Same words, different meaning

What remains after splitting text into words?

Analysis begins by deciding boundaries. The string contains words, numbers, multiplication signs, a unit and punctuation. These can be identified as separate elements — tokens. Segmentation depends on the tool’s rules: a hyphen serves different purposes in an ordinary word, the A6-20 product code and a numerical range.

Tokenisation provides addressable spans of the original string. We can now show where a number came from, which span was identified as a model and where negation occurred. But the sequence of elements alone does not tell us what was measured.

Another operation maps word forms to lemmas. In the Russian example, “колёсами” and “колёс” are inflected forms of “колесо” (“wheel”). Linking them to that lemma helps find related expressions without listing every form manually. The prepositions “с” (“with”) and “без” (“without”) must remain available: they change the relationship to the object.

The spaCy documentation describes tokenisation, lemmatisation and dependency parsing separately. These are distinct operations in a particular library, not a mandatory universal pipeline for every system. For our task, their results are best kept separate: normalised forms for matching, and original text and relationships for claim verification.

Names and codes need special handling. “Model 8 Pro” must not become “Model 8” during normalisation if Pro identifies another variant. Nor should a material name be replaced with a broader concept without checking. Otherwise, processing intended to aid comparison will start merging different objects.

Which relationships need to be reconstructed?

In the example “The suitcase dimensions were measured with the wheels installed: 55 × 40 × 20 cm”, we need to identify at least the object, parameter, numerical value, unit and measurement method. The numbers describe dimensions, the dimensions belong to the suitcase, and “with the wheels installed” specifies the configuration measured. The shorter “suitcase with wheels” could merely name a product type, which is why the main example states the measurement method explicitly.

Dependency parsing establishes relationships between words. In Universal Dependencies, a function word marking a relationship, such as a preposition, may be linked to the corresponding noun. But a syntactic label does not itself mean “measurement condition”. That domain role must be determined from the construction and the task.

This record allows precise questions to be put to the data. Are handles included? Which number is height, width or depth? Three unlabelled values do not justify automatically assigning axes. If this matters for comparison, a source clarification or measurement diagram is needed.

Multiword expressions also matter. Russian “ручная кладь” means “carry-on baggage”, a transport context, rather than two independent properties corresponding to “ручной” (“hand” or “manual”) and “кладь” (“baggage”). In another category, “сухой остаток” (“dry residue” or “solids content”) names a composition parameter, not an instruction to dry some leftovers. A domain glossary helps detect these constructions, but using it still requires context.

Where negation applies

“Suitable for fabric” and “not suitable for fabric” contain almost the same words. The difference is decisive when choosing a product. During text transformation, negation must remain linked to the property it modifies.

The task is more complex than finding “not”. “Mass no more than 2 kg” and “mass no less than 2 kg” contain the same object, number and unit but set opposite range limits. Both cannot be stored simply as “mass: 2 kg”.

In “does not require preliminary priming”, negation concerns the necessity of an action. That is a different claim from “not suitable for a primed surface”. The shared root relating to priming does not make them interchangeable.

Even negative advice is not always a prohibition: “not recommended” must not be strengthened to “prohibited” without clarification. Meaning can also be lost in the other direction. If a source says “not permitted”, the softer “best avoided” changes the force of the restriction.

In UD v2, negation is represented by the feature Polarity=Neg alongside an ordinary dependency relation; the separate universal neg relation from the previous version was removed. The feature helps retain linguistic information. Scope and domain meaning still require analysis of the construction, while the validity of the prohibition itself requires checking the source.

A condition cannot be separated from its result

The first article showed how washing brushes before the paint dries became a general promise about washing paint off. Linguistically, the object being cleaned and the time boundary were lost. Retaining “water” and “paint” is not enough for such an instruction.

Time may relate to different events: starting work, surface drying, full curing or permitted storage. The same “after 30 minutes”, without a link to an event, is not a ready-made answer to “when can the item be used?”

Numerical restrictions must be read in full. “Up to 30 minutes”, “at least 30 minutes” and “about 30 minutes” differ in their limits and certainty. Retaining the number does not necessarily retain the recommendation.

Words expressing possibility, necessity and conditions matter too. “Can be wall-mounted using bracket X” does not establish that the bracket is included. “Bracket X is required” does not mean that any bracket with the same appearance will work. Every additional relationship needs evidence.

During extraction, a condition should remain beside the action it qualifies. A separate general list of restrictions may aid retrieval, but without links it becomes unclear what each restriction applies to. If a paragraph describes two operating modes, a condition for one must not automatically be applied to the other.

What does the Russian pronoun “её” refer to?

“Кружка поставляется с крышкой. Её можно мыть в посудомоечной машине” — “The mug comes with a lid. It can be washed in a dishwasher.” In Russian, both “кружка” (mug) and “крышка” (lid) are feminine nouns, so the feminine pronoun “её” can refer to either. The nearest noun offers a candidate link, but does not guarantee that the author meant the lid.

The distinction matters for the working dossier: permission for the lid does not establish permission for the mug, or vice versa. Without more context, the ambiguity should be retained or clarification requested. Substituting the most convenient object adds meaning that the text does not convey unambiguously.

Other constructions can be clearer. “The set contains two containers. Each holds 500 ml” links the volume to one container. A total of one litre follows from a separate calculation. If the pronoun link is not resolved, 500 ml may be recorded incorrectly as the capacity of the whole set.

Labels, list structure, the neighbouring paragraph and domain consistency can help. The surrounding text may itself contain an error, however. An analyser needs to distinguish “this is the likely textual link” from “this property is supported for the product”.

A record that makes the analysis checkable

After processing, it must still be possible to return to the original phrase. A working representation for the suitcase example might look like this:

This is a proposed domain structure, not a universal data schema. Clothing requires other parameters; consumables need application and consumption conditions. The shared requirement is to preserve relationships that affect the answer.

One way to detect an error is to reconstruct a claim from the extracted data and compare it with the source. Check separately that the object, parameter and condition are linked correctly. Reconstructing “external dimensions including wheels and handles” would already add handles here. “Suitable as carry-on baggage in all cases” would add universal compatibility too.

Review needs to cover both omissions and added relationships. A system may identify every word correctly yet attach a condition to another object. An analysis of a specific discrepancy is therefore useful: original passage, resulting relationship and required clarification.

“What are the dimensions including wheels?” and the label “dimensions excluding protruding parts” concern a closely related topic. Before using the label as an answer, we need to establish which parts the measurement excludes. This determines whether the formulations can be grouped together or require separate answers.