Abstract
Jacob Eisenstein, James Clarke, Dan Goldwasser, Dan Roth. Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing.
2009.
Introduction
Machine learning performance is often limited by the quality of data representation. Traditionally, human experts manually engineer "features"—the specific attributes or relationships a model analyzes—to make complex tasks learnable. This manual engineering creates a significant bottleneck, requiring expertise in both the specific domain and machine learning architecture.
This paper introduces "Reading to Learn," a framework that bypasses this bottleneck by enabling agents to automatically discover high-level semantic features from human-oriented instructional text. By "reading" manuals or tutorials, the system extracts logical structures that are more informative than raw sensory data.
Methodology: Multidocument Semantic Abstraction
The core of the approach is a generative model that assumes natural language sentences are stochastically generated from underlying logical formulas (conjunctions of literals), such as card(x) ∧ freecell(y) ∧ empty(y).
Syntax as a Semantic Signal
A key innovation is the use of dependency parsing. The model assumes that the physical structure of a sentence—its dependency tree—provides a strong constraint on how logical variables are linked. If two predicates share a variable in a logical formula, their corresponding words are likely to appear close together in the sentence's syntax tree.
The model uses "lexical glosses"—a small dictionary mapping words to sensors (e.g., "foundation" maps to the homecell predicate)—to bridge the gap between text and logic. To find the best formulas without labeled data, the system employs a Dirichlet Process to discover latent formulas across multiple documents. This "multidocument" approach allows the model to filter out stylistic noise and focus on consistent logical structures.
Experimental Evaluation
The researchers evaluated the framework in the domain of Freecell solitaire. The world model consisted of 11 basic predicates (sensors) like card(x) and on(x,y). Instructional text was gathered from five different online sources. The system's task was to extract features that could help an Inductive Logic Programming (ILP) learner (ALEPH) distinguish between legal and illegal game moves.
Results
The "Full Model" was compared against several baselines, including a "Sensors-Only" model and a "No-Syntax" version.
As shown in the results, the features extracted by the Full Model consistently outperformed the raw sensors. With only 15 training scenarios, the Full Model achieved 82.89% accuracy, compared to 79.12% for sensors alone. Furthermore, the "No-Syntax" baseline performed significantly worse, validating the hypothesis that leveraging dependency structures is crucial for correctly linking variables in extracted features.
Conclusion
The "Reading to Learn" framework demonstrates that machines can effectively leverage human-written instructions to build better internal representations. By treating syntax as a signal for semantic structure and aggregating information across multiple documents, agents can recover complex, useful features like empty(x1) ∧ freecell(x1) (identifying an empty freecell).
Future work aims to automate the creation of lexical glosses and integrate the "reading" phase more tightly with environmental exploration, allowing agents to resolve textual ambiguities through direct interaction.
Summary updated