Research On Immune Multi-word-agent Autonomy Learning Based Sentiment Analysis

Sentiment texts refer to comments on persons, things, or events with sentimentpolarities. Sentiment information in these comments reflecting people’s attitudescontain evaluation holders, evaluation objects, evaluation words, and their modifiers.Sentiment analysis on sentiment texts refers methods to identifying and extractingsentiment information from sentiment texts with Natural Language Processingtechnologies. However, characteristics of sentiment texts give a challenge to thesentiment analysis on sentiment texts. For example, firstly, sentiment polarities ofwords affected by contents cannot be judged accurately. Secondly, it is difficult toconstruct accurate relations between evaluation objects and evaluation words.Finally, sentiment expressions are various. In recent years, many researchers havecarried out research of sentiment analysis on texts at different levels such as words,phrases, sentences, and texts and tried to apply it to Product Recommendation (PR),Question Answering (QA), and Information Retrieval (IR) to improve performance.Although statistical machine learning based sentiment analysis on texts attractsextensive attention as the mainstream approaches, there are still deficiencies of theseapproaches. Firstly, the large-scale labeled corpus is required for training andevaluating models. Then, these models cannot learn continuously to improve theirperformance. An adaptive immune theories based machine learning model ispresented to overcome these two deficiencies. A semi-supervised model based onset-similarity joins is presented to mine sentiment words and sentences. Althoughaccuracies of mining sentiment words and sentences were improved, it revealeddeficiencies of statistical machine learning based sentiment analysis models fully,where overcoming these deficiencies became aims of constructing the new machinelearning model; Then, inspired by similarities between the human immune systemand the human language system, an autonomy learning model is presented based onadaptive immune theories, which included two steps:(1) the artificial immunesystem based on multi-agent system modeling is constructed with the plasmanegative regulation mechanism as the foundation platform for constructing the newmachine learning model;(2) a multi-word-agent autonomy learning model isconstructed through simulating words as immune cellular and molecular to optimizerelations between words by interactions between immune word-agents; Finally,sentiment analysis on collocations is realized by the autonomy learning model; Themain contents of our research work include four aspects as follows.Firstly, in the research of sentiment analysis based on statistical machinelearning methods, set-similarity joins based semi-supervised sentiment analysis is presented. The model is built based on the flow graph of sentiment candidates.Firstly, sentiment candidates are extracted from corpus and used to construct theflow graph with a sentiment lexicon, where nodes in the graph are these candidates,and edges are semantic relations between candidates and primary sentimentpolarities. Then the graph is cut into sub-graphs with the Ford-Fulkerson algorithm.Finally nodes in these sub-graphs can be merged into positive and negative sets byset-similarity joins. And the performance is improved further by the self-trainingbased semi-supervised method.Secondly, in the research of constructing the foundation platform, the plasmanegative regulation mechanism based artificial immune system is proposed. Plasmacan bind T cells to kill them so as to reduce the diversity of the T cells population,which improves the efficiency of interactions between T cells and B cells in order toimprove the efficiency of immune responses. An artificial immune system isconstructed by simulating immune cells and molecular as agents with theories suchas clonal selection, negative selection, idiotypic immune network, and the plasmanegative regulation mechanism and the cellular automaton based complex systemmodeling method. Experimental results show that not only processes of adaptiveimmune responses can be simulated, but also plasma negative regulation mechanismcan improve efficiencies of these responses.Thirdly, in the research of constructing the new machine learning model, wepresent the adaptive immune theories based multi-word-agent autonomy learningmodel. Words are simulated as immune cells and molecular to construct immuneword-agents with adaptive immune theories and autonomy oriented computingbased complex system modeling, where relations between words are simulated asspecific relations between receptors of these cells and molecular, and strengths ofthese relations are measured by matched degrees called affinities of these receptors.These immune word-agents can learn continuously to optimize relations betweenwords through communicating among them, cloning, mutating, and selecting.Finally, on the basic of all above researches, we adopt the multi-word-agentautonomy learning model to realize sentiment analysis on collocations. Firstly, thesystem object function for sentiment analysis on collocations is constructed. Thenwords are simulated as B cells and antigens involved in adaptive immune responses,and immune word-agents are built by simulating behaviors, states, and interactivepolicies of B cells and antigens. Finally, relations between words are optimized byadaptive immune responses so as to optimize relations between evaluation objectsand evaluation words continuously, which is to optimize the object function.In conclusion, the main object of our research is to construct an autonomylearning model based on adaptive immune theories through simulating behaviors,states, and policies of immune cells and molecular in adaptive immune responses, and the model is applied to overcoming deficiencies of present sentiment analysismodels. This research has achieved some preliminary results. We believe that thedeep research on the model and discoveries of immune theories can promote thedevelopment of the model in the future.