Research On Product Attribute Extraction Based On Multimodal Theory

As an objective description of products,product attribute is one of the basic features of ecommerce system,which is widely used in various scenarios of e-commerce.As the main means of product attribute acquisition,product attribute extraction plays an extremely important role in the supplement of product attribute.Improving the precision and recall rate of product attribute extraction model can save a lot of labor costs and improve the service level of e-commerce platform.Conventional product attribute extraction methods are generally based on text and a small number of attributes,which cannot be extended to large-scale product attribute extraction and do not make full use of additional information such as product pictures.Based on this,this thesis proposes a multimodal product attribute extraction method based on question answering,which solves the problem of large-scale attribute extraction,effectively extracts the product picture information,and improves the model effect.Specifically,it mainly includes the following two aspects:· Question answering based product attribute extraction framework.The traditional method based on sequence annotation cannot be applied to large-scale product attribute extraction because it is a multi-classification model.And question answering based product attribute extraction framework regards the product title and image as the context,and the attribute name as a question.The model obtains the corresponding attribute value by annotating the answer to the current question in the context,and transforms the traditional multi-classification problem into binary classification problems,large-scale attribute extraction can be realized in parallel.At the same time,attribute information provides prior knowledge and improves the model effect.· Attribute-Aware Visual Attention mechanism and Filtration Gate.Product picture contain both useful information and noise.Effectively extracting picture information through a reasonable method can improve the effect of the model.By combining product title,attribute and picture,the mechanism enables the model to focus on the most relevant part to the current attribute extraction due to the addition of different attributes of the same title and picture.The Filtration Gate controls the proportion of the image information,which combines features from different signals that represent the information needed to solve the problem better.