Where one need ends and another begins

What holds a group together?

For a product page, a useful group of formulations shares a sufficiently narrow question. A common topic name does not establish that. “Water” can cover thinning, brush care, stain removal and resistance of the finished coating.

Grouping should be evaluated against the task. If we need topics for a comprehensive manual, “thinning” works as a section. To assess product-page completeness, “what can I thin it with?” and “in what ratio?” are better counted separately: answering the first does not answer the second.

Before merging, check whether the formulations require the same information under the same essential conditions. An answer sufficient for one must remain sufficient for the other. The object, state, action and intended result matter. A long paragraph answering several questions does not itself prove that there is only one need.

The structure may be hierarchical: broad topic, individual question, necessary clarifications. Choose the grouping level before calculating coverage. Otherwise, splitting or merging groups will change the final score without any change to the text itself.

Similarity helps retrieve candidates

With many strings, manually comparing every pair is impractical. Retrieving similar formulations allows candidates to be selected first and then examined for domain meaning. Dictionaries, lemmas, rules and vector representations can all be used.

Sentence-BERT describes a way to obtain sentence representations and compare them using measures including cosine similarity. The Sentence Transformers documentation demonstrates text-similarity scoring. A score depends on the model and measure selected. It is neither a percentage of identical meaning nor a probability that merging is correct.

Keep two decisions separate: “these strings are worth comparing” and “these strings may be treated as one question”. Related topics support the first. The second requires checking differences that change the answer.

A numerical threshold should not be chosen simply because it neatly separates a few familiar examples. Cases from the relevant language and category are needed, including negation, neighbouring tasks and ambiguous wording. Changing the model or text preparation can alter scores and require recalibration.

The Cross-Encoder model catalogue lists similarity, duplicate-question detection, relevance and logical-relation tasks separately. This matters: a tool that retrieves a shared topic has not necessarily been trained to check whether two queries need one answer.

How a chain expands a group

A pairwise similarity threshold can produce an unexpected final group. Suppose we connect every pair with cosine similarity above 0.8, then group connected strings. If A is close to B and B to C, the algorithm may put all three into one component. This does not guarantee direct similarity between A and C.

Three unit vectors in a plane are enough to demonstrate this: A points at 0°, B at 30° and C at 60°. Their cosine values are calculated from the angles between their directions:

This is a mathematical illustration, not a measurement of queries or a recommended threshold. It shows why a link through an intermediate string cannot replace checking the boundaries of the final group.

In practice, such a chain might lead from routine thinning to changing consistency, then to restoring dried material. The intermediate formulation connects topics but does not make the instructions interchangeable. Reviewing a group’s most different questions and cases that change the buyer’s action is useful.

The review method depends on the algorithm. Some approaches check all pairs; others compare with a representative and group constraints; others require an editorial decision for disputed links. Each approach needs testing for its own errors: a representative can also be too broad.

Shared topic, shared question and entailment

Two phrases can be similar yet express incompatible claims. “Suitable for fabric” and “not suitable for fabric” concern the same property of the same material. In description review, they cannot count as the same answer. Negation and its scope must be retained.

Some relationships are directional. “The set contains six tubes” entails “the set contains tubes”. The presence of tubes does not entail their number. Cosine similarity is symmetric: swapping the vectors does not change its value. One such number therefore cannot, by itself, represent the direction of entailment.

Question grouping needs another relationship: whether queries can be treated as duplicates for the chosen task. “What can I thin it with?” and “How much water should I add?” may sit next to each other in a manual but ask for different information. Both should be counted separately on the needs map if they matter to choice. The availability of supported answers is checked in the next step.

Specify exactly which relationship the decision seeks. “Group similar items” leaves that to the person or system doing the work. “Find formulations of the same question about the permitted thinner for this formulation” defines the subject and boundaries.

A query group is not yet a product group

The thinning question may be shared by many artists’ paints. Its answer still depends on the particular formulation. Semantic grouping of queries does not authorise extending one model’s recommendation to the whole category.

It is therefore useful to store intent and applicability as separate attributes. One records what the person wants to know. The other records which product or product class a possible answer concerns and whether enough data is available to publish it.

For the educational A6-20, “can I paint on fabric?” remains significant even though suitability is unconfirmed. It can be included on the needs map and referred for clarification. Fabric must not be added as a positive property merely to match the query.

Essential qualifications within a group must remain too. If questions cover cotton, synthetics and washing, the broad heading “for fabric” does not erase those differences. A supported answer about cotton alone does not cover the entire group.

Two errors that need separate counts

A false merge hides differences. Restoring dried material receives an answer about routine thinning; full curing receives the time for surface drying. A generator may then write fluent text while retaining the mistaken task structure.

Excessive splitting causes another problem. Several versions of one question require the same information but are counted as independent items. Answers are repeated, the assignment grows and the completeness score is distorted.

A review needs examples of both errors and borderline pairs. “What can I thin acrylic with?” and “What can I use to dilute acrylic paint?” are candidates for merging when formulation and context match. “What can I thin it with?” and “In what ratio?” are neighbouring but distinct questions. “How do I wash it off a brush before it dries?” and “How do I remove dried paint from fabric?” differ in object and state.

Record the decision together with its grounds: which difference was considered significant, which context was used and what remains unknown. If wording is ambiguous, retaining “clarification needed” is more useful than forcing it into a predefined group.

A test set should not consist only of obvious synonyms. It needs similar-looking questions with different answers and differently worded versions of the same task. An error on one pair does not describe the quality of the whole system; it identifies a case for investigation and further testing.

Retain the question with its boundaries

After grouping, the original strings must remain retrievable. Useful group information includes the main question, essential conditions, applicability, excluded cases and example formulations. A broad topic is stored separately from the questions against which an answer will be assessed.

Merging queries merely to shorten the list does not establish the result’s usefulness. Review must show that the differences do not change the required information or hide restrictions. In some cases, splitting is the correct result rather than a more compact tree.

An accepted group has a recorded shared question; disputed formulations remain candidates pending clarification. Once questions about thinner, ratio and restoring dried material are separated, each has its own data requirements. We can now ask which answers can be supported for the selected paint.