Table Textualization Insights

• Table textualization is the process of converting structured tabular data into natural language while preserving key semantics and evidential context.
• Different methods include direct table-to-text, retrieval-oriented conversion, and logical-form based approaches, each optimizing fidelity and explainability.
• Advanced systems like TabT5 and TlT leverage table-aware encoders and intermediate semantics to improve content selection and output quality.

Table textualization denotes the transformation of tabular content into natural-language or text-like representations that preserve enough structure, semantics, or evidence to support generation, retrieval, reasoning, verification, or reading. Recent work spans direct table-to-text generation, rule- or LLM-based conversion of tables into retrieval corpora, region- and cell-level evidence extraction, and neighboring pipelines that reverse the direction by turning long text into structured tables. Taken together, these studies suggest that table textualization is best understood not as a single benchmark, but as a family of structure-to-language transformations whose objectives vary among fidelity, interpretability, and downstream utility (Bao et al., 2018, Deng et al., 2024).

1. Scope and task family

Across papers, “table textualization” is used in more than one operational sense. In the narrow sense, it refers to generating natural-language descriptions from rows, highlighted regions, or entire tables. In a broader systems sense, it refers to converting tables or sub-tables into textual evidence that an LLM can retrieve, align, verify, or reason over more reliably than over raw tabular layout. This suggests a spectrum from full verbalization to selective evidence rendering (Bao et al., 2018, Han et al., 11 Sep 2025).

Mode	Representative papers	Typical output
Direct table-to-text	(Bao et al., 2018, Andrejczuk et al., 2022, Chen et al., 2023)	sentence, summary, description
Retrieval-oriented textualization	(Han et al., 11 Sep 2025, Carballo et al., 2022, Wei et al., 2024)	narrative chunks, embeddings, missing text fields
Selective evidence textualization	(Wu et al., 18 May 2025, Ho et al., 12 Jun 2025, Wang et al., 26 Feb 2026)	regions, cell sets, aligned highlights
Inverse structured summarization	(Deng et al., 2024)	tables derived from long text

A recurrent technical issue is whether tables should be flattened into text at all. Several systems accept linearization as a practical interface to pretrained LLMs, while others argue that linearization obscures row-column relations, harms reasoning, or weakens explainability. The literature therefore alternates between textualization as a solution and structure preservation as a corrective to overly aggressive textualization (Alonso et al., 2023, Wang et al., 13 Jan 2026).

2. From row description to structure-aware generation

An early neural formulation appears in “Table-to-Text: Describing Table Region with Natural Language” (Bao et al., 2018). That work restricts the input to a single row of a regular table, with attributes, cells, and caption, and uses a table-aware encoder-decoder with attention and a flexible copying mechanism. The copy gate lets the model choose between generating from the vocabulary and copying table-specific values such as names and numbers. On the newly introduced WIKITABLETEXT benchmark, which contains 13,318 explanatory sentences for 4,962 tables, Table2Seq++ reached a BLEU-4 score of 38.23; the same line also reported 40.26 on WIKIBIO and 38.85 on SIMPLEQUESTIONS (Bao et al., 2018).

Subsequent work retained the seq2seq framing but made the encoder more table-aware. TabT5 augments T5 with row and column embeddings and table-aware pre-training on 6.2M Wikipedia tables, including denoising and the task-specific “ToTTification” objective (Andrejczuk et al., 2022). On ToTTo test, T5-base scored BLEU 47.10 and PARENT 56.17, while TabT5-base reached BLEU 48.80 and PARENT 57.60; TabT5 with ToTTify reached BLEU 49.20 and PARENT 57.25 (Andrejczuk et al., 2022). TASD takes a different route, adapting GPT-2 with a three-layered multi-head attention network for table structure understanding and a multi-pass decoder for deliberative refinement; on NumericNLG and ToTTo it improved BLEU, METEOR, and ROUGE-L over fine-tuned GPT-2 and other baselines (Chen et al., 2023).

A separate design axis concerns discourse form rather than encoder structure. Tabular-TX focuses on highlighted-cell interpretation in Korean tabular data and generates summaries in a Theme-Explanation structure, where the Theme Part is derived from title-like context and the Explanation Part verbalizes the relation among highlighted cells (Kwack et al., 17 Jan 2025). Its preprocessing converts tables into key-value form, duplicates merged cells across rowspan and colspan spans, and filters to highlighted cells plus directly related headers. On the reported benchmark, EXAONE 3.0 7.8B with Tabular-TX achieved an average score of 0.45, compared with 0.12 under plain in-context learning and 0.17 with LoRA (Kwack et al., 17 Jan 2025).

3. Beyond plain linearization

A major line of research argues that high-quality textualization requires explicit intermediate semantics. TlT, short for Table-to-Logic-to-Text, inserts an executable logical form between the table and the generated statement (Alonso et al., 2023). Its Table2Logic stage predicts a tree-structured logical form, and its Logic2Text stage verbalizes that form. The paper’s most consequential result is on human fidelity: a direct generator with content selection achieved 44.9% faithful outputs, whereas TlT reached 75.0%, and TlT with gold logical forms reached 82.4%; annotator agreement was reported as Fleiss’ $\kappa = 0.84$ (Alonso et al., 2023). The authors further argue that content selection is necessary but not sufficient, and that logical-form structure itself contributes materially to fidelity.

Another response to linearization is to abandon text as the primary input representation. PixT3 renders tables as images and treats table-to-text generation as an image-to-text problem (Alonso et al., 2023). Built on Pix2Struct, it uses a structure-learning curriculum over 135,400 synthetic tables and a visual encoder with a budget of 2,048 patches of $16 \times 16$ pixels each; the reported optimal down-scaling factor is $\gamma = 0.39$ (Alonso et al., 2023). The paper distinguishes TControl, LControl, and OpenE settings and reports that PixT3 is strongest in LControl, where the full table and highlighted cells are both visible, while remaining competitive in OpenE and weaker in TControl (Alonso et al., 2023).

Other papers argue that when reasoning is the goal, textualization should be limited or avoided. “Augment before You Try” adds external knowledge as a second table and answers with standard SQL over the original and augmenting tables, rather than flattening both into prose (Liu et al., 2024). On WikiTQ, the proposed method achieved Exact Match 55.80, compared with 54.10 for Binder+CoT and 53.60 for Program-of-Thought (Liu et al., 2024). “Beyond Linearization” goes further by replacing full-table serialization with an Attributed Table Graph and Question-Guided Personalized PageRank; the reported gain reaches up to 9.7% in accuracy, with average results of 78.8% on WikiTQ and 93.6% on TabFact for the full-graph variant (Wang et al., 13 Jan 2026). These results are frequently interpreted as evidence that textualization is useful only insofar as it does not erase the table’s relational structure.

4. Retrieval, knowledge fusion, and imputation

In retrieval-augmented systems, table textualization functions less as summarization than as preprocessing. The goat-farming knowledge assistant based on Retrieval-Augmented Generation explicitly defines table textualization as a two-stage mapping,

$f_{\text{text}}(T)=L(R(T)),$

where $R(\cdot)$ is a rule engine and $L(\cdot)$ is an LLM-driven semantic parser (Han et al., 11 Sep 2025). The rule engine performs header identification, cell-level mapping, and data validation; the semantic parser converts rows or fragments into coherent text while preserving semantics. The system used Qwen3-235B to transform 56 tables, of which 51 were in nutrition management, 2 in disease prevention and treatment, and 3 in feeding management (Han et al., 11 Sep 2025). The reported overall validation gain from adding table textualization to local retrieval was small, from 77.44% to 78.04%, because table questions were only 3.09% of the validation set; however, Table Q&A accuracy rose from 58.62% to 88.79%, a gain of 30.17 percentage points (Han et al., 11 Sep 2025).

TabText applies a comparable intuition to healthcare prediction, but its output is not free-form prose generation at inference time (Carballo et al., 2022). Each patient-day is converted into a paragraph-like serialization, one sentence per column, with contextual metadata such as table descriptions and the prediction objective. The best-performing construction used descriptive language, omitted missing values, replaced numerical values with text categories, and included metadata; the resulting representation is encoded into a 768-dimensional embedding and concatenated with standard tabular features (Carballo et al., 2022). Across nine healthcare tasks, the paper reports that augmenting pre-processed tabular data with TabText representations can improve the average and worst-case AUC performance of standard machine learning models by as much as 6% (Carballo et al., 2022).

TTITA treats missing textual columns as a table-to-text problem: given numeric, categorical, and textual fields in a row, it generates a missing text attribute with a transformer decoder (Wei et al., 2024). Its encoder builds a fixed-size context vector from heterogeneous column types, and a multi-task variant jointly imputes numeric, categorical, and text targets via

$$L_{\text{total}} = L_{MSE} + L_{CE}^{cat} + L_{CE}^{text}.$$

On Amazon Gift Cards reviews, TTITA achieved METEOR 0.4216, ROUGE-1 F1 0.4769, and BLEU-1 0.4517, outperforming LSTM, GRU, and prompted Llama2 baselines; the paper emphasizes that the improvement is more significant when the target sequence is longer (Wei et al., 2024).

5. Faithfulness, explanation, and human-centered textualization

A substantial recent shift is from full verbalization toward evidence-centered textualization. Table-R1 operationalizes this as region-based reasoning, in which the model first identifies a relevant table region

$T_{reg} = \{[c_1,\dots,c_i],[r_1,\dots,r_j]\}$

and then answers using textual, symbolic, or program-based traces (Wu et al., 18 May 2025). Region-Enhanced Supervised Fine-Tuning inserts $T_{reg}$ into Direct Prompting, Textual CoT, Symbolic CoT, and Program-of-Thought settings, while TARPO uses mixed rewards for both region quality and answer correctness. The paper reports an average performance improvement of 14.36 points across multiple base models and a 67.5% reduction in response token consumption relative to GRPO (Wu et al., 18 May 2025). This suggests that one productive form of table textualization is not sentence generation, but region abstraction.

The same explanatory turn appears in scientific claim verification. “Table-Text Alignment” extends SciTab with human-annotated cell-level rationales and evaluates both claim labeling and evidence selection (Ho et al., 12 Jun 2025). The aligned dataset contains 372 samples, of which 195 are supported and 177 refuted. GPT-4o achieved 88.4 Macro-F1 for claim labeling in zero-shot mode, but the best evidence-selection result was much lower: Qwen 2.5 72B with CoT prompting reached 50.8 Macro-F1, compared with 78.0 Macro-F1 human agreement (Ho et al., 12 Jun 2025). The paper’s central claim is that models often predict correct labels without recovering human-aligned rationales, so accuracy alone is insufficient evidence of faithful table-grounded reasoning (Ho et al., 12 Jun 2025).

TableTale extends this logic into reading interfaces. Instead of paraphrasing tables, it constructs a document-level linking schema between paragraphs, sentences, mentions, and table regions in scientific papers (Wang et al., 26 Feb 2026). The formative study analyzed 132 validated paragraph-table pairs from 59 computer science papers and identified 1,388 mentions, with semantic, numeric, and structural linking mechanisms (Wang et al., 26 Feb 2026). In a within-subject study with 24 participants, TableTale reduced cognitive workload and improved reading efficiency: average reading time fell from 4:56 under the Basic condition to 4:12 under TableTale (Wang et al., 26 Feb 2026). Here the “textualization” is neither full summarization nor retrieval corpus construction; it is the progressive surfacing of aligned evidence at paragraph, sentence, and mention granularity.

6. Subjectivity, inverse direction, and open problems

Most table-to-text work has treated the target as objective description, but “Facts-and-Feelings” argues for subjectivity-aware textualization (Dey et al., 2024). The Ta2TS dataset contains 3,849 instances across finance, weather, and sports, with references explicitly mixing objective information and subjective interpretation. On this benchmark, T5-base with contextual prefix reached BERTScore 85.15 and METEOR 24.98, while GPT-3.5 1-shot reached METEOR 26.28; the paper also notes that low BLEU does not necessarily indicate poor generation because the same meaning may be expressed with different words (Dey et al., 2024). This widens the problem definition from fact rendering to narrated interpretation.

A neighboring research direction reverses the mapping and thereby clarifies what textualization alone cannot do. T$^3$, the Text-Tuple-Table pipeline, treats text-to-table generation as extraction, integration, and realization rather than formatting (Deng et al., 2024). On LiveSum, a dataset of 3,771 football commentary examples paired with summary tables, baseline GPT-4 achieved RMSE 2.273 and Error Rate 46.32%, while full T$16 \times 16$0 reduced these to RMSE 0.929 and Error Rate 25.27%; Claude 3 Opus with T$16 \times 16$1 reached RMSE 0.438 and Error Rate 14.04% (Deng et al., 2024). The paper’s error analysis attributes all sampled GPT-4 + T$16 \times 16$2 errors to stage 1 tuple extraction, with 78% missing event tuples and 21% wrong event tuples (Deng et al., 2024). The broader implication is that table generation and table textualization both depend on reliable intermediate representations when information is distributed, aggregated, or implicit.

Several open problems recur across the literature. Content selection remains the largest bottleneck for high factual correctness in logical-form-based generation (Alonso et al., 2023). Aggregate improvements can understate module value when evaluation sets contain few table-specific questions, as in the goat-farming RAG system (Han et al., 11 Sep 2025). Very large tables still exceed the visual budget in pixel-based models (Alonso et al., 2023), while long serialized tables remain vulnerable to lost-in-the-middle effects that graph-based reranking explicitly targets (Wang et al., 13 Jan 2026). A plausible synthesis is that future table textualization systems will combine selective structure preservation, explicit intermediate semantics, and explanation-oriented outputs rather than relying on unrestricted linearization alone.