiguous without source context. Ambiguous records should be reviewed rather than guessed into a consistent-looking format.
Identify ownership
Each important field needs a business owner who can decide disputed meaning and a source owner who understands how values are produced. These may be different roles. A cleaning operator should know where to send unresolved questions.
Record who can approve a rule change. If one report needs a different interpretation, consider a derived field rather than silently redefining a shared one and breaking other uses.
Connect rules to validation
Translate definitions into checks where practical. A permitted-status list can become a validation rule; a required relationship may need a cross-field check. Keep the rule understandable alongside any technical implementation.
Classify failures by consequence. A missing optional description may be a warning, while an invalid transaction reference may block processing. Avoid a single error label that treats all issues as equally serious.
Maintain the dictionary with the process
Review definitions when systems, products or reporting requirements change. Preserve versions so historical data can be interpreted using the rules in effect at the time. Tell downstream users when a field's meaning or permitted values change.
Use questions raised during cleaning to improve the dictionary. Repeated uncertainty is evidence of a missing definition, not simply operator inexperience. A reliable dictionary gives the organisation a common language for data and makes later quality work easier to explain, reproduce and audit.
Resolve an ambiguous field before bulk correction
Imagine a file containing a field called active date. Sales considers it the first purchase date, while operations uses it for the start of service. Both teams have populated values that make sense within their own interpretation. Formatting the column consistently would not resolve the underlying conflict.
Bring the owners together to decide whether one definition should apply or two fields are needed. Review downstream reports before changing meaning. Historical values may require a source-specific interpretation rather than a single automatic conversion.
Add the approved definition, examples and transition rule to the dictionary. Then design validation that reflects the decision. Keep unresolved records in a review route and preserve their original values. This example illustrates why semantic agreement precedes cleaning. A tidy dataset can remain misleading when different contributors use the same label for different events. The dictionary provides a place to make those distinctions explicit and maintain them when systems or teams change.
