Text Reasoning Meaning Representation (TRMR)

• TRMR is a formal framework that decomposes natural language into explicit, compositional reasoning steps using structured graphs.
• It integrates graph-based, trajectory distributional, and neuro-symbolic models to link text spans with clear, operational reasoning processes.
• TRMR enhances transparency in tasks like reading comprehension and natural language inference, though it may underperform on ambiguous cases compared to end-to-end models.

Text Reasoning Meaning Representation (TRMR) is a formal approach for representing, operationalizing, and reasoning about semantic and inferential relations explicitly in natural language understanding systems. TRMR models are designed to bridge the gap between opaque, end-to-end models and rigid symbolic logic by capturing explicit, machine-readable records of the reasoning steps—often structured as graphs or distributional objects—that lead from unstructured text to answers or logical conclusions. The core TRMR paradigm encompasses a family of frameworks spanning graph-based compositional operations, trajectory distributions over LLM generations, knowledge-primitive algebraic signatures, and neuro-symbolic logic derivations.

1. Formal Definition and Structural Elements

TRMR formalizes the decomposition of natural language reasoning into explicit and compositional structures that support transparency and interpretability. In the canonical formulation (Wang et al., 2020), each instance consists of a triple: $$\mathrm{TRMR}(P,Q) = (G_{parse}, G_{ret}, G_{deriv})$$ where:

• $G_{parse}$: An acyclic directed graph representing the recursive decomposition of the question into sub-operations, with nodes as atomic operations (e.g., count, filter, sort) and argument nodes (linked to spans of the question or outputs from other operations).
• $G_{ret}$: A bipartite mapping associating argument nodes with text spans from the passage, serving as an explicit information retrieval plan.
• $G_{deriv}$: An acyclic directed graph detailing how outputs of operations and retrieved text spans combine through the same operation set to produce the final answer.

The TRMR grammar is given in Backus–Naur Form, capturing operation application, span and argument referencing, and the mapping from spans to derivations. Each atomic operation is paired with variable-arity denotational semantics, often in the style of lambda calculus, for precise interpretation.

Examples of primitive operations in the canonical TRMR vocabulary include:

• $\mathrm{filter}(\mathrm{cond}, S)$
• $\mathrm{count}(S)$
• $\mathrm{sum}(S_1, S_2, \ldots)$
• $\mathrm{sort}(\mathrm{order}, S)$
• $\mathrm{time\text{-}span}(\mathrm{event})$
• $\mathrm{more}(S_1, S_2)$

Within this framework, answers are produced by traversing the derivation graph, applying the respective compositional operations as functions over sets or sequences, where each argument and step is explicitly linked to text-span evidence (Wang et al., 2020).

2. TRMR Instantiations: Graphical, Distributional, and Logic-based Models

Several paradigms operationalize TRMR:

• Graph-based TRMR (Wang et al., 2020): Employs explicit parse and derivation DAGs to model reasoning chains for question answering over unstructured text, with strong alignment between spans in question, passage, and intermediate values.
• Trajectory-based Distributional TRMR (Liu et al., 2023): Represents meanings as probability distributions $\mu_x: \mathcal{A}^* \rightarrow [0,1]$ over all text continuations $y$ from an initial string $x$ using an autoregressive LLM $M$. This meaning function enables algebraic inference, including pointwise partial order, meet ($\wedge$), join ($\vee$), and asymmetric semantic relations (e.g., logical entailment and hypernymy).
• Knowledge-driven Algebraic TRMR (Basu et al., 2021): Maps sentences to finite conjunctions of ground atoms over a fixed set of binary/ternary knowledge primitives (e.g., Agent, Theme, Location, Time, Cause), usually derived via rule-based mappings from syntactic parses and external lexicons (e.g., VerbNet), supporting transparent logic-program inference.
• Neuro-symbolic TRMR via AMR and Logic (Feng et al., 2024): Translates sentences to Abstract Meaning Representation (AMR) graphs, then to ground propositional logic, enabling automated reasoning (recognizing entailment/contradiction) using SAT solvers augmented by relaxation and "forgetting" operations.

Thus, TRMR encompasses a diverse class of meaning representations, united by their compositionality and explicit support for intermediate reasoning transparency.

3. Annotation, Datasets, and Evaluation

The R3 benchmark (Wang et al., 2020) exemplifies the graph-based TRMR approach, providing over 60,000 annotated question–answer pairs with gold TRMR graphs. The annotation pipeline involves:

1. Problem Parsing: Annotators select from a catalog of 30+ atomic operations, recursively building parse DAGs.
2. Information Retrieval: Annotators link arguments in the parse to specific spans in the passage.
3. Answer Derivation: Annotators or system auto-generate the derivation DAG, potentially correcting it.

Quality assurance is enforced through initial training and test sets (≥90% exact-match to gold parses), ongoing gold checks, and validation by consensus among independent validators. Metrics such as inter-annotator agreement (Cohen's κ ≈ 0.87) are reported.

The R3 dataset covers a wide spectrum of reasoning phenomena:

Reasoning Type	Examples	Fraction (%)
Count / Sum	Numerical aggregation	46
Filter / Sort	Selection and ordering	30
Comparison	More/Less	12
Temporal	After/Before, event intervals	6
Span / Time-span	Span extraction, duration	6

Explicit annotation of the reasoning chain and input–output relations target both explainability and robust evaluation of compositional reasoning.

4. Algorithmic and Semantic Foundations

TRMR models in all paradigms share a focus on denotational semantics, applying formal mappings from language (or parse/tree/graph) to operations or algebraic objects.

• Lambda-calculus Semantics (Wang et al., 2020):

$$\llbracket \mathrm{count}(S) \rrbracket = |S|,\quad \llbracket \mathrm{filter}(\mathrm{cond}, S) \rrbracket = \{x \in S \mid \mathrm{cond}(x) \}$$

• Distributional Algebra (Liu et al., 2023):

The trajectory distribution $\mu_x(y) = P_M(y|x)$ is manipulated pointwise; for entailment, $\mu_v \leq \mu_u$ iff for all $y$, $\mu_v(y) \leq \mu_u(y)$. Semantic distances and asymmetric relations leverage this structure for zero-shot inference.

• Logic-based Reasoning (Basu et al., 2021, Feng et al., 2024):
  • Mapping sentences to conjunctions of knowledge primitives or to ground logic formulas.
  • Logic programs are constructed to support non-monotonic, goal-directed inference, including support for defaults, negation-as-failure, and coreference resolution.
  • Neuro-symbolic approaches use semantic similarity (via embeddings) and symbolic relaxation in SAT-based theorem proving.

These algorithmic frameworks enable compositional and robust mechanization of text-based reasoning tasks.

5. Applications and Empirical Results

TRMR has been applied to a range of tasks:

• Reading Comprehension QA: R3 (Wang et al., 2020) enables supervised multitask models that jointly predict answers and intermediate graphs, as well as explicit comparison of system-produced rationales with gold graphs.
• Natural Language Inference (NLI): Neuro-symbolic pipelines (Feng et al., 2024) map text to AMR, then propositional logic, applying SAT-based reasoning and relaxation methods, giving fully explainable verdicts on entailment, contradiction, or neutrality—albeit currently with lower accuracy than state-of-the-art neural models.
• Knowledge-based NLU Systems: Knowledge-driven representations (Basu et al., 2021) power systems like SQuARE and StaCACK, achieving 99.9% and 100% accuracy respectively on bAbI-style QA/dialog benchmarks, with step-wise justifications.
• Semantic Similarity and Multimodal Reasoning: Distributional TRMR (Liu et al., 2023) outperforms other prompt-free methods in STS and hypernymy tests, supports cross-modal alignment, and enables general compositional inference.

Empirical results consistently demonstrate that TRMR models yield superior interpretability and explicit traceability, while neural models (for now) tend to outperform TRMR in raw prediction accuracy, particularly for neutral or ambiguous inference categories (Feng et al., 2024).

6. Strengths, Limitations, and Future Directions

Strengths:

• Direct capture of compositional, multi-hop, and set-theoretic reasoning steps.
• Enforcement of text–evidence alignment eliminates “hallucinated” or invalid rationales.
• Transparency: all intermediate inferences and operations are explicitly available for analysis.

Limitations:

• Fixed vocabulary of compositional operations or knowledge primitives can restrict expressivity; advanced logic, hypotheticals, and implicit commonsense are not captured.
• Neural-symbolic relaxations in logic-based TRMR rely on sub-symbolic similarity, which can impact determinacy.
• Generally lag behind end-to-end neural models in accuracy for ambiguous or underspecified tasks (e.g., NLI-neutral cases).

Future Directions:

A plausible implication is further fusion between explicit TRMR graphs and distributional (e.g., trajectory-based) representations, cross-modal alignment, and dynamic extension of operation/core primitive schemas. Potential applications include explainable systems for knowledge base QA, textual entailment, and multimodal reasoning. Bootstrapping end-to-end systems that generate or consume TRMR as a rationale substrate remains a key challenge and target for research.

7. Connections and Comparative Frameworks

TRMR sits at a novel intersection between fully symbolic logic-based approaches and black-box neural models. It is distinguished from traditional semantic parses (e.g., AMR in isolation) by its explicit operational and compositional structure and from latent neural representations by the accessibility and interpretability of its intermediate states.

Comparative approaches include:

• Trajectory-based distributional semantics: meaning as use, modeled via conditional generation distributions.
• Denotational logic mapping: compositional semantic algebra over syntactic frames.
• Neuro-symbolic logic: integrating AMR graph grounding, propositional abstraction, and transformer-based similarity for robust NLI.

The TRMR paradigm thus provides a unifying and extensible framework for explicit, step-wise text reasoning and meaning representation across multiple reasoning and understanding domains (Wang et al., 2020, Liu et al., 2023, Basu et al., 2021, Feng et al., 2024).