Finding New Values

Once you assigned labels to some of the tokens in Unigraph screen, you may use the New Values button to run the machine learning algorithm that finds new attributes and recommends labels based on the labels that you have assigned and approved so far. Alternatively, you may go to the Digraph screen and label some digraphs before running the algorithm. Note that assigning and approving unigraph and digraph labels and running the algorithm to find new values is an iterative process; you can repeat these two steps as frequently as you want until you are satisfied with the quality of the results and number of attributes extracted. If the machine learning algorithm does not find enough new values, it is an indication that it requires more information to discover the patterns and be able to recommend new values, and you might want to label some more unigraphs and digraphs before re-running the algorithm.

To run the machine learning algorithm, first select a mode from the drop-down menu next to the New Values button on the top right. You can select from the following three modes:

Table 6-4 Machine Learning Algorithm Modes

Mode	Description
Random	Randomly partitions the data into a training set and a test set to be used by machine learning algorithm.
By attribute	If you select this mode, you will be prompted to select one attribute. The data is then partitioned in such way that all description strings that have one or more tokens labeled as the selected attribute are used as the training set. The remaining descriptions strings are used as test set.
By annotation	The description strings that are 80% labeled are used as training set. The remaining descriptions strings are used as test set.

The By Attribute mode is most effective in the early stages, where very few attribute values have been identified.

Note that after you click the New Values button, it may take few seconds for the machine learning algorithm to run and find new attributes. Then a pop-up message shows how many new values were found. The Machine Label and New Discovery columns in the Annotation section on the top left will be populated to show the new labels and indicate which tokens were labeled by the machine in the most recent run.