How to find tenders when buyers use unfamiliar words

Miss

Begin with a notice your current words failed to find

Define the capability and freeze the existing text query before adding anything. Then identify at least one official notice that a knowledgeable reviewer considers relevant but the current query did not retrieve. The notice may have been found through a classification code, buyer watch, colleague, award lineage or semantic search. Record that discovery route. It proves a miss without pretending that every unseen relevant notice is known. The task is to learn what language would have found this record prospectively.

Read the title, procedure and lot descriptions, requirements and public attachments. Find the smallest phrase that carries the intended concept in context. For a document-digitisation offer, the useful evidence may be “archive image capture and metadata indexing,” not the isolated word “capture.” Preserve the whole sentence and its field. Also record words that looked similar but described a different object. One missed notice creates a candidate phrase, not an approved market-wide synonym.

Corpus

Collect independent buyer usage before generalising a phrase

Assemble a dated corpus from official notice data, not search-engine snippets. TED’s public Search API exposes published notices for analysis and returns links to available formats and languages. The UK central platform also publishes structured notice data for reuse. Use such sources to collect a manageable sample with stable identifiers, notice stage, language, jurisdiction, buyer and lot. Include known irrelevant records that use similar words. A contrast set is essential for discovering when a phrase changes meaning.

Count independent uses carefully. Several lots copied from one framework do not equal several buyers. A buyer may reuse a standard specification across years. Mark shared templates and linked procedures. Two independent authorities using the same expression for the same deliverable provide stronger transfer evidence than twenty duplicated notices. Diversity of context also matters: a phrase used in both a health authority and a municipal archive may travel better than one tied to a single programme name. Do not set a universal minimum; expose the evidence count and distribution.

• Use official notice bodies or attached public documents.
• Retain publication number, version, lot and source language.
• Separate independent buyers from repeated templates.
• Include relevant, irrelevant and unjudged examples.
• Do not mix planning, competition and award corpora without labels.

Phrase map

Organise exact phrases by the concept they express

Extract multi-word units that a buyer could plausibly repeat: purchased object, service action, operational outcome, affected asset, user group, deliverable, statutory programme and established abbreviation. Keep each phrase in its original language and capitalization where those details matter. Attach the phrase to a concept ID written in plain language. “Records conversion,” “archive scanning” and “image capture with indexing” may all sit under the concept “turn physical records into searchable digital files,” but only after their passages support that reading.

Use controlled vocabularies as orientation, not proof of buyer usage. EuroVoc distinguishes concepts from the terms used to label them and supports preferred, alternative, broader, narrower and related relationships. That is a useful discipline: a related concept is not automatically an alternative term. Procurement classification codes belong in another layer because they classify the purchase rather than reproduce free text. The phrase map can link a code for exploration while keeping code evidence and natural-language evidence separate.

Experiment

Test each phrase in the portal that will run the alert

Freeze the portal or API, query interface version, publication dates, geography, notice stages and any code filter. Run the anchor by itself, then one candidate phrase per arm. Save the exact syntax and ask the API to validate it when that facility exists. TED’s Search API supports expert queries and a syntax-check option; its field list shows that notice title, classifications and many procedure or lot fields are distinct. A phrase searched across all text is not equivalent to the same phrase restricted to a lot description. Test the behavior you will actually deploy.

Deduplicate by stable notice or procedure identity before judging additions. For each arm, identify records absent from the current approved set and review them against the same capability rule. Record unique relevant additions, unique false positives, duplicates and unjudged tail. Do not call the result recall unless the complete relevant universe is known. The practical question is whether the phrase finds useful notices the approved search missed at an acceptable review cost. A phrase with no unique additions can still be retained as a resilience path only if that choice and cost are explicit.

Noise

Constrain ambiguous buyer language without hiding mixed contracts

When a phrase finds the right concept only in a particular setting, store the setting as part of the rule. “Image capture” may need a records, archive or indexing co-term. “Migration” may need data, application or platform depending on the capability. Prefer a positive qualifier that establishes the intended object. A negative term is harder to govern because a valid mixed contract can contain both the target and the excluded concept. Test exclusions on every known relevant notice before automation.

Distinguish polysemy from breadth. A broad phrase names a real parent need and may be useful with qualification. A false friend uses the same word for another concept. A related term describes work that could accompany the capability but does not buy it. These deserve different states. An agent should show the passages that led to its classification and ask for review when contexts conflict. It must not solve noise by silently raising a semantic-score threshold whose meaning reviewers cannot reproduce.

• Prefer positive object or outcome qualifiers.
• Test every negative term against mixed known-relevant notices.
• Label broad, ambiguous, related and false-friend phrases differently.
• Keep query syntax specific to the target portal.
• Return conflicting contexts rather than averaging them.

Operation

Publish a phrase set an agent can execute and dispute

For every entry publish concept, exact phrase, language, source notices, independent-buyer count, representative passage, scope field, ambiguity state, paired terms, exclusions, executable query, test window, result judgement, decision, owner and next review trigger. Preserve rejected candidates with reasons. Without that history, another model will rediscover the same attractive false friend. Store the approved set separately from the research queue so untested suggestions cannot leak into production alerts.

An agent may mine new official notices for candidate expressions, cluster passages and run bounded read-only tests. It may not approve its own semantic inventions merely because they are close in embedding space. Require a cited buyer use and an observed retrieval contribution. Reopen the set when a known relevant notice is missed, an approved arm’s noise exceeds the review budget, a phrase changes meaning across buyers, or the portal changes query behavior. Phrase expansion ends at discovery. Qualification still decides whether a retrieved procurement merits pursuit.

• Separate approved, conditional, rejected and untested phrases.
• Expose evidence passages and independent-buyer counts.
• Version the search surface with its literal query syntax.
• Name drift triggers and a human review owner.
• Do not let vocabulary expansion authorize qualification or outreach.