Research On Online Reviews Auto-reply Based On Deep Semantic Matching

With the popularity of the Internet and the promotion of e-commerce,more and more consumers choose online shopping.Online reviews and merchant replies can help consumers learn more about merchants and commodities,which affects the reputation of the merchants and has a certain impact on consumers' purchase intentions.However,there are still few studies on intelligent automatic responses to online reviews of e-commerce platforms.Based on this,this paper studies the automatic reply technology of online reviews.The main research contents include text clustering analysis,construction of sentiment analysis model and text matching method based on semantics.This paper proposes a clustering method based on Canopy + Kmeans clustering,and based on topic model with minimal domain knowledge for topic feature word expansion.First,use Canopy to perform a rough cluster analysis to obtain the number of topics in the online reviews,and then use this number as k values to perform a cluster analysis using k-means clustering algorithm to obtain the topic feature words of each topic.Taking these topic feature words as domain knowledge,and using the Anchored Cor Ex model to set anchor words and anchor strength,you can find feature words that are not easily found in these topics.In this way,the feature words of each topic in online reviews can be better found.In the construction of sentiment analysis model,this paper uses BERT pre-training model to obtain the semantic representation vectors of sentences,which are used ad the input of BiLSTM.Then use BiLSTM model for context feature extraction and classification training to get BERT-BiLSTM sentiment analysis model.After testing on the validation data set,better results are achieved than the BERT fine-tuning model and classic machine learning classification algorithms.This paper chooses a method based on the combination of text similarity and multidimensional sentiment matching in text matching,the purpose is to find the online reviews in the FAQ database that are most similar to the semantics of the input text.Specifically,the BERT pre-training model is first used to obtain the semantic representation vectors of sentences and then the cosine similarity between the input text and the sentence vector in the FAQ database is calculated to find online reviews with high semantic similarity.At the same time,the topical feature words obtained by clustering and the clause method combined with dependency parsing were used to divide online comments into short sentences containing topic attributes,and then sentiment analysis was performed on them to obtain the multidimensional sentiment of online reviews.Finally,the sentence vectors were combined Cosine similarity and matching degree of multidimensional sentiment analysis for text matching.In order to ensure the diversity of the response content,after Easy data augmentation techniques EDA operation is performed on the matched online reviews corresponding to the merchant response,the sentence with the highest cosine similarity in the sentence vector is selected as the automatic response content.Based on the above research,this paper designs and implements a prototype system that automatically responds to online reviews.The system is tested by examples to verify the effectiveness and practicability of the system.