Caption Alignment and Structure-Aware Attention for Scientific Table-to-Text Generation
The task of table-to-text generation involves summarizing and creating natural language descriptions of tables. Previous approaches have used sequence-to-sequence generation methods, which view tables as linear sequences that only capture the structure of the tables. However, these methods are not e...
Saved in:
| Main Authors: | , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
IEEE
2024-01-01
|
| Series: | IEEE Access |
| Subjects: | |
| Online Access: | https://ieeexplore.ieee.org/document/10770202/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | The task of table-to-text generation involves summarizing and creating natural language descriptions of tables. Previous approaches have used sequence-to-sequence generation methods, which view tables as linear sequences that only capture the structure of the tables. However, these methods are not effective when dealing with tables containing complex structures, such as multi-level headers commonly found in scientific articles. Additionally, entities in scientific articles, like “model”, “task”, “dataset”, and “metric”, are important in table-to-text generation tasks. To address these challenges, we introduce a framework called Caption Header Alignment and Structure-Aware attention table text generation framework (CAS), which incorporates Caption Header alignment, Structure-Aware attention, and scientific table named entity recognition (NER) to align table headers with Captions, capture structural information in scientific tables, and recognize scientific table entities for text generation. Our experimental results demonstrate that our proposed method outperforms all previous baselines. Moreover, we also combine TableLlama-7b with CAS and achieve state-of-the-art performance on SciGen and numeric-NLG datasets. |
|---|---|
| ISSN: | 2169-3536 |