Fine-grained Opinion Mining For Commodities Based On Deep Learning

In the era of big data,shopping is gradually changing from offline mode to the mode of combining online and offline.Online commodity evaluation can reflect consumers' attitudes and opinions on commodities to a great extent.This paper makes an in-depth study from three aspects,including the extraction of aspect words in the evaluation,the prediction of the evaluation dimension,and the emotional analysis at the level of the evaluation dimension,so as to dig out the fine-grained subjective emotions in the commodity evaluation.In combination with the deep learning algorithm,the advantages and disadvantages of target commodities in multiple evaluation dimensions are analyzed from the perspective of consumers,so as to help consumers get better shopping experience.First,the Soundex phonetic tags and Brown clustering tags are proposed for fuzzy processing of the common problems in evaluations,such as "erroneous characters","nearphonetic characters" and "colloquial expressions".Experiments show that adding label features to text expression can improve the results predicted by the model,increasing the recall rate by 3% and macro F value by nearly 1%.Secondly,the dimensionality recognition model of commodity evaluation proposed in this paper combines the design of full connection layer and sliding window in the hidden layer mapping,which also makes the convergence efficiency of the model higher.On the premise of not affecting the prediction results,the training time is reduced by nearly half.Finally,combining with the extracted evaluation terms and context,this paper innovatively proposes the use of pooling and attention mechanism to carry out multidimensional weight learning,so as to realize fine-grained classification of commodity review emotions.Experiments on SemEval's aspect-based emotion analysis task standard dataset,obtain macro F values of 77.95% in catering field and 69.59% in notebook field.The result was 2% higher than TD-LSTM model and 0.8% higher than the best ATAELSTM model.But it is only 0.7% lower than the TD-LSTM model on the Twitter dataset.However,the Twitter data set belongs to the coarse-grained emotion analysis corpus,which also indicates that the model proposed in the text is more advantageous in finegrained emotion analysis.Through the above three aspects of work,the paper classifies the scattered commodity evaluation opinions and forms the commodity opinions at the evaluation dimension level.The research results can be applied to the research fields such as emotion analysis,relationship extraction and commodity recommendation.The further work of this paper includes the following two aspects: solving the problem of missing words in evaluation and the problem of diversity of evaluation dimensions.