Data-to-text Generation: Methods & Applications

• Data-to-text generation is the process of automatically transforming structured data into tailored, human-readable narratives and reports.
• It employs a modular pipeline that includes signal analysis, content selection, and surface realization to ensure clarity and coherence.
• Both rule-based and trainable methods are used, with performance evaluated using metrics like BLEU, ROUGE, and content ordering measures.

Data-to-text generation (DTG) is a subfield of Natural Language Generation (NLG) concerned with the automatic transformation of non-linguistic, structured data—such as time series, tables, knowledge graphs, and attribute–value pairs—into coherent, contextually appropriate natural language text. DTG systems replace or complement traditional data visualization by producing textual summaries, reports, or descriptions, targeting domains as varied as weather forecasting, finance, healthcare, sports analytics, and e-commerce. Central to DTG is the challenge of content selection: determining which facts to express and how to verbalize them in a manner that aligns with user needs and preserves the fidelity and clarity of the information.

1. System Architecture and Processing Stages

The canonical architecture of data-to-text systems consists of a modular pipeline comprising the following primary stages—originally formalized in Reiter (2007) and consistently reflected in research up to the present:

1. Signal Analysis: Processes raw input (numerical data, sensor streams) to extract trends, anomalies, or other patterns.
2. Data Interpretation: Establishes higher-order relationships, such as correlations or causal links, among the extracted patterns.
3. Document Planning: The content selection module—tasked with identifying what subset of information should be conveyed and in what order. This includes discourse structuring and logical ordering to ensure coherence.
4. Microplanning and Surface Realisation: Converts selected content into grammatically correct, fluent text through surface realization. This stage may be implemented using rigid templates, grammar-based systems, or trainable neural LLMs.

This architecture is realized across a spectrum of design paradigms, from strict rule-based pipelines to fully end-to-end neural models in which these steps are latent or softly integrated (Gkatzia, 2016).

2. Content Selection: Rule-based and Trainable Methods

Content selection is foundational in bridging raw data and user-personalized narrative. The surveyed literature delineates two principal methodological families (Gkatzia, 2016):

• Rule-Based Methods: These systems depend on expert knowledge and domain rules to select salient data. Mechanisms involve:
  • Explicit enumeration or thresholding of trends and outliers.
  • Domain and communication reasoners that apply handcrafted policies (e.g., Gricean Maxims) regarding relevance, brevity, and informativeness.
  • Techniques such as clustering, thresholding, and domain-specific decompression models.
  • Robustness and transparency are strengths, but scalability to new domains is a limitation.
• Trainable (Statistical/Machine Learning) Methods: These cast content selection as a multi-label classification or sequential labeling problem:
  • Models include Hidden Markov Models (HMMs), structured perceptrons, Support Vector Machines, Integer Linear Programming, and neural network classifiers.
  • Recent approaches integrate content selection and surface realization, often via joint inference, reinforcement learning, or probabilistic graph models capturing inter-fact dependencies.
  • Data-dependency is a key consideration—models require large, aligned datasets, but they generalize more flexibly to novel domains.

The following table summarizes the core characteristics of both families:

Approach Type	Strengths	Limitations
Rule-Based	Robust, interpretable	Limited scalability, manual tuning
Trainable/Statistical	Scalable, adaptable	Requires large, aligned datasets

In both families, the granularity of selection (sentence-level, word-level, or discourse-level), the ability to capture omitted and redundant information, and the integration with downstream linguistic realization remain critical research axes.

3. Personalization, User Adaptation, and Output Structuring

Advanced DTG systems emphasize user adaptation—modulating content selection and surface realization to suit preferences, expertise, and context of the target audience. Approaches include:

• Adaptive content filtering and ordering based on user profiles.
• Learning-to-rank or dynamic narrative structuring, as in multi-label classification settings.
• Incorporation of user feedback loops to enhance system interactivity and output relevance.

Document planning functions not only to filter irrelevant facts, but also to assemble and structure content into well-formed paragraphs, logically ordered sentences, and coherent narratives. Rule-based systems employ ordering heuristics or discourse relations, while trainable systems may deploy sequence modeling with attention over data representations.

4. Evaluation Criteria and Task-Specific Metrics

Evaluating DTG systems requires both intrinsic and extrinsic metrics due to the multiplicity of goals (factuality, fluency, utility, user satisfaction):

• Intrinsic Metrics: BLEU, ROUGE, and F-measure assess surface-level n-gram overlap with gold summaries. These are supplemented with more targeted measures:
  • Relation Generation (RG): Counts and precision of factually correct mentions.
  • Content Selection (CS): Precision and recall of selected content relative to human reference.
  • Content Ordering (CO): Normalized edit distances (e.g., Damerau-Levenshtein distance) on content orderings.
• Extrinsic Metrics: User task success rates, downstream decision effectiveness, and subjective satisfaction ratings via human studies.

A dual focus on both intrinsic content quality and real-world impact is recommended for comprehensive assessment (Gkatzia, 2016).

5. Domain Dependence, Data Requirements, and Method Selection

Selecting an appropriate DTG approach depends crucially on domain size, data/knowledge availability, and the goals of the deployment:

Factor	Rule-Based Recommendation	Trainable Recommendation
Domain size	Small, well-constrained	Large, open-ended
Expert knowledge	Readily available/encodable	Scarce, with abundant data
Alignable corpus	Unavailable	Available

In low-data, tightly constrained settings, expert-driven pipelines are justified. Conversely, large domains with sufficient aligned (data + text) corpora benefit from data-driven, trainable architectures (Gkatzia, 2016).

6. Opportunities and Research Frontiers

Several prominent directions are outlined for future progress in DTG:

• Transfer Learning: Enhancing portability of content selection and language realization modules across domains and languages.
• Multi-modality: Hybrid systems integrating textual and visual modalities for richer user support, particularly when conveying uncertainty.
• Handling Data Uncertainty: Methods to verbalize probabilistic or uncertain data (e.g., “likely,” “possible”) in a manner understandable to users.
• Evaluation: Developing extrinsic metrics and user-centric evaluation protocols that better capture meaningfulness and decision support.

The paper further advocates for systematic evaluations that encompass both output quality and user impact (Gkatzia, 2016).

7. Conclusion

Data-to-text generation constitutes a multidisciplinary research domain at the interface of NLG, knowledge representation, and adaptive user interfaces. Approaches span from transparent, rule-based architectures suited to expert-driven, small-scale tasks, to scalable statistical and neural methods adapting rapidly to data-rich environments. Content selection remains the pivotal component dictating narrative clarity and informativeness. The balance between domain specificity, training data sufficiency, and desired generalization informs system design. Future advances are expected in cross-domain generalization, multi-modal integration, user adaptation, and holistic evaluation methodologies.