Resolving Structural Ambiguity with Semantic Relation Taxonomies

15 concepts13 papers1 proposal

Natural language is highly compressed. A single grammatical construction can mask completely different underlying meanings. Consider the English possessive: "John’s car" denotes ownership, "John’s arm" denotes a part-whole relationship, and "John’s story" could mean authorship. Syntactically, these phrases are identical.

To build systems capable of deep reasoning—whether solving mathematical word problems or drawing maps from textual descriptions—computational models cannot rely on surface syntax alone. Instead, a long-standing research direction focuses on mapping uniform grammatical constructions to explicit semantic relation taxonomies through multi-stage parsing pipelines. By decoupling syntactic connectivity from semantic meaning, this approach provides a structured framework for machine understanding.

The Strategy: Taxonomies as Bridges to Meaning

Why does this approach work? Instead of attempting to jump directly from a raw sentence to a complex logical formula, this paradigm introduces a structured intermediate layer. By defining a finite, discrete taxonomy of semantic relations, researchers can isolate and solve specific ambiguities.

For instance, Tratz et al. (2013) tackled the English possessive by establishing a 17-category taxonomy encompassing relations like kinship, part-whole, and subjective roles. This formal framework allowed machine learning models to classify the "hidden" meaning of the 's marker with high accuracy.

Similarly, Ravin (1990) targeted the highly ambiguous preposition "with" in verb phrases (e.g., "to cut with a knife" vs. "to run with haste"). By categorizing the preposition into five distinct semantic relations—such as instrument use, manner of action, state alteration, and co-agency—the system could systematically distinguish between adjuncts (determined by the noun) and complements (dictated by the governing verb).

This challenge is cross-linguistic. In Japanese, the highly versatile adnominal particle no creates identical structural ambiguities in noun-noun phrases ("A no B"). Shimazu et al. (1987) resolved this by treating nouns as functional "functors" that take other nouns as arguments, mapping the no construction into a structured taxonomy of predicates, arguments, attributes, and case roles.

By formalizing these hidden relationships, systems can isolate core semantic structures from auxiliary linguistic elements—a principle that Nivre and Fang (2017) demonstrated is essential for evaluating cross-linguistic dependency parsers fairly via content-word relations.

The Mechanism: Multi-Stage Parsing Pipelines

Mapping surface text to these semantic taxonomies is too complex for a single monolithic step. Instead, successful systems employ multi-stage parsing pipelines or cascades that progressively strip away surface variation, isolating specific linguistic challenges before final semantic validation.

This multi-pass architecture dates back to foundational natural language processing systems. Schank and Tesler (1969) proposed a Conceptual Dependency parser that handled lexical lookup, syntactic rule inversion, idiom elimination, and abstract noun rewriting sequentially. This structured decomposition ensured that syntactic constraints were resolved before the system attempted to build an abstract meaning network.

For problem-solving systems, this decomposition is vital for translating language into execution. Gelb (1971) built a multi-pass pipeline that used transformational grammar to standardize text and break down complex sentences into simple declarative forms before applying a formal grammar to generate a hierarchical semantic representation of events.

Modern iterations of this idea scale up the complexity of the transformations to handle multi-sentence discourse. Hopkins et al. (2017) designed a system to solve Math SAT problems using a cascade of tree transducers. By passing structural hypotheses through successive layers of syntactic and semantic translation, the system could resolve complex discourse anomalies and anaphora (such as tracking what "the numbers" refers to across sentences), translating natural language into formal mathematical code.

Surface Text 
     │
     ▼
[Syntactic Parsing & Standardization] ──► Identifies structural connections
     │
     ▼
[Lexical & Semantic Categorization]   ──► Consults dictionaries/word classes
     │
     ▼
[Relation Taxonomy Mapping]           ──► Resolves "hidden" roles (e.g., Instrument, Kinship)
     │
     ▼
Formal Semantic Representation        ──► Ready for downstream reasoning/execution

The Engine: Bootstrapping with Lexical Knowledge

How do these multi-stage systems determine which specific relation in a taxonomy applies to a given phrase? They rely on the relational properties and semantic classes of the constituent words.

As von Glasersfeld (1969) argued, traditional intrinsic lexical definitions are insufficient for machine understanding; instead, word classes must be derived by analyzing how words participate in relational structures and syntactic constructions. This insight is echoed in Preprint No. 22 of the 1969 COLING conference, which proposed classifying words by integrating distributional semantics with their typical usage patterns to infer meaning and function. Even fundamental semantic properties like synonymy and antonymy can be organized into axiomatic equivalence classes to systematically structure these underlying lexical databases, as shown by Edmundson and Epstein (1969).

When explicit semantic classes are missing, systems can dynamically extract them by leveraging machine-readable lexical resources. Kurohashi and Sakai (1999) solved the ambiguity of Japanese noun phrases by automatically parsing the definitions of nouns within a children's dictionary. Because children's dictionaries use a constrained vocabulary, the system could easily extract functional links (such as agency or composition) to determine the relationship between a modifier and a head noun.

Similarly, in spatial reasoning, a system designed for interpreting Japanese street directions decomposed sentences by matching verbs against predefined dictionaries of case relations (such as origin, goal, direction, and distance). This proved that anchoring text to structured lexical properties is the key to bridging the gap between surface language and downstream spatial or logical representations.

Key Foundational Work

Semantics and the Syntactic Classification of Words (1969) — contributed 2 concept elements to this direction.
Computer Aided Research on Synonymy and Antonymy (1969) — contributed 1 concept elements to this direction.
International Conference on Computational Linguistics COLING 1969: Preprint No. 22 (1969) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_semantic_relation_taxonomy_f2ea420a,
  title = {semantic relation taxonomy},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/semantic-relation-taxonomy/rdr_f2ea420a95bc48a8817701974306a10c},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_semantic_relation_taxonomy_f2ea420a_v1,
  title = {semantic relation taxonomy},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/semantic-relation-taxonomy/rdr_f2ea420a95bc48a8817701974306a10c/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake