Carleton University - School of Computer Science Honours Project
Extracting Mechanical Structure from Industrial Designs
Natural Language Processing is applied to a practical problem facing the manufacturing industry, where industrial design files are not conducive to managing materials and components in a machine-readable format. Free text from diverse design files is analyzed, and the structure of the design is inferred based on known examples of text. Principal observations are: 1. Given robust training data, Bayesian classification of part numbers by three-character trigram performs well. 2. Part numbers cluster poorly when grouped by Levenshtein Distance or Longest Common Subsequence. 3. Technical shorthand and abbreviations follow a grammar which is machine-readable. 4. A standard English Corpus, such as the Brown Corpus, when augmented by a technical dictionary and training data, can be used to parse this grammar. 5. Abbreviations, mixed-character alphanumeric strings, and non-standard English terms combine to form quasi-word collocations, which can be identified by statistical analysis in the same way as English word collocations. 6. The type of material or item, if any, described by a collocation discovered in this way, can be learned by a machine, given robust training data. 7. The parent-child relationship between types of materials can be learned through statistical analysis.