Clément Rebuffel, Laure Soulier, Geoffrey Scoutheeten, Patrick Gallinari
Transcribing structured data into natural language descriptions has emerged as a challenging task, referred to as "data-to-text". These structures generally regroup multiple elements, as well as their attributes. Most attempts rely on translation encoder-decoder methods which linearize elements into a sequence. This however loses most of the structure contained in the data. In this work, we propose to overpass this limitation with a hierarchical model that encodes the data-structure at the element-level and the structure level. Evaluations on RotoWire show the effectiveness of our model w.r.t. qualitative and quantitative metrics.
| Task | Dataset | Metric | Value | Model |
|---|---|---|---|---|
| Text Generation | RotoWire (Relation Generation) | count | 21.17 | Hierarchical Transformer Encoder + conditional copy |
| Text Generation | RotoWire (Content Ordering) | BLEU | 17.5 | Hierarchical Transformer Encoder + conditional copy |
| Text Generation | RotoWire | BLEU | 17.5 | Hierarchical transformer encoder + conditional copy |
| Data-to-Text Generation | RotoWire (Relation Generation) | count | 21.17 | Hierarchical Transformer Encoder + conditional copy |
| Data-to-Text Generation | RotoWire (Content Ordering) | BLEU | 17.5 | Hierarchical Transformer Encoder + conditional copy |
| Data-to-Text Generation | RotoWire | BLEU | 17.5 | Hierarchical transformer encoder + conditional copy |