Reconstructing Surface Details from Underspecified Semantic Graphs

This direction is split into other directions: semantic underspecification. Its last version stays readable and citable here.

Version 1 of 121 papers28 conceptsfrozen 2026-10-04

In natural language generation (NLG), the goal is often to transform a computer-readable "meaning" into a human-readable sentence. However, the most popular semantic formalisms—such as Abstract Meaning Representation (AMR)—are intentionally "lossy." They focus on core concepts (who did what to whom) while stripping away "surface" details like tense, articles, and function words.

This creates a fundamental challenge: underspecification. A generator cannot simply translate the graph; it must act as an inference engine to reconstruct the linguistic scaffolding that the semantic representation left behind.

The Gap Between Meaning and Surface

Abstract semantic representations are designed to be concise. For example, in an AMR graph, the concepts for "the boy," "a boy," and "boys" might all be represented by the same node: boy. Similarly, the distinction between "is eating," "ate," and "will eat" is often collapsed into a single predicate.

As Chengqing Zong, Wenjie Cao, and Bo Xu 0002 note in Approach to interchange-format based Chinese generation, this abstraction forces the generation component to perform complex inference. To produce a fluent sentence, the system must deduce missing attributes like mood, tense, and number from the surrounding context or domain knowledge. Timo Schick highlights in Transition-Based Generation from Abstract Meaning Representations that this gap is the primary hurdle to achieving human-level fluency from semantic inputs.

Generation as Inference and Optimization

If the input is missing information, how does a model decide which words to add? Researchers have approached this as a problem of "filling in the blanks" through various mechanisms:

1. Rule-Based and Hybrid Inference

Early systems, such as those described by Chengqing Zong and Hua Wu in Chinese Generation in a Spoken Dialogue Translation System, used explicit domain knowledge. If a semantic representation in a hotel booking domain was missing a tense, the system could infer it based on the dialogue state (e.g., a request for a future room is likely in the future tense).

2. Combinatorial Optimization

Rather than just predicting the next word, some models treat generation as a global search. Linfeng Song and Yue Zhang reformulated the task in AMR-to-text generation as a Traveling Salesman Problem. By partitioning the semantic graph into fragments, they used a TSP solver to find the most "efficient" (grammatical) sequence of translations, effectively letting the optimization process "discover" where function words and prepositions should live.

3. Transition-Based Transformations

Another approach is to treat generation as a series of incremental actions. In Transition-Based Deep Input Linearization, Ratish Puduppully and Yue Zhang 0004 show that a system can jointly handle word ordering and "morphological inflection." As the model moves through the graph, it triggers actions to insert the necessary function words and adjust word endings (like changing "eat" to "eaten") to satisfy grammatical constraints.

Factorizing Meaning and Syntax

A significant breakthrough in handling underspecification involves separating the what from the how. In Factorising AMR generation through syntax, Kris Cao and Stephen Clark argue that because a single semantic graph can correspond to many valid sentences, the model should first generate a "syntactic skeleton" (a constituency tree). This skeleton provides the slots for tense and function words, which are then filled with content from the semantic graph. This factorization allows the model to explore different phrasings for the same abstract meaning.

The Alternative: Richer Representations

While much work focuses on "fixing" underspecified graphs during generation, some researchers argue for using more detailed semantic formats from the start.

Valerie Hajdik and Emily M. Bender demonstrated in Neural Text Generation from Rich Semantic Representations that using Minimal Recursion Semantics (MRS)—which preserves tense and number—leads to much higher accuracy (BLEU scores over 80) compared to AMR. Similarly, Abelardo Carlos Martinez Lorenzo and Roberto Navigli introduced the BabelNet Meaning Representation (BMR), which explicitly encodes tense and aspect to avoid the "guessing game" required by more abstract formats.

Summary

Generating text from abstract meaning is not a simple translation; it is a reconstruction. Whether through auxiliary graph-reconstruction tasks (Tianming Wang et al., Better AMR-To-Text Generation with Graph Structure Reconstruction) or by using the original source text to "nudge" the generator (Hardy and Andreas Vlachos 0001, Guided Neural Language Generation), the core of the research direction is bridging the gap between sparse semantic concepts and the rich, redundant details of natural language.

Papers in this version

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_reconstructing_surface_details_from_unde_d4a5ee23,
  title = {Reconstructing Surface Details from Underspecified Semantic Graphs},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/reconstructing-surface-details-from-underspecified-semantic-graphs/rdr_d4a5ee23642d44b6b5611fb1862e3371},
  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_reconstructing_surface_details_from_unde_d4a5ee23_v1,
  title = {Reconstructing Surface Details from Underspecified Semantic Graphs},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/reconstructing-surface-details-from-underspecified-semantic-graphs/rdr_d4a5ee23642d44b6b5611fb1862e3371/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}