Introduction
The design of complex Very Large Scale Integration (VLSI) chips relies heavily on Computer-Aided Design (CAD) systems. However, these systems are limited by the quality and depth of their knowledge bases. Currently, a significant bottleneck exists in manually translating technical specifications from hardware manuals into formal models. These documents are often ambiguous, rely on implicit causal relationships, and span multiple levels of abstraction. The hmU (hardware manual Understander) system addresses this by automating the extraction of formal hardware models from natural language specifications.
System Architecture
The hmU system employs a two-part architecture to bridge the gap between unstructured text and functional hardware models:
- Natural Language (NL) Analyzer: Utilizing Lexical Functional Grammar (LFG), this component translates sentences into an intermediate representation. This is then refined into a "canonical form" that identifies specific events, agents (like the CPU), and objects (like memory).
- Model Builder: This core component constructs an automaton-based model using a hierarchical "cause-effect" framework. It integrates "common sense" domain knowledge regarding timing, synchronization, and hardware behavior to resolve linguistic ambiguities.
Hierarchical Modeling and 'Indicates' Relations
A primary challenge in hardware manual comprehension is the multi-level nature of descriptions. A single high-level event, such as a "memory read request," corresponds to a specific sequence of low-level signal changes. The hmU system manages this through "indicates" relations, which link functional events at higher levels to physical signal transitions at lower levels.
As shown in Figure 1, the system maps a high-level CPU request to Level 2 functional events (e.g., *MREQ becoming active) and ultimately to Level 1 signal changes synchronized with the system clock. This hierarchical structure allows the model to maintain logical consistency across different granularities of description.
Causal Reasoning and Diagram-Text Fusion
Hardware components are treated as "cooperating agents" that operate in parallel. To model their interactions, the system must identify causal links that are often implicit in the text but explicit in timing diagrams. The authors propose a "Diagram-Text Fusion" strategy where timing charts provide the structural skeleton of the model, which is then enriched by the natural language text.
Figure 6 illustrates how the system extracts causal dependencies from timing charts. By identifying which signal transitions trigger others (e.g., a clock edge triggering a data sample), the Model Builder can construct a robust causal network. To handle incomplete or vague information, the system employs a reasoning process similar to a Truth Maintenance System (TMS), allowing it to revise the model as more specific data is processed.
Evaluation and Results
The hmU system was evaluated using the Z80-CPU Technical Manual. Qualitative results demonstrated that the system could successfully map complex sentences (e.g., "The CPU turns off the *RD signal") into formal model structures. Furthermore, it effectively linked signal transitions at the pin level to functional operations, validating the hierarchical cause-effect approach.
Limitations and Future Directions
While the prototype demonstrates the feasibility of automatic model construction, several limitations remain. Currently, timing diagrams must be manually encoded into symbolic expressions, as the system lacks a direct "diagram analyzer." Future work aims to automate this process and improve the system's ability to handle more intricate discourse analysis and complex canonical transformations. Scaling the system to handle arbitrary, highly complex LSI chips remains a long-term goal.
Summary updated