When initiating a new design, designers usually begin with a database to look for historical design solution, past experience and available techniques through design documents. This database is a collection of labeled design documents under predefined categories, and covers the concerned aspects from product structure, functionality, manufacturing processing and materials, in which all have an effect on the final design quality. To achieve a desired product design, Lakshminarayanan et al. built a database with historical operating conditions and product quality, and presented a methodology to analyze it [1]. Ferguson suggested product designers to access a range of corporate databases [2], in particular, customer complaints, product material features, and R&D testing, with the use of data mining techniques. This investigation enabled information from later life cycle stages to be used by earlier ones, and made this information understandable and useable to other product functions. In order to effectively use past design knowledge, Wu built up a design knowledge database through an elaborate process of data collection, mining, integration, storing, management and maintenance [3]. It included design standards and requirements, product samples, experimental data, computational models, designers' experiences, and failure information. Product designers need material information relevant with product issues to adjust design solutions. To meet their needs, Kesteren investigated the current utilization of material properties database, and proposed some strategies for database developer to improve information presentation [4]. All these databases are becoming electronically accessible and growing at an explosive rate.
The applications of text mining technique to support product design increase dramatically in the last five years. It provides an effective tool to process and analyze huge amount of design documents and to eventually gain valuable insight into product design solutions. In most of these applications, a labeled database is essential. However, manually labeling these databases is of high cost [5]. The alternative is to manually label the relatively small number of design documents, and use them as training samples to build a set of text classifiers with artificial intelligence techniques. These classifiers can categorize the whole design document database automatically and yield appropriate labels to each document.
Several research papers have been focused on generation of labeled datasets. Adami et al. proposed a semi-automatic process Computers in Industry 61 (2010) 66-74