In addition to the Averbis Multi-Label Classifier, Averbis also offers the option of a binary classifier. This classifier has already been used for relevant/irrelevant detection.
To make it easier to control its use, the binary classifier has now been integrated into the Classifier / Automatic Classification. This means that the binary classifier can be trained on two folders. The result is a trained classifier that assigns patents to either one folder or the other (never to both folders).
In the following example, a binary classifier was created for the relevant/irrelevant check. To do this, the classifier is trained with two folders. One folder contains the relevant patents from a monitoring task, while the other folder contains the “irrelevant” patents from a monitoring task. The confidence level was set relatively high at 80%. Retraining was excluded (retraining frequency: Never).
After training is complete (“Start training”), an evaluation appears showing which patents were used for training (used) and which were not (not used). This depends on whether sufficient information, such as machine translations of the selected texts, is available for this simple patent family. A false/positive analysis is also provided. After training, this analysis sends all documents to the classifier, allows them to be assigned, and compares the result with the actual assignment. The correctly assigned patents are displayed in the Positive column and the incorrectly assigned patents in the Negative column. The number displayed is clickable and opens the corresponding list of results for the documents.
Automatic Classification
If a binary classifier is selected in Automatic Classification, the user interface is adjusted accordingly. The two binary folders (e.g., for relevant and not relevant) must then be selected. Clicking on Execute will run the automatic classification directly.
The binary classifier can thus be used specifically for a relevant/irrelevant check. To do this, two folders with relevant and irrelevant patents must be created for training (folder 1 contains the relevant patents). An automatic classification is then created that refers to the binary classifier. The target folder of the automatic search is used as the input. The result (folder 1) is written to the new input folder of the monitoring task. The non-relevant patents (folder 2) should be made available to the expert for review (write permissions for the monitoring task group).
This is also possible with the new extended workflow model.