A Study Of Methodologies For Adolescent Reading Recommending And Guiding Towards E-Bibliotherapy

Bibliotherapy has proven its effectiveness in dealing with mental disorders through the long-term academic research and clinic practice.Nowadays,with respect to the increasing adolescent psychological problems,this paper studies how to ease their stress by means of e-bibliotherapy and proposes new methodologies for reading recommending and reading guiding towards adolescents.This thesis mainly consists of two parts:1)Building adolescent reading recommender system named Teen Read,to automatically recommend multiple articles to users every week based on their stress,interests and reading behaviors.Teen Read unify the article profile and the user profile into a common stress-interest feature space,and make recommendations in a contentbased filtering way.After a reasonably long time of reading,Teen Read will finetune the user profile to reflect her/his real stress and interest status according to the adjustment of user' stress setting and the dynamic reading behaviors?2)Generating instructive questions for the multiple recommended articles to guide reading.Such a question shall a)attract teens' attention,b)convey the essential message of the recommended articles so as to improve teens' active comprehension;and most importantly c)highlight teens' stress to enable to generate emotional resonance so as to improve the stress easing effect.We proposed four summary encoder-decoder solutions to tackle the task and constructed a topic-specific naturally annotated large-scaled QA dataset named Teen QA to train and test the model.In order to examine the quality of recommendations,we conducted two user studies involving about 30 volunteers.The results showed that users' stress level decreased by22% after six weeks' reading meanwhile their interests to the recommended articles were all scored above 3.2(out of 5).We also evaluated the quality of the generated instructive questions with human and automatic evaluation metrics.It showed that our ED-So CF(Encoder-Decoder with Summary on Contexts with Feature-rich embeddings)solution could generate fair good instructive questions in the human evaluation metrics(all scored above 3.2)and achieved the state-of-the-art performance in the automatic evaluation metrics(especially in semantic similarity metrics).