Unifying Fragmented Semantic Formalisms for Graph Parsing

84 concepts70 papers1 proposal

While syntactic parsing has largely converged on unified standards like Universal Dependencies, semantic parsing has historically remained deeply fragmented. Different linguistic schools define sentence meaning using fundamentally incompatible graph structures. This "balkanization" of computational semantics means that a model trained to understand "who did what to whom" in one framework cannot easily leverage data from another, severely limiting cross-framework generalization.

To overcome this fragmentation, recent research has shifted toward Cross-Framework Meaning Representation Parsing (MRP). By mapping disparate linguistic formalisms into a unified graph abstraction, researchers have designed architectures capable of learning universal semantic representations from disjoint data.

The Challenge of Semantic Structural Divergence

The primary obstacle to a unified semantic parser is the sheer structural diversity of existing frameworks. As outlined by Stephan Oepen et al. in the SemEval 2014 Task 8 roadmap, frameworks diverge significantly in how they represent core linguistic phenomena like coordination, compounding, and scope.

This divergence typically manifests along two major axes:

  1. Anchoring: Bi-lexical dependency frameworks (such as DM or PSD) anchor every semantic node directly to surface words in the text. Conversely, conceptual abstractions (such as Abstract Meaning Representation (AMR) or UCCA) utilize unanchored nodes to represent abstract concepts that lack a 1:1 textual mapping, as noted by Daniel Hershcovich et al. in their work on TUPA.
  2. Spans vs. Dependencies: Semantic Role Labeling (SRL) is traditionally split down the middle. Some corpora annotate full phrase spans, while others identify single head-word dependencies (Rocco Tripodi et al., UniteD-SRL).

When neural models are trained in isolation on these individual frameworks, they become captive to the specific theoretical lens of the dataset. Probing experiments by Ilia Kuznetsov et al. in A matter of framing demonstrate that a model's internal representations can appear highly syntactic or purely semantic depending entirely on the framework used to train and evaluate it.

The Common Denominator: Unified Graph Abstractions

To resolve this, the community established the Meaning Representation Parsing (MRP) shared tasks (Omri Abend, MRP 2019). The core insight of MRP is that despite theoretical disagreements, nearly all semantic formalisms can be mathematically generalized into a single directed graph structure.

In this unified view, parsing becomes a task of predicting nodes (semantic predicates or concepts) and directed edges (the relationships between them). Frameworks like the one proposed by Gaku Morio et al. formalize this via an interactive "Plain Graph Notation" (PGN), which decouples the underlying parsing engine from framework-specific engineering (i-Parser).

[Surface Text] ──> [Contextual Encoder (BERT/BiLSTM)] ──> [Unified Graph Representation]
                                                                │
                                        ┌───────────────────────┼───────────────────────┐
                                        ▼                       ▼                       ▼
                                  [DM Graph]              [UCCA Graph]            [AMR Graph]
                              (Word-Anchored)          (Phrase-Structure)        (Abstractive)

Multi-Task Parsing Mechanisms

With a unified graph abstraction in place, researchers can design architectures that parse multiple semantic frameworks simultaneously. Two main algorithmic paradigms make this work: transition-based parsing and graph-based parsing.

Transition-Based Unification

Transition-based parsers convert graph generation into a sequence of state-machine actions (such as shifting words onto a stack, reducing them, or creating edges). To handle cross-framework parsing, systems like TUPA (Daniel Hershcovich et al.) generalize these transition actions so they can construct arbitrary graph topologies, whether anchored or unanchored.

To make this computationally efficient, Wanxiang Che et al. introduced an "operation buffer" technique that allows parallel training of traditionally sequential stack operations (HIT-SCIR at MRP 2019). When frameworks diverge too drastically, hybrid transition systems deploy specialized sub-routines: a transition-based parser for text-anchored structures and an iterative graph-sequence inference model for abstract conceptual frameworks (Longxu Dou et al., HIT-SCIR at MRP 2020).

Graph-Based Biaffine Attention

Rather than building a graph step-by-step, graph-based parsers predict the global existence of nodes and edges directly. This approach relies heavily on deep contextualized embeddings (like BERT) paired with biaffine attention layers.

As demonstrated by Seung-Hoon Na et al., "multi-level" biaffine attention fuses representations from both shallow and deep layers of an encoder (JBNU at MRP 2019). This architecture works because lower layers preserve the token alignments needed for anchored frameworks (like DM), while higher layers capture the abstract semantic relationships required by frameworks like UCCA or AMR. Yue Zhang et al. proved that this direct edge-prediction strategy allows a single multi-task network to outperform isolated, framework-specific models by exploiting shared underlying linguistic structures (SUDA-Alibaba at MRP 2019).

Why It Works: Compositional Normalization

Multi-task learning on fragmented semantic data is notoriously difficult because frameworks often use conflicting conventions for functional words (like auxiliaries, prepositions, or determiners). Why, then, do unified graph parsers succeed?

The deeper intuition lies in compositional normalization. In Normalizing Compositional Structures Across Graphbanks, Jonas Groschwitz demonstrated that by using a formal graph algebra (the AM+ algebra), the underlying compositional structures of seemingly disparate graphbanks can be mathematically aligned. This normalization reveals that many cross-framework differences are superficial design choices rather than fundamental linguistic disagreements.

This alignment mirrors historical breakthroughs in computational semantics, where formal proofs bridged competing underspecification formalisms. For example, Joachim Niehren established a quadratic-time translation between Minimal Recursion Semantics (MRS) and Dominance Constraints (Bridging the Gap Between Underspecification Formalisms), while Alexander Koller et al. proved that Hole Semantics could be solved using dominance constraint algorithms (Underspecification formalisms).

By exposing these shared algebraic and logical foundations, modern joint parsers can successfully bridge formalisms. As shown by Hao Peng et al., treating missing annotations as latent variables allows models to learn a unified representation of meaning even when trained on completely disjoint sentences with entirely different formalisms (Learning Joint Semantic Parsers from Disjoint Data).

Key Foundational Work

Case Structure Systems (1973) — contributed 3 concept elements to this direction.
On Logical Formulation of the Computation Process in Semantical Systems (1973) — contributed 1 concept elements to this direction.
Semantic for Text Processing (1980) — contributed 1 concept elements to this direction.
History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 72 papers
  2. Version 1 · 2026-10-04 · 96 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_semantic_framework_fragmentation_f3c9ea3b,
  title = {semantic framework fragmentation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/semantic-framework-fragmentation/rdr_f3c9ea3b024f406d8e69d429825e688c},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 2, frozen text):

@misc{lacuna_semantic_framework_fragmentation_f3c9ea3b_v2,
  title = {semantic framework fragmentation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/semantic-framework-fragmentation/rdr_f3c9ea3b024f406d8e69d429825e688c/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake