Data‑to‑Text Generation: Controllability, Reliability, and Planning Strategies

Task Background – Data‑to‑text generation converts structured data (tables, key‑value pairs, RDF triples, etc.) into natural language descriptions for applications like sports commentary, e‑commerce product descriptions, and medical reports. Challenges include domain‑specific terminology, cross‑domain coverage, and selecting the most important information.

Common Datasets – Typical datasets range from NBA match statistics to Wikipedia infoboxes and newer datasets that include logical expressions. The goal is to model the conditional probability P(y|x), where y is the generated text and x is the structured input.

Structured Data Representation – Two main encoding strategies are used: (1) treating key‑value pairs as tokens concatenated with table content, and (2) appending attribute information after the table content. Both can be fed into RNN, Transformer, or pretrained encoder‑decoder models.

Explicit Planning – Involves three stages: content selection, sentence ordering, and sentence‑level planning. Planning can be represented as a text plan (z) that guides a description generation network. Examples show how swapping plan tokens changes the generated sentences while preserving coherence.

Implicit Planning – Treats the plan z as a latent variable (discrete or continuous). Training uses variational methods (VAE, EM, HMM) with supervised and unsupervised losses. Sampling different z values yields diverse outputs; continuous latent variables can be modeled with VAE, while discrete sequences can be handled with semi‑Markov models and Viterbi decoding.

Model Architectures – Both explicit and implicit approaches can use Bi‑LSTM, Seq2Seq, or pretrained models. Implicit methods benefit from flexibility and diversity but may be unstable during training.

Advantages & Disadvantages – Explicit planning offers fine‑grained control and interpretability but requires aligned data‑text pairs. Implicit planning requires only latent structure, enabling diverse generation, but training can be more complex.

Q&A Highlights – Variational training is harder but can achieve better quality when stable; KL annealing helps mitigate posterior collapse. Pre‑training the description network improves cold‑start performance.

Overall, the work demonstrates how integrating planning—either explicit or implicit—enhances controllability and reliability in data‑to‑text generation systems.