A transfer learning approach to predict shipment description quality

International shipments always have a harmonized system code (HSCode) associated with them, to determine the tariff for the custom declaration. The HSCode is derived from the shipment description that the customer provides, which makes the quality of the description important to assign the correct code. When the description is too generic or incomplete, the logistic company will have to contact the customer in order to find out the content of the shipment. Due to the fact that there is no effective way to identify the quality of description, we developed a description quality evaluation model, based on deep learning combined with domain knowledge. By using a 2000 shipments data set with scores ranging from 0 to 4 provided by experts, where 4 represents the best quality possible, the developed model can classify 45.17% of the data correctly and 43.95% of the data with 1 score difference(i.e predict label 1 as 2 or 0) from the human annotated ground truth. This model can be used for historical data analysis, and potentially giving customers on-site feedback when they are providing a bad description for the shipment content.