Abstract
This paper introduces a unified, unsupervised framework for concept-to-text generation—the task of transforming structured data into natural language. By reimagining generation as a global parsing problem using a Probabilistic Context-Free Grammar (PCFG), the authors move away from brittle, hand-engineered pipelines toward a domain-independent model that jointly decides what to say and how to say it. Evaluated across three diverse domains, the model achieves competitive performance without requiring manual rules or expensive labels.
Introduction
Traditionally, generating text from a database (e.g., a weather report from sensor data) involves a "pipeline" of three distinct stages: content planning (deciding what to mention), sentence planning (determining structure and lexical choices), and surface realization (generating actual sentences). These pipelines are typically hand-coded for specific domains, making them expensive to build and difficult to port. Furthermore, because each stage makes local decisions, errors often propagate through the system.
The authors address this "pipeline bottleneck" by creating a model that optimizes the entire process globally. By learning from unaligned data—databases paired with text descriptions—the model eliminates the need for manual alignment or domain-specific heuristics.
Methodology: Generation as Parsing
The core innovation is recasting generation as parsing in reverse.
The PCFG Backbone
The authors define a Probabilistic Context-Free Grammar where non-terminals represent database records and fields, and terminals are the words. This grammar captures the inherent structure of the data and its mapping to language.
Joint Optimization and Learning
Instead of sequential steps, the model finds the most likely "derivation tree" that connects a set of database records to a sequence of words. This allows the model to weigh content selection and word choice simultaneously. Weights for the grammar rules are learned using the Expectation-Maximization (EM) algorithm from a corpus of database records and their corresponding (but not word-for-word aligned) text descriptions.
The Extended Decoder
To ensure output fluency, the PCFG is intersected with external models: an n-gram language model and an unsupervised dependency model (DMV). To manage the resulting computational complexity, the authors employ cube pruning, an efficient search technique borrowed from machine translation.
Datasets and Domain Portability
The model was evaluated across three distinct domains to demonstrate its portability:
- ROBOCUP: A sportscasting dataset consisting of soccer game events and short commentary.
- WEATHERGOV: A large dataset of weather scenarios (temperatures, wind speeds, etc.) paired with multi-sentence forecasts.
- ATIS: A dataset of flight-related queries mapped to structured meaning representations.
Key Insights
- Global vs. Local Decisions: By avoiding a greedy search over local decisions, the model can recover from "bad" content choices if they lead to ungrammatical text, and vice versa.
- Unsupervised Syntactic Knowledge: By using an unsupervised dependency model (DMV), the generator learns long-range relationships (e.g., subject-verb agreement) that standard language models often miss.
- Domain Independence: Because the grammar structure is based on the general concepts of "records and fields," the same mathematical framework works for soccer, weather, and air travel without modification.
Results
The model achieved results competitive with, and sometimes superior to, state-of-the-art systems relying on domain-specific constraints. On the WEATHERGOV domain, the joint model (k-BEST-LM-DMV) achieved a BLEU-4 score of 34.18, significantly outperforming a 1-best baseline (8.64).
Human evaluations further validated these results. In "judgment elicitation" studies, humans rated the model's output as significantly more fluent and semantically correct than baseline systems. In the ATIS domain, the model's output was found to be nearly indistinguishable from human-written text.
Conclusion and Future Work
The proposed global model demonstrates that concept-to-text generation can be effectively handled by a unified, unsupervised framework. While the model currently requires a pre-defined output length and faces computational challenges during decoding, it offers a robust path forward for domain-independent generation. Future work may extend the PCFG to capture document-level discourse patterns or "flip" the model to perform semantic parsing, turning text back into structured data.
Summary updated