Source-linked AI summary
Table-to-text Generation by Structure-aware Seq2seq Learning
Tianyu Liu, Kexiang Wang, Lei Sha, Baobao Chang, Zhifang Sui
TL;DR
Table-to-text generation requires descriptions that capture both the content and structure of complex factual tables. The paper introduces a structure-aware seq2seq model with field-gated encoding and dual attention, and reports large-margin improvements over competitive baselines on WIKIBIO.
Problem
Existing table-to-text work used limited pre-defined schemas, while complex tables require comprehensive representation of both content and structure.
Method
A structure-aware seq2seq architecture uses a field-gating LSTM encoder and dual word-level and field-level attention for local and global addressing.
Results
The model outperforms competitive baselines by a large margin on WIKIBIO, with visualizations and case studies showing coherent, informative descriptions.
Takeaways & Limitations
The approach supports table-to-text generation by jointly modeling table content and structure.
Abstract
from arXiv · showhide
Table-to-text generation aims to generate a description for a factual table which can be viewed as a set of field-value records. To encode both the content and the structure of a table, we propose a novel structure-aware seq2seq architecture which consists of field-gating encoder and description generator with dual attention. In the encoding phase, we update the cell memory of the LSTM unit by a field gate and its corresponding field value in order to incorporate field information into table representation. In the decoding phase, dual attention mechanism which contains word level attention and field level attention is proposed to model the semantic relevance between the generated description and the table. We conduct experiments on the \texttt{WIKIBIO} dataset which contains over 700k biographies and corresponding infoboxes from Wikipedia. The attention visualizations and case studies show that our model is capable of generating coherent and informative descriptions based on the comprehensive understanding of both the content and the structure of a table. Automatic evaluations also show our model outperforms the baselines by a great margin. Code for this work is available on https://github.com/tyliupku/wiki2bio.
Introduction
Table-to-text generation must represent both the content and complex structure of factual tables, beyond limited pre-defined schemas and small datasets. This paper proposes a structure-aware seq2seq framework for generating biographies from Wikipedia infoboxes.
- Table-to-text generation produces natural-language descriptions from structured tables represented as field-value records.
- Earlier approaches used limited pre-defined schemas and datasets such as WEATHERGOV and ROBOCUP with only a few tens of thousands of records.
- WIKIBIO contains 700k biographies from Wikipedia, providing a challenging benchmark based on infoboxes with many field-value records.
- The proposed structure-aware seq2seq framework models table content and structure through local and global addressing.
- The architecture uses a field-gating encoder to incorporate field information and a decoder with dual attention for content and structure.
- Experiments on WIKIBIO report substantial improvement over baselines.
Related Work
Prior table-to-text systems separately addressed content selection and surface realization, often using statistical models or schema-specific representations. The paper positions its approach against limitations in long-range dependency modeling and complex table representation.
- Generation systems commonly separate content selection from surface realization.
- Content selection approaches align records with sentences or use hierarchical semi-Markov methods before generating descriptions.
- Lebret, Grangier, and Auli modeled Wikipedia infoboxes with an n-gram model using local and global conditioning.
- The n-gram approach is inefficient for modeling long-range dependencies between words during description generation.
- Mei, Bansal, and Walter used a seq2seq aligner for weather records with one-hot representations suited to relatively simple, highly structured data.
Task Definition
The task represents a table as field-value records and generates a token sequence describing it. Inference seeks the sequence with maximum conditional probability given the records.
- A table T is represented as n field-value records {R1, R2, · · ·, Rn}.
- Each record contains a sequence of words and corresponding field representations.
- The generated description S contains p output tokens, with wt denoting the word produced at time t.
- Inference generates the sequence w*1:p that maximizes P(w1:p|R1:n).
Structure-aware Seq2seq
The structure-aware seq2seq architecture encodes table content together with field and position information, then generates descriptions using local and global addressing through dual attention.
- Field representation: A Wikipedia infobox is represented as field-value records, with context embeddings for word segments and field embeddings identifying field names and token positions.Each token can be distinguished by its field and positions from the beginning and end of the field.
- Field-gating Table Encoder: The field-gating encoder incorporates each table word and its field embedding into an LSTM hidden state while updating cell memory with field information.The field gate controls how much proposed field information is retained in the cell memory.
- Field-gating Table Encoder: Concatenating word and field embeddings serves as a baseline but treats field information as an additional token label and loses structural information.The proposed field-gating variation instead updates the LSTM cell memory with a field gate and corresponding field value.
- Description Decoder with Dual Attention: The decoder uses dual attention: word-level attention models relevance to table content, while field-level attention locates the relevant field-value record.The dual attention weight combines field-level and word-level attention to produce the attention vector used for decoding.
- Local and Global Addressing: Local addressing models information within field-value records, whereas global addressing models relevance among records in the table.The seq2seq structure supports word-level context modeling for field contents, while field-gating representation and field-level attention support structural addressing.
Experiments
Experiments evaluate structure-aware table-to-text generation on WIKIBIO using BLEU-4 and ROUGE-4 against statistical and vanilla seq2seq baselines. The model improves generation quality and is more robust when infobox records are randomly disordered.
- Dataset and setup: WIKIBIO contains 728,321 English Wikipedia articles, with each article’s first sentence used as the corresponding infobox description.Tables average 53.1 tokens, compared with 26.1 tokens in descriptions.
- Dataset and setup: The experiments compare structure-aware seq2seq with statistical language models and vanilla seq2seq, using BLEU-4 and ROUGE-4 for automatic evaluation.The baselines include KN, Template KN, NLM, Table NLM, and vanilla seq2seq.
- Generation assessment: Neural models outperform statistical language models, and structure-aware seq2seq further improves table-to-text generation over vanilla seq2seq.Dual attention boosts performance by over 1 BLEU compared with vanilla attention.
- Research on disordered tables: The disordered-table experiment randomly shuffles infobox records without changing the content within individual fields.This tests whether models can generate descriptions when record order no longer matches the original infobox arrangement.
- Research on disordered tables: All neural models perform worse on disordered tables, but structure-aware seq2seq experiences smaller BLEU and ROUGE decreases than the other neural models.The authors attribute this result to the efficiency of the global addressing mechanism.
Qualitative Analysis
Qualitative analyses show that dual attention connects generated words to relevant table content and fields, while global addressing improves coverage of multi-field information. Case studies also reveal that vanilla seq2seq models may produce unsupported or incomplete details.
- Dual Attention: Dual attention properly links generated name, birthdate, nationality, and occupation information to the corresponding table content.The visualization attributes these references to aggregated dual attention.
- Case Study: Local addressing enables all three neural network generators to produce coherent and understandable sentences.In one case, each generator inferred “baseball” from the table despite its absence from the infobox.
- Case Study: Vanilla seq2seq models generate unsupported phrases such as “major league baseball” or “major leagues.”Without global addressing, the model selects the most probable league associated with a baseball player in Wikipedia.
- Case Study: Global addressing enables a 15-word generated segment to combine information from five table fields.The segment covers position, debutdate, finaldate, debutteam, and finalteam.
- Case Study: Vanilla seq2seq biographies omit the team served and the playing period contained in the infobox.The proposed structure-aware model covers nearly all information mentioned in the table in this case.
Conclusions
The paper presents a structure-aware seq2seq architecture that jointly represents table content and structure through field gating and dual attention. Visualizations, case studies, and generation assessments show substantial advantages over competitive baselines.
- Conclusions: The architecture combines a field-gating encoder with a description generator using word-level and field-level attention.The field gate incorporates field information into the encoder LSTM, while dual attention supports local and global addressing.
- Conclusions: Visualizations, case studies, and generation assessments show that the model outperforms competitive baselines by a large margin.