Research On Sentiment Analysis And Dialogue Generation In Social Media

With the rapid development of the Internet,the way people use the Internet has been changed—people deliver information via the Internet rather than a reader.Massive information with sentiment is broadcast,fermented,aggregated and collided in social media,and further change the real world.Analyzing,understanding and generating text with sentiment now are becoming hot topics in natural language processing researches.The current phase of textual sentiment computation on social media is moving from simple sentiment classification to sentiment understanding based on user profiling,but it usually lacked researches on stance detection combining external knowledge and causal inference.At the same time,the study of dialogue generation for guiding social media is also moving from the simple dialogue scene to complex dialogue scenes integrating sentiment and stance.To this end,this research proposes a comprehensive framework of sentiment analysis,stance detection,sentimental dialogue generation and stance-oriented dialogue generation.To address the above four questions,this paper focuses on the following research:For sentiment analysis,traditional neural attention mechanism is usually lack of global perception capability.However,multi-head self attention mechanism has overly large number of parameters,thus runs in low efficiency.To address thses issues,by inspiring human's template attention during reading procedure in neuroscience,this research proposes a novel convolution-based attention mechanism.The proposed convolutionbased attention mechanism uses one-dimensional convolutional operation to simulates human attention and augment the perception field by only introducing limited parameters.Based on the proposed attention mechanism,this research proposes a convolution-based neural attentional network.At the same time,making full use of the global emotion perception capability of the convolutional attention mechanism,this research investigates a continuous textual emotion expression extraction algorithm.Experimental results on sentence and document-level sentiment analysis datasets show that the proposed convolutionbased neural attentional network significantly outperforms traditional models.The user's stance is usually implied in the semantics expressed by the text and the target.Only by understanding the semantics of the text and the target can the stance be correctly determined.To solve the problem of understanding the semantic of the target,this research introduces target-specific embedding and attention mechanism,and proposes a convolution-based neural attentional model combined with the target.The proposed model can fuse information of target in both word and sentence level.To solve the problem of understanding the semantic of the target,this research introduces large-scale structured external knowledge as the background and evidence to greatly improve the performance of stance detection.Specifically,aiming at the heterogeneity of text and knowledge representation,a Knowledge Enhanced Neural Memory Network based on post-fusion mechanism is proposed.Separated memory units are used to represent text and external knowledge respectively,and the complementary attributes between text and knowledge are used to efficiently select external knowledge related to stance detection.Aiming at the noise and sparseness of external knowledge in the knowledge fusion,a Multi-view Knowledge Aware Network is proposed,in which the knowledge entity is used as the modeling perspective to model the external knowledge at a higher level.Through the fusion network and attention mechanism at the view level,the influence of noise in external knowledge is reduced.The experimental results on two stance detection datasets show that the proposed convolutional attention model combined with position object achieves certain performance improvement;the two position analysis methods combined with external knowledge further improve the performance of position analysis,reaching the highest known performance.For sentimental dialogue generation,current models usually focus on the capability of semantic expression and sentimental expression.To jointly improve this two capacility,this research proposes a variational autoregressive autoencoder.The proposed model incorporates the process of variational inference to each generation step,which globally models the semantic information of sentiment and dialogue context,thus improve the fluency and relatedness of generated responses.By leveraging the autoregressive property of the proposed model,this research proposed a novel model incorporating the continuous sentimental expression with variational autoregressive autoencoder.In this model,the continuous sentimental expressions are extracted by the convolutional attention-based model proposed before.By using continuous expression,the proposed model is able to generate more fluent and sentimental responses.Experimental results of automatic evaluation of relatedness,fluency,and diversity,and human annotation on sentimental dialogue generation datasets demonstrates the performance of the proposed model is superior to the baselines model.Analysis of further experiments shows that responses generated by the proposed model are more fluent and accurate from the sentimental aspect.Based on sentimental dialogue generation and knowledge-grounded dialogue generation,by introducing the targets and stance labels,this research proposes and defines the stance-oriented dialogue generation problem for the first time,and decomposes the problem into two sub-problems: the knowledge-grounded dialogue generation and the dialogue generation with stance labels.To solve the problem of the imbalance between knowledge retelling and dialogue interaction,this research proposes the implicit variable variational interpolation method to explicitly model the dialogue context and the influence of external knowledge and proposes a variational interpolational encoder-decoder model to balance the relationship between them,which greatly improves the readability and knowledge utilization of the generated responses.In order to solve the problem of fusing target and stance labels,this research proposes a novel attention mechanism,which integrates the given target and stance labels into the dialogue generation.Inspiring by the idea of Generative Adversarial Network,this research proposes to leverage stance discriminator to distinguish the stance label of generated responses and uses the policy gradient method to backpropagate the supervision signal to the generator,which effectively improves the stance-oriented expression.The experimental results on a large-scale stance-oriented dialogue generation dataset show that the proposed model performs significantly better than the baseline model when evaluated by the fluency and stance accuracy of the generated text.