Unified Meaning Representation (UMR)

• UMR is a graph-based semantic representation that models both sentence-level and document-level meaning using nodes, edges, and explicit attributes.
• It extends AMR by integrating morphosyntactic and pragmatic features along with document-level discourse relations, aiding multilingual and low-resource applications.
• UMR parsing leverages approaches such as fine-tuned Transformer models and UD-bootstrapping, achieving high metrics (e.g., SMATCH ~90) on standard corpora.

Unified Meaning Representation (UMR) is a graph-based, cross-linguistic semantic formalism engineered to capture both sentence-level and document-level meaning in a uniform, annotation-flexible schema adaptable to the world’s languages, including extremely low-resource and morphologically complex ones. UMR extends the structural principles of Abstract Meaning Representation (AMR) and introduces key innovations for addressing multilinguality, document context, and explicit encoding of morphosyntactic and pragmatic attributes (Markle et al., 17 Feb 2025, Markle et al., 8 Dec 2025, Wein, 13 Feb 2025).

1. Formal Definition and Graph Structure

A UMR graph for a sentence $S$ is a tuple $G_S = (V, E, \ell_V, \ell_E)$:

• $V$: set of nodes, each representing a concept (predicate, entity, attribute, or reference)
• $E \subseteq V \times V$: set of directed edges
• $\ell_V: V \to \Sigma_V$: node labeling function (maps nodes to concept vocabulary, e.g., “buy-01”, “person”, “aspect”)
• $\ell_E: E \to \Sigma_R$: edge labeling function (maps edges to semantic role labels; e.g., ":ARG0", ":aspect", ":modstr")

UMR supports reentrancy for coreference and shared argument structures. Node-level attributes (e.g., :refer-person, :refer-number, :aspect, :modstr) capture inflectional, referential, and modal/pragmatic features. UMR graphs are generally represented in PENMAN notation, as in:

(s / buy-01 :ARG0 (p / person :refer-person 3rd :refer-number Plural) :ARG1 (c / car :ARG1-of (n / new-01) :refer-number Singular) :aspect Activity :modstr FullAff)

(Wein, 13 Feb 2025, Markle et al., 8 Dec 2025)

At the document level, UMR introduces a second graph layer that connects sentence-level graphs with cross-sentential relations: coreference arcs, temporal chains, modality, and alignment annotations (Markle et al., 17 Feb 2025).

2. Extensions Over AMR and Universal Dependencies

UMR generalizes and supersedes AMR in several respects:

• Multilingual flexibility: UMR employs a lattice-style taxonomy enabling both coarse- and fine-grained annotation stages, eliminating the bespoke rolesets required for each language in AMR. It is directly applicable to low-resource and typologically diverse languages (Markle et al., 17 Feb 2025, Markle et al., 8 Dec 2025).
• Morphosyntactic and pragmatic scope: Attributes for aspect, modality, referential number/person are first-class nodes, whereas AMR typically omits or externalizes these categories.
• Document-level semantics: UMR explicitly encodes cross-sentence discourse phenomena, such as coreference, discourse connectives, and temporal structure, which AMR lacks (Markle et al., 17 Feb 2025).
• Compatibility with syntactic frameworks: While Universal Dependencies (UD) provides cross-lingual syntactic trees, UD lacks explicit modal, scope, and multi-sentence semantic information. UMR can be bootstrapped from UD parses while enriching them with higher-order semantics (Markle et al., 8 Dec 2025).

3. Annotation Workflow and Data Resources

UMR annotation is staged to maximize transferability and minimize language-specific constraints:

• Stage 0: Roleset creation/extension for low-resource languages
• Stage 1: Span identification and mapping to concept inventory
• Stage 2: Argument edge assignment and role labeling
• Stage 3: Node attribute attachment

The largest available corpus is UMR v2.0, encompassing >210,000 sentences and 8 languages, including English, Chinese, and several indigenous languages. Data normalization includes alignment with dialogue tags, tokenization for UD/PENMAN interoperability, and linearization for sequence models (Markle et al., 8 Dec 2025).

4. UMR Parsing and Generation Methodologies

4.1 Text-to-UMR Parsing

Two principal methodologies are established for English text-to-UMR parsing (Markle et al., 8 Dec 2025):

• Fine-tuning AMR parsers: Existing Transformer encoder–decoder models (e.g., BiBL, AMRBART, amrlib) are extended to output UMR vocabulary and graph linearizations, with modifications for UMR roles and node attributes. SETUP (a fine-tuned BiBL) attains AnCast=84.35, SMATCH=88.82, and SMATCH++=90.98 on the English UMR v2.0 test set.
• UD-bootstrapped + Completion: UD parses are converted to partial UMR graphs by rules, then completed to full UMR by a Transformer which contextualizes both the sentence and its partial graph.

4.2 UMR-to-Text Generation

Three paradigms (Markle et al., 17 Feb 2025):

• Pipeline (UMR → AMR → Text): Converts UMR into AMR and leverages mature AMR-to-text models, discarding UMR-specific features. SMATCH=0.63 on held-out English UMRs.
• Fine-tuning pretrained LLMs: Models like mT5, mBART, and Gemma are fine-tuned end-to-end on UMR-linearized graphs paired with reference texts. Data is limited to a few hundred UMR graphs per language.
• Fine-tuning AMR-to-text models on UMR directly: Existing models retain their graphical encoders and are fine-tuned with UMR graphs, achieving English BERTScore=0.825 and Chinese BERTScore=0.882.

Table: Representative Parsing Evaluation (English, v2.0)

Model	AnCast	SMATCH	SMATCH++
BiBL (SETUP)	84.35	88.82	90.98
AMRBART	81.70	86.24	88.70
amrlib (T5)	79.39	85.24	88.84

(Markle et al., 8 Dec 2025)

5. Downstream NLP Applications

5.1 Low-Resource Machine Translation

Incorporating UMR graphs into prompts for translation by GPT-4 measurably improves translation accuracy for Arapaho, Navajo, and Kukama (Wein, 13 Feb 2025). In an experimental regime evaluating four prompting protocols (Zero-Shot, ZS+UMR, Five-Shot, FS+UMR), the addition of gold UMR in both zero- and five-shot paradigms yields consistent, statistically significant gains, especially in chrF and BERTScore metrics.

Protocol	Arapaho chrF	Kukama chrF	Navajo chrF
ZS	12.97	13.96	15.36
ZS+UMR	16.20	16.82	17.91
FS	32.91	40.82	24.61
FS+UMR	35.67	43.54	25.87

Statistical testing confirms significance (paired t-test, $p<0.05$ in 9/12 comparisons), demonstrating the utility of UMR as a semantic scaffold for extremely low-resource MT.

5.2 Aspect-Category Sentiment Analysis (ACSA)

UMR is embedded as an intermediary in Chain-of-Thought (CoT) prompting for ACSA, where the LLM is guided through UMR parsing, aspect/opinion extraction, category mapping, and sentiment assignment (Ventirozos et al., 22 Dec 2025). On complex datasets (e.g., Laptop16, 67 classes), Qwen3-8B shows micro-F1 improvements of 14.2 percentage points by leveraging UMR-based CoT prompts over standard CoT. However, the method’s effectiveness is model- and dataset-dependent, with no significant effect in a three-way ANOVA ($F(1,6)=0.42, p=0.543$), and occasional performance drops (e.g., Gemini-2.5-Pro).

6. Strengths, Limitations, and Evaluation Metrics

UMR delivers explicit, principled semantic abstraction, multilingual generality, and is data-efficient (state-of-the-art generation and parsing with as few as ~100–200 training graphs in high-resource languages) (Markle et al., 17 Feb 2025, Markle et al., 8 Dec 2025). Evaluation employs metrics such as SMATCH, SMATCH++, AnCast, AnCast++, BERTScore, BLEU, and METEOR, with AnCast prioritizing accurate alignment of semantic roles, modality, and coreference.

Reported weaknesses include:

• Annotation data scarcity, especially in low-resource domains and for document-level phenomena
• Some pipeline approaches lose UMR-specific information (aspect, modality) in AMR conversion (Markle et al., 17 Feb 2025)
• Prompt-based UMR usage is limited by few available in-context annotated instances (Ventirozos et al., 22 Dec 2025)
• Model-dependent effectiveness in downstream reasoning and generation tasks

7. Future Directions

Current challenges and research priorities include:

• Expansion of multilingual, domain-diverse UMR annotation corpora (Markle et al., 8 Dec 2025)
• Development of dedicated UMR-to-text parsers/generators that natively capture document-level structure and all UMR-specific attributes
• Investigation of joint or neuro-symbolic approaches that unify parsing, representation, and generation, inherently handling UMR’s multi-level graph architecture
• Improved evaluation protocols, especially for low-resource and indigenous languages (Markle et al., 17 Feb 2025, Wein, 13 Feb 2025)

These advances are expected to further position UMR as a universal, semantically interoperable representation for multilingual NLP, low-resource technologies, and linguistically informed reasoning.