Source-linked AI summary
Order-Planning Neural Text Generation From Structured Data
Lei Sha, Lili Mou, Tianyu Liu, Pascal Poupart, Sujian Li, Baobao Chang, Zhifang Sui
TL;DR
Neural text-generation approaches for structured data lack explicit modeling of content order, although field order can matter in biographies. This paper plans table-to-text generation with learned field links, hybrid attention, and copying; on WikiBio, it reports gains across BLEU, ROUGE, and NIST over prior results.
Problem
Neural approaches to table-to-text generation lack explicit modeling of content order, although biography summaries typically mention nationality before occupation.
Source: Introduction · p. 1Method
The model learns links between table fields to plan content order, combines content- and link-based attention with a gate, and copies rare or unseen words.
Source: Introduction · p. 1, Conclusion and Future Work · p. 7Results
On WikiBio, the model outperforms previous results in BLEU, ROUGE, and NIST; the full model gains +2.5 BLEU and ROUGE and +0.3 NIST.
Source: Overall Performance · p. 5, Conclusion and Future Work · p. 7Takeaways & Limitations
WikiBio ablations support combining content- and link-based attention and copying, while the proposed order-planning method performs better than naive field-embedding alternatives.
Source: Different Ways of Using Field… · p. 6, Conclusion and Future Work · p. 7
Abstract
from arXiv · showhide
Generating texts from structured data (e.g., a table) is important for various natural language processing tasks such as question answering and dialog systems. In recent studies, researchers use neural language models and encoder-decoder frameworks for table-to-text generation. However, these neural network-based approaches do not model the order of contents during text generation. When a human writes a summary based on a given table, he or she would probably consider the content order before wording. In a biography, for example, the nationality of a person is typically mentioned before occupation in a biography. In this paper, we propose an order-planning text generation model to capture the relationship between different fields and use such relationship to make the generated text more fluent and smooth. We conducted experiments on the WikiBio dataset and achieve significantly higher performance than previous methods in terms of BLEU, ROUGE, and NIST scores.
Introduction
Neural table-to-text models capture language but lack explicit planning of content order, a feature illustrated by the typical ordering of nationality before occupation in biographies. The paper proposes field-link-based order planning with a copy mechanism and reports improved WikiBio results across three metrics.
Introduction
- Earlier human-designed rules and templates generated inflexible texts, motivating neural approaches to structured-data generation.Source: Introduction · p. 1
- Neural encoder-decoder systems can capture complex language but lack explicit modeling of content order during generation.The paper contrasts word-by-word generation conditioned on previous words and table information with deliberate planning of content order.Source: Introduction · p. 1
- In biographies, nationality is typically mentioned before occupation, illustrating how table fields can guide content order.Source: Introduction · p. 1
- A learned field-linking mechanism models relationships between table fields to plan what to say first, while a copy mechanism handles rare words.The model uses an encoder-decoder with RNN text synthesis and attention to table entries.Source: Introduction · p. 1
- On WikiBio, the approach significantly outperforms previous state-of-the-art results in BLEU, ROUGE, and NIST.The introduction also reports ablation tests of the model components and visualization analysis of the order-planning mechanism.Source: Introduction · pp. 1–2
Approach
The model converts a table into a natural-language summary with an encoder–dispatcher–decoder architecture. Its dispatcher combines content relevance with learned field-order information, while the decoder can copy rare table words.
Approach
- The model processes an input table through an encoder, a dispatcher that plans what to generate next, and an RNN decoder.The decoder also incorporates a copy mechanism to cope with rare words.Source: Approach · p. 2
Encoder: Table Representation
- The encoder reads a sequence of table contents alongside their corresponding field names using an LSTM-RNN.Field names receive separate embeddings, with each field represented by one embedding; the RNN integrates content words, including sequences or sets of words.Source: Encoder: Table Representation · p. 2
Dispatcher: Planning What to Generate Next
- The dispatcher combines content- and link-based attention to plan which table information to generate next.The link matrix learns how likely one field is to be mentioned after another, and its scores are weighted by previous attention before normalization.Source: Dispatcher: Planning What to Generate… · p. 2, Link-Based Attention · p. 3
Decoder: Sentence Generation
- The decoder uses hybrid attention over table contents and a copy mechanism that can handle rare or unseen words.The copy mechanism allows a word to be generated from the vocabulary or copied directly from the source table.Source: Approach · p. 2, Decoder: Sentence Generation · p. 3, Copy Mechanism · p. 4
Training Objective
- Training minimizes sentence negative log-likelihood with an ℓ2 penalty, using end-to-end backpropagation and Adam optimization.Source: Training Objective · p. 4
Experiments
These experiments evaluate order-planning on WikiBio biographies, comparing the full model with prior methods and testing its components and alternative uses of field information. The reported results favor the full model and adaptive hybrid attention; the case study illustrates nationality-before-occupation ordering.
Dataset
- Each WikiBio example pairs an infobox with the biography’s first sentence as the target, which averages 26.1 tokens and spans multiple fields.The dataset contains 728,321 biographies.Source: Dataset · p. 4
Settings
- The experiments use an 80% training and 10% test split, with model selection on a 1,000-sample validation subset based on BLEU-4.Content-word and generation-candidate vocabulary size was 20,000, despite more than 900,000 out-of-vocabulary words, motivating use of copying.Source: Dataset · p. 5, Settings · p. 5
Baselines
- The baselines include traditional KN language models and neural Table NLM variants; performance is reported with BLEU-4, ROUGE-4, and NIST-4.Perplexity is omitted because copying makes vocabulary size vary across samples, preventing comparison across approaches.Source: Baselines · p. 5
Results
- Adding all proposed components yields +2.5 BLEU and ROUGE improvement and +0.3 NIST improvement, achieving new state-of-the-art results.The comparison is with previous work on overall performance.Source: Overall Performance · p. 5, Table 2 · p. 5
- The copy mechanism is consistently effective across different types of attention, while link-based attention alone is less effective than content-based attention.The hybrid of content- and link-based attention with an adaptive gate performs better than either alone across both copy settings and BLEU, ROUGE, and NIST.Source: Ablation Test · p. 5, Table 3 · p. 5
- The best fixed interpolation reaches 42.89 BLEU at z = 0.4, below the self-adaptive gate.Fixed z values were tested from 0 to 1 in increments of 0.1; the gate uses decoding state and table information to balance the two attention types.Source: Effect of the gate · p. 6, Figure 3 · p. 6
- Feeding field information into content-attention computation degrades performance, while feeding it to the decoder RNN slightly improves performance; both alternatives trail the proposed method.The comparison tests alternative uses of weighted field embeddings.Source: Different Ways of Using Field… · p. 6, Table 4 · p. 6
Case Study and Visualization
- The case study shows link-based attention directing generation from the Nationality field to Occupation, producing “American economist” in that order.Content-based attention alone focuses on occupation and is described as confusing where “American” belongs.Source: Case Study and Visualization · p. 6
Related Work
Earlier text-generation work separated planning from realization or used templates and shallow learned models, while neural generation from structured data still lacked explicit content-order modeling. This paper proposes hybrid content- and link-based attention and a differentiable model in which planning and realization can be learned end-to-end.
Related Work
- Traditional generation divided content planning, sentence planning, and surface realization into separate steps, with templates and shallow statistical models used for realization.Prior planning work included collective classification and generative semi-Markov alignment, but planning and realization were generally separate.Source: Related Work · p. 6
- RNNs captured complex patterns through end-to-end training and were applied to tasks including machine translation, dialog, and summarization.Structured-data generation work used coarse-to-fine attention or copying, but did not explicitly model content order.Source: Related Work · p. 7
- The proposed approach combines content- and link-based attention, using a transition-like link matrix to model how likely one field is to follow another.It adapts the DNC’s hybrid content- and location-based addressing, replacing heuristic location addressing with a field-transition model.Source: Related Work · p. 7
- The model is differentiable, allowing planning and realization to be learned end-to-end.Source: Related Work · p. 7
Conclusion and Future Work
The paper combines link-based field-order planning with content attention and copying to generate text from tables. It reports strong WIKIBIO results and identifies multi-table generation and other structured data as future directions.
Conclusion and Future Work
- A link matrix supports link-based attention to model content order, while a self-adaptive gate balances it with content attention and copying addresses rare or unseen words.The generation process uses an RNN with attention to table contents.Source: Conclusion and Future Work · p. 7
- The approach outperforms previous WIKIBIO results by a large margin on BLEU, ROUGE, and NIST scores.Ablations support the copy mechanism and hybrid attention; comparisons favor the proposed order-planning mechanism over naively feeding field embeddings.Source: Conclusion and Future Work · p. 7
- The paper also provides a case study and visualizes attention scores to examine the model.Source: Conclusion and Future Work · p. 7
- Future work proposes hierarchical attention for selecting among multiple tables and fields, and applying the method to other structured data such as knowledge graphs.Source: Conclusion and Future Work · p. 7