In Submission · 2026

Semantic Atoms for Scalable Text Analysis in Data Analytics in the Era of LLM-Based Agents

An agent-facing systems abstraction for text analysis in the era of LLM-based agents

Modern tables mix text with structured columns, and LLM agents now drive the analysis. TADA asks the systems question underneath: when the executor already controls a symbolic runtime, what semantic interface should the data system expose?

Jinxiang Xie, Zihao Li, Rui Ding, Bartuer Zhou, Shi Han, Dongmei Zhang

Nanjing University | Microsoft Research

Two placements of the semantic boundary. Operator-style interfaces (left) fuse an LLM judgment into each analytical action, leaving no reusable semantic state. TADA (right) confines semantic judgment to two atoms: Tagging (N→N) adds row-aligned fields keyed by rid; Categorization (N→M) creates a detached category relation keyed by cid. Both materialize ordinary table state for deterministic filtering, grouping, ranking, joining, and regression.

Overview

Existing LLM-powered data systems extend query languages with natural-language operators — semantic filters, semantic aggregates, semantic joins. These are semantic–OLAP macros: an LLM judgment welded to one analytical action. That is convenient for a human query author, but unnecessarily coarse for an agent that already knows how to filter, group, join, and fit a model.

TADA proposes a finer boundary built on two semantic atoms over text columns. Tagging is an N→N operation producing typed, row-aligned fields keyed by row id. Categorization is an N→M corpus-level operation creating a detached relation of discovered abstractions. Both outputs are materialized as ordinary table state — inspectable, reusable, and composable with the full symbolic runtime. The LLM is confined to semantic judgment; everything else stays deterministic.

At corpus scale, a Map–Reduce execution harness lets the same two atoms operate beyond a single reliable context window, with a top-k preservation analysis making the approximation boundary explicit.

Key Contributions

Agent-Facing Abstraction

Text analysis inside analytical programs is formulated as an interface problem: two typed, materialized semantic contracts separate probabilistic judgment from symbolic execution.

Scalable Realization

TADA-Agent implements the abstraction with validated, keyed state materialization and a Map–Reduce harness for bounded-context execution, backed by an analysis of top-k category preservation under partitioning.

Artifact-Backed Evaluation

TABench: 26 complete semantic–symbolic workbook requests derived from a literature-grounded taxonomy, probing both the strength and the boundary of the two-atom framework — plus public extraction benchmarks and a scaling study.

Honest Accounting

Missing, unreadable, or protocol-mismatched evidence is retained as limitations rather than converted into favorable scores; repeated runs are distinguished from single-run baselines throughout.

Composing Semantic and Symbolic Steps

Three workloads that outgrow task-specific semantic operators. Orange boxes are semantic LLM steps; teal boxes are deterministic symbolic steps. (a) A row-aligned sentiment field is materialized, filtered symbolically, then feeds corpus-level topic discovery. (b) The same Categorization atom is invoked inside ordinary group-wise control flow. (c) Discovered aspects become row-aligned features entering an ordinary regression.

Public Document-Extraction Benchmarks

All displayed metrics are higher-is-better. (a) On CUAD, TADA-Agent reaches 0.787 average F1, versus 0.753 for DocETL and 0.619 for LOTUS under the same model and evaluator. (b) On BioDEX, the unchanged two-atom interface also leads the reproduced alternatives on RP@5, RP@10, and term recall.

TABench: End-to-End Workbook Outcomes

Query-specific artifact evaluation over 26 complete workbook requests. A Good result must satisfy all mandatory criteria for output structure and analytical content. TADA-Agent produces 21/26 Good outputs (80.8%), compared with 16 for Codex, 15 for Shortcut AI, 14 for Claude in Excel, and 2 for Excel Copilot.

Inside the Agent Loop

TADA atoms inside a ReAct analytical loop

An end-to-end case study: given a natural-language request over a review workbook, the agent interleaves reasoning with three action types. Categorization induces a detached issue taxonomy (N→M), Tagging assigns row-aligned issue fields (N→N), and the deterministic runtime joins, groups, and computes regional refund effects. Each intermediate result is materialized state that the next reasoning turn can inspect.

Scalable Execution

On three approximately 2.6K-row open-vocabulary joint-topic workloads, moving from (n,r) = (60,10) to (500,19) cuts complete-execution LLM calls by 85.3% and wall time by 30.0%, with no observed decrease in the reported global-topic or row-level metrics.

Relation to CAST

The two atoms were introduced in CAST (ACL 2026 Findings) as stable summarization and tagging operations. TADA adopts those implementations and recasts their typed outputs as an agent-facing systems abstraction — proving the compositional properties (sufficiency for scoped TADA workloads, scalable bounded-context execution) that CAST left open.

Status

In submission. A unified cross-system benchmark campaign (all baselines under three-repeat self-consistency with a position-debiased LLM judge and human calibration) is in progress.