Research On Construction And Alignment For Cross-lingual Product Knowledge Graph

With the continuous development of e-commerce in the global market,the product knowledge graph plays an important role and is widely used in platform governance,brand operation,shopping guide and other core businesses.There are some differences between the product knowledge graph and the general knowledge graph due to the various types of products and the huge attribute system in e-commerce.Therefore,this thesis mainly studies the construction of product knowledge graph based on product attributes,as well as the product entity alignment and integration for cross-lingual product knowledge graph.The product knowledge graph mainly describes the attribute and attribute value information about products.In the early research work,the rule-based methods were used to extract the attribute,in which experts designed domain related vocabularies.Some works regarded the task as a special Named Entity Recognition(NER),but unable be applied to the real e-commerce environment with huge attribute system.Therefore,the first work of this thesis proposes the attribute-comprehension open tagging model,which treats attribute beyond NER type alone but leverage its contextual representation.Experimental results show that the model can effectively deal with tens of thousands of attributes,even new attributes that the model has never seen before.Meanwhile,we construct a real large-scale product datasets,and construct English product knowledge graph base on it.There are many small languages in e-commerce globalization.Due to the limited number of users and products,these small languages lacks of corresponding labeled data to train effective tagging model,so it is difficult to build the product knowledge graph of low resource languages.Therefore,the second work of this thesis proposes the adversarial multi-task learning model,which leverage the rich labeled data of high resource language to help the model training of low resource language.The high resource language is regarded as auxiliary task and low resource language as main task.Two independent neural networks are used to capture language related features respectively,and adversarial learning is introduced to extract language independent features.Experiments on three low resource language datasets show that the proposed model can effectively improve the performance of low resource language tagging model,and ablation experiments also prove the effectiveness of multi task learning and adversarial learning.Due to the diversity of products,product knowledge graph in different languages has both intersection and difference.If the graphs of different languages are aligned and integrated,the product information will be greatly enriched.Therefore,the third work of this thesis proposes the attribute-enhanced entity alignment model,which integrates information on different granularity from attribute level and attribute value level respectively according to the different information content.Then we leverage the graph neural network to get each product entity feature,and then calculate the entity aligned with it.We conduct experiments on the multilingual product graphs obtained from the first two works,the results show that the proposed model can effectively model the product graph and the performance is better than all the benchmark systems.Furthermore,ablation experiments verify the reasonability and validity of our model.This thesis studies several hot topics in product knowledge graph,including the graph construction of high resource language and low resource language,as well as the entity alignment of cross-lingual graphs,and verifies the validity of our model on the real product datasets,which promotes the application and development of product knowledge graph.