Source-linked AI summary
Table-to-Text: Describing Table Region with Natural Language
Junwei Bao, Duyu Tang, Nan Duan, Zhao Yan, Yuanhua Lv, Ming Zhou, Tiejun Zhao
TL;DR
Generating table descriptions requires representing table structure, producing fluent relevant sentences, and using informative low-frequency words. The paper encodes rows and uses a copying mechanism; its BLEU-4 scores improve on both synthetic datasets, and the authors report a strong WIKITABLETEXT result.
Problem
Generating table descriptions requires representing table structure, producing fluent and relevant sentences, and using informative low-frequency words such as names and numbers.
Source: Introduction · p. 1Method
An encoder-decoder represents a table row as a continuous vector and uses a copying mechanism to generate rare words from cells, attributes, and captions.
Source: Introduction · p. 1Results
BLEU-4 rose from 34.7 to 40.26 on WIKIBIO and from 33.32 to 39.12 on SIMPLEQUESTIONS.
Source: Introduction · pp. 1–2Takeaways & Limitations
The copying mechanism boosts performance and can selectively replicate appropriate table contents in generated sentences.
Source: Introduction · p. 1BLEU-4 decreases as the number of unseen attributes increases; handling unseen attributes is left for future work.
Source: Table-to-Text Generation with… · p. 6
Abstract
from arXiv · showhide
In this paper, we present a generative model to generate a natural language sentence describing a table region, e.g., a row. The model maps a row from a table to a continuous vector and then generates a natural language sentence by leveraging the semantics of a table. To deal with rare words appearing in a table, we develop a flexible copying mechanism that selectively replicates contents from the table in the output sequence. Extensive experiments demonstrate the accuracy of the model and the power of the copying mechanism. On two synthetic datasets, WIKIBIO and SIMPLEQUESTIONS, our model improves the current state-of-the-art BLEU-4 score from 34.70 to 40.26 and from 33.32 to 39.12, respectively. Furthermore, we introduce an open-domain dataset WIKITABLETEXT including 13,318 explanatory sentences for 4,962 tables. Our model achieves a BLEU-4 score of 38.23, which outperforms template based and language model based approaches.
Introduction
The paper frames table-to-text generation as describing a table row in natural language while addressing table understanding, fluent relevance, and rare table words. It introduces an encoder-decoder with copying, evaluates it on three datasets, and releases WIKITABLETEXT.
Introduction
- Table-to-text generation could support search engines and conversational agents by turning structured table information into natural-language descriptions.Source: Introduction · p. 1
- The task must represent table structure, generate fluent and relevant sentences, and use informative low-frequency words such as names and numbers.Source: Introduction · p. 1
- The encoder-decoder represents a table row as a continuous vector and copies rare words from cells, attributes, and captions into generated sentences.Source: Introduction · p. 1
- WIKITABLETEXT contributes 13,318 explanatory sentences for 4,962 tables as an open-domain resource for table-to-text generation.Source: Introduction · pp. 1–2
- 38.23 BLEU-4 on WIKITABLETEXT outperformed template-based and neural language-model approaches by an order of magnitude; BLEU-4 rose from 34.7 to 40.26 on WIKIBIO and from 33.32 to 39.12 on SIMPLEQUESTIONS.Source: Introduction · pp. 1–2
Task Formalization and Dataset
The task maps a selected table row to a descriptive sentence, with larger regions left for future work. The paper also describes how WIKITABLETEXT was assembled and distinguishes its open-domain scope from prior datasets.
Task Formalization
- A table is formalized through its attributes, cells, and caption; the model input is a selected row, and only some row information need appear in its description.Source: Task Formalization and Dataset · p. 2
- The task excludes multi-row regions and irregular tables, defined as tables with merged attributes or cells or unequal counts of cells and attributes.The paper leaves larger regions and irregular tables to future work.Source: Task Formalization and Dataset · p. 2
Dataset Construction
- Three rows were sampled from each of 5,000 Wikipedia tables with at least 3 rows and 2 columns; annotators describe at least two cells without external lookup.The resulting 13,318 row-text pairs were split into 10,000 training, 1,318 development, and 2,000 test instances.Source: Task Formalization and Dataset · p. 2
Dataset Scope
- WIKITABLETEXT is presented as the first open-domain table-to-text dataset, unlike domain-restricted WEATHERGOV and ROBOCUP; this work uses row input rather than multiple-row reasoning.Source: Task Formalization and Dataset · p. 2
Background: Sequence-to-Sequence
Sequence-to-sequence learning encodes a source sequence as a continuous vector and decodes it into a target sequence. Its encoder represents the input, while its decoder generates output words from that representation and previous words.
Background: Sequence-to-Sequence
- The encoder maps a variable-length source sequence to a fixed-length continuous vector, using its last hidden state as the sequence representation.An RNN updates its hidden vector recursively from the current word and previous hidden state.Source: Background: Sequence-to-Sequence · p. 3
- The decoder generates each target word from the source representation and the previously predicted words.The decoder is typically an RNN that produces one word at each time step.Source: Background: Sequence-to-Sequence · p. 3
- Gated RNN units such as LSTM or GRU are commonly used to address the gradient-vanishing problem in standard RNNs.The passage names LSTM and GRU as alternatives for the decoder hidden-state function.Source: Background: Sequence-to-Sequence · p. 3
Approach: Table-to-Sequence
The approach encodes a table row using its cells, attributes, and caption, then uses a table-aware decoder to generate a descriptive sentence. Attention and a flexible copying mechanism incorporate table content, including rare words, into generation.
Table Aware Encoder
- The encoder represents each row column using cell and attribute embeddings, treating the caption as a special cell with a virtual attribute.It averages the column vectors to initialize the decoder, making the representation invariant to exchanging columns.Source: Table Aware Encoder · p. 3
Approach: Table-to-Sequence
- A GRU-based sequential decoder generates one word at a time and is extended to incorporate table structure and produce table-related descriptions.Figure 2 illustrates the proposed table-to-text generation approach.Source: Table Aware Decoder · p. 3, Figure 2 · p. 4
Table Aware Decoder
- Column attention incorporates the previous time step’s attention, ideally helping the decoder remember used columns and avoid attending to them again.The resulting attention weights form a context vector that influences the hidden state at each time step.Source: Table Aware Decoder · pp. 3–4
- The decoder conditions word generation on global table information and local information from the attribute corresponding to the preceding output word.The global information is the encoder output, while an unknown-attribute symbol is used when the preceding word did not come from the table.Source: Table Aware Decoder · p. 4
- The decoder uses a neural gate to trade off generating from the target vocabulary and copying a word from a table cell.When a word appears in both sources, its probability combines the vocabulary and cell-copy contributions.Source: Table Aware Decoder · p. 4
- Table2Seq++ extends cell copying by also using attention over attribute embeddings as evidence to copy attribute words.The cell-copying model is named Table2Seq; the attribute-enhanced version is Table2Seq++.Source: Table Aware Decoder · p. 4
Training and Inference
- The model is trained end to end with back-propagation and uses beam search at inference to generate the top-K results.K is the beam size.Source: Training and Inference · pp. 4–5
Experiment
Experiments evaluate the model on WIKITABLETEXT, WIKIBIO, and SIMPLEQUESTIONS; on WIKITABLETEXT, copying is the strongest component, though generated examples still reveal errors in column interpretation and cell relations.
Experiment
- The experiments assess the approach on three datasets: WIKITABLETEXT, WIKIBIO, and SIMPLEQUESTIONS.Source: Experiment · p. 5
Implementation Details
- The three datasets share Gaussian parameter initialization, 300-dimensional word and attribute embeddings, 500-dimensional decoder states, beam size 5, and BLEU-4 evaluation.Ada-delta adapts the learning rate; the dev set halves it after six epochs without improvement.Source: Implementation Details · p. 5
Table-to-Text Generation with WIKITABLETEXT
- 37.90 to 5.41: removing copying sharply reduces WIKITABLETEXT performance, and the model outperforms baselines by an order of magnitude.Table 2 reports BLEU-4 scores; the authors attribute copying's impact to rare words in table cells.Source: Table-to-Text Generation with… · p. 5, Table 2 · p. 5
- Attention peaks on the source cells for copied words, the learned gate can instead generate vocabulary words, and copying from both cells and attributes improves results.The caption is the second most important factor on WIKITABLETEXT; global and local factors have little influence there.Source: Table-to-Text Generation with… · pp. 5–6, Figure 3 · p. 6
- Generated examples can be fluent yet add or omit details or misread cell relations; BLEU-4 decreases as the number of unseen attributes rises.The authors leave effective handling of unseen attributes for future work.Source: Table-to-Text Generation with… · p. 6, Figure 5 · p. 7
Infobox-to-Biography Generation with WIKIBIO
- Table2Seq outperforms the WIKIBIO baselines, but removing copying has limited impact; the authors relate this to longer reference sentences averaging 26.1 words.Even without copying, the model performs better than Table NLM, which the authors attribute to the GRU decoder with recurrent attention.Source: Infobox-to-Biography Generation with… · p. 6, Table 5 · p. 6
Fact-to-Question Generation with SIMPLEQUESTIONS
- On SIMPLEQUESTIONS, the model outperforms baselines by a wide margin and generates questions from knowledge-base facts when training data is rich.The baselines generate a question template and replace its placeholder with the subject; the model learns where to copy from.Source: Fact-to-Question Generation with… · p. 7, Table 6 · p. 7
Related Work
Prior work generates text from varied structured data, while this paper focuses on web tables and uses attention and copying to replicate source contents.
Related Work
- This work targets web tables, unlike prior generation from database records, infoboxes, or Freebase facts.Related approaches use scenario records (Angeli, Liang, and Klein 2010; Mei, Bansal, and Walter 2015), infoboxes (Lebret, Grangier, and Auli 2016), and Freebase facts (Serban et al. 2016).Source: Related Work · p. 7
- The approach differs from neural encoder-decoder methods that do not copy source contents into the output.Mei, Bansal, and Walter (2015) and Serban et al. (2016) use neural encoder-decoders without copying; this paper combines attention and copying.Source: Related Work · p. 7
- Wiseman, Shieber, and Rush (2017) study long-text generation from tables and report challenges in data-to-text generation.Source: Related Work · p. 7
Conclusion
The paper presents an encoder-decoder model for describing table rows, with flexible copying, and reports experiments demonstrating its accuracy and copying capability.
Conclusion
- The encoder-decoder model represents table rows using table semantics and copies content from cells, attributes, and captions into generated sentences.Source: Conclusion · p. 7
- The authors release WIKITABLETEXT, an open-domain dataset for table-to-text generation.Source: Conclusion · p. 7
- Experiments demonstrate the approach's accuracy and the power of its copying mechanism.Source: Conclusion · p. 8