[Paper Review] Table-to-Text Generation with Effective Hierarchical Encoder on Three Dimensions (Row, Column and Time)

Heng Gong, Xiaocheng Feng+2 more|arXiv (Cornell University)|Sep 5, 2019

Advanced Text Analysis Techniques20 references4 citations

TL;DR

This paper proposes a novel hierarchical encoder for table-to-text generation that models table cells across three dimensions—row, column, and time—using self-attention mechanisms. It introduces a cell fusion gate to dynamically combine salient representations from each dimension, significantly improving factual consistency and fluency, achieving state-of-the-art results on the ROTOWIRE benchmark with a 2.66 BLEU improvement over strong baselines.

ABSTRACT

Although Seq2Seq models for table-to-text generation have achieved remarkable progress, modeling table representation in one dimension is inadequate. This is because (1) the table consists of multiple rows and columns, which means that encoding a table should not depend only on one dimensional sequence or set of records and (2) most of the tables are time series data (e.g. NBA game data, stock market data), which means that the description of the current table may be affected by its historical data. To address aforementioned problems, not only do we model each table cell considering other records in the same row, we also enrich table's representation by modeling each table cell in context of other cells in the same column or with historical (time dimension) data respectively. In addition, we develop a table cell fusion gate to combine representations from row, column and time dimension into one dense vector according to the saliency of each dimension's representation. We evaluated our methods on ROTOWIRE, a benchmark dataset of NBA basketball games. Both automatic and human evaluation results demonstrate the effectiveness of our model with improvement of 2.66 in BLEU over the strong baseline and outperformance of state-of-the-art model.

Motivation & Objective

• To address the limitation of existing table-to-text models that encode tables as one-dimensional sequences or sets of records, which neglects structural and temporal relationships.
• To improve factual fidelity in generated text by modeling table cells in the context of their row, column, and historical time-series data.
• To enhance content selection and summarization by jointly encoding row, column, and time representations using a learnable fusion mechanism.
• To model player performance over time, such as identifying 'double-double' trends, which requires access to historical data beyond the current table.

Proposed method

• Uses three separate self-attention networks to encode each table cell in its row, column, and time dimensions independently.
• Applies a learnable cell fusion gate to combine the three-dimensional representations into a single, dense vector based on the saliency of each dimension.
• Employs mean pooling over row-level representations to form a row-level context representation, followed by a content selection gate to filter less relevant rows.
• Uses a standard attention-based decoder with a content selection gate to attend selectively to important table rows during text generation.
• Treats time-series data by modeling each cell’s representation using historical values from the same player across previous games.
• End-to-end training is performed on the ROTOWIRE dataset, with automatic and human evaluation to assess fluency, coherence, and factual consistency.

Experimental results

Research questions

• RQ1 Can modeling table cells across row, column, and time dimensions improve the quality of generated text in table-to-text generation?
• RQ2 How does jointly encoding row, column, and time representations affect content selection and factual consistency?
• RQ3 Can a learnable fusion gate effectively combine multi-dimensional representations to improve model performance?
• RQ4 Does incorporating historical performance data enhance the model’s ability to generate accurate summaries of player trends (e.g., 'double-double' streaks)?
• RQ5 How does the proposed model compare to strong baselines and state-of-the-art models in terms of automatic and human evaluation metrics?

Key findings

• The model improves BLEU score from 14.19 (baseline) to 16.85, a relative improvement of 18.75%.
• The model achieves 91.46% in relation generation (RG) F1 score, up from 74.80% in the baseline, a 22.27% relative improvement.
• Content selection (CS) F1 increases from 32.49% to 41.21%, a 26.84% relative improvement, indicating better selection of relevant table records.
• Content ordering (CO) score improves from 15.42 to 20.86, a 35.28% relative improvement, showing better structural coherence in generated text.
• Human evaluation confirms that the model generates more grammatical, coherent, and factual text than both neural baselines (CC, NCP) and template-based systems.
• Qualitative analysis shows the model correctly identifies key players like Al Jefferson and accurately summarizes historical performance trends, such as 'second double-double in last three games'.