Generating Text from Underspecified Semantic Representations

16 concepts10 papers2 proposals

When we translate human thought into a formal semantic representation—such as an Abstract Meaning Representation (AMR) graph or a Discourse Representation Structure—we intentionally strip away the surface-level idiosyncrasies of natural language. We discard language-specific word order, omit articles ("the", "a"), drop prepositions, and often ignore grammatical features like tense, number, and mood. The goal is a clean, language-agnostic representation of meaning.

However, when a Natural Language Generation (NLG) system attempts to turn this abstract meaning back into text, it encounters a fundamental hurdle: semantic underspecification. Because the input representation lacks surface-level syntactic details, the generation process is not a simple deterministic mapping. Instead, it is a one-to-many problem where a single semantic graph can correspond to dozens of grammatically distinct, yet valid, surface sentences. To produce fluent text, generation systems cannot just translate; they must act as inference engines that reconstruct the missing linguistic features.

The Anatomy of the Semantic-to-Surface Gap

To understand why this problem requires active inference, consider what is missing from a typical deep semantic representation. As Dyer et al. explain in Generation from Abstract Meaning Representation using Tree Transducers, formalisms like AMR deliberately omit morphological details such as tense and number. A single concept node like boy does not specify whether the output should be "the boy," "a boy," or "boys."

This structural omission creates an inherent ambiguity during the realization process. As Manning notes in A Partially Rule-Based Approach to AMR Generation, the generation system must autonomously determine word order and synthesize missing syntactic elements to build a fluent sentence. If the generator makes an incorrect guess about a missing feature, the output becomes ungrammatical or shifts the original meaning.

[Semantic Graph: boy -> eat -> apple]
                 │
                 ▼ (Underspecification Resolution)
  ┌──────────────────────┼──────────────────────┐
  ▼                      ▼                      ▼
"The boy eats an apple." "Boys ate apples." "A boy will eat the apple."

This gap becomes even more pronounced in machine translation and dialogue systems. When moving between languages, a target language often strictly requires grammatical information that the source text never explicitly encoded. Kameyama et al. term these discrepancies "translation mismatches" in Resolving Translation Mismatches with Information Flow. If the source representation is underspecified, the generator must find a way to inject the missing details required by the target language's grammar.

Resolving Underspecification via Context and Knowledge

How can a system logically deduce features that are completely absent from its input? The answer lies in leveraging context, domain knowledge, and external linguistic priors.

Information Flow and Situation Theory

One foundational approach treats generation as a process of "reactive reasoning" within a multi-layered context. In Resolving Translation Mismatches with Information Flow, Kameyama et al. utilized Situation Theory to model the broader communicative environment, including the discourse state and the speaker's perspective. By tracking how information flows through the dialogue, the system can infer missing details—such as honorifics or implicit pronouns—that are absent from the surface form of the source language but mandatory in the target language.

Domain-Driven Inference

In constrained environments, such as spoken dialogue systems, common-sense reasoning and domain knowledge can fill in the blanks. The work of Zong et al. in Chinese Generation in a Spoken Dialogue Translation System and Approach to interchange-format based Chinese generation demonstrates how a generation pipeline can use a rule-based inference layer to reconstruct logical structures. For instance, if a user's spoken input is fragmented or poorly transcribed, the system can leverage domain-specific predicate-argument frames to deduce the missing tense, mood, or category attributes before the final linguistic realization phase.

Algorithmic Solutions: From Optimization to Graph Encoders

As the field shifted from rule-based systems to statistical and neural models, the mechanisms for resolving underspecification evolved from explicit logical deduction to structural and probabilistic modeling.

Combinatorial Optimization

One way to handle the missing pieces is to frame generation as an optimization problem over semantic fragments. In AMR-to-text generation as a Traveling Salesman Problem, Song et al. reformulated the challenge as an Asymmetric Generalized Traveling Salesman Problem (AGTSP). By partitioning an AMR graph into fragments and finding the optimal sequence to traverse them, the system uses a language model-driven solver to naturally insert the function words, articles, and tenses required to make the sequence fluent, effectively letting a language prior resolve the underspecification.

Structural Graph Encoders

Because deep semantic representations omit function words like prepositions, the model must learn to infer these tokens from the geometric relationships within the input structure. Hong et al. addressed this in Improving Language Generation from Feature-Rich Tree-Structured Data with Relational Graph Convolutional Encoders. By employing Relational Graph Convolutional Networks (R-GCNs), their model encodes the structural context of dependency trees directly. This allows the network to autonomously synthesize necessary grammatical markers based on the structural configuration of the concepts, rather than relying on a flat, linear sequence.

   (Concept: Eat)
    /        \
  :ARG0     :ARG1
  /            \
(Concept: Boy) (Concept: Apple)

  ▲ Neural Graph Encoder captures structural context
  │
[Model infers missing function words: "The", "an", and tense: "-s"]
  │
  ▼
"The boy eats an apple."

Turning Underspecification into a Multilingual Advantage

While underspecification complicates generation, it can also be leveraged as a powerful asset for cross-lingual applications. Because abstract semantic graphs omit language-specific surface constraints, they can serve as a universal pivot. Fan et al. demonstrated this in Multilingual AMR-to-text Generation by training a single "one-to-many" model to generate text in 21 different languages from a single English-centric AMR graph. The model succeeds because it delegates the task of synthesizing language-specific grammatical markers entirely to the generation phase, utilizing cross-lingual embeddings to map the shared abstract semantics to distinct surface realizations.

The Evaluation Dilemma

The one-to-many nature of underspecified generation creates a unique challenge for evaluation. If a semantic representation can be correctly realized in multiple valid ways, standard automated evaluation metrics become unreliable.

As early as 1993, Dorrepaal argued in On the notion of uniqueness that standard discourse theories lacked the granularity to capture how specific linguistic structures impose unique interpretations, leading to semantic asymmetries that confuse generation systems.

Decades later, Manning et al. confirmed this evaluation gap in A Human Evaluation of AMR-to-English Generation Systems. They showed that while automated metrics like BLEU can roughly rank different systems, they fail to capture nuanced errors. Because there is no single "correct" surface form, automated metrics often penalize perfectly valid grammatical variations or miss critical errors like hallucinations and omissions. Human evaluation remains essential to determine whether a system has successfully navigated the ambiguity of an underspecified input to produce an accurate, fluent output.

Key Foundational Work

Resolving Translation Mismatches with Information Flow (1991) — contributed 1 concept elements to this direction.
On the notion of uniqueness (1993) — 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_underspecification_5ad4017b,
  title = {semantic underspecification},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/semantic-underspecification/rdr_5ad4017b898043cea81d46717e0e40e1},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_semantic_underspecification_5ad4017b_v1,
  title = {semantic underspecification},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/semantic-underspecification/rdr_5ad4017b898043cea81d46717e0e40e1/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake