Research On Text Title Generation Method Based On Eye Movement Attention Mechanism

With the continuous popularity of online network platforms,the texts data generated from online platforms such as online shopping,e-education,and digital entertainment has exhibited dramatical growth,users can browse hundreds of millions of texts every day.However,excessive data inevitably brings the problem of information overload and users feel tedious to read the verbose texts,but also obtain information with less value.Unfortunately,it is challenging to select and absorb the required information from a large number of texts.Automatic text headline generation methods refers to generation of a text headline with only one sentence for a given text.In view of the fact that the headline not only guides users to select the information they are interested in,but also can provide users with basic text information.Particularly,it can save users time of browsing the text.Consequently,researchers intend to mine the key text information by generating high-quality headlines to solve the problem of information overload effectively.In recent years,with the rapid development of the artificial intelligence technology,More and more researchers are constructing neural network models to automatically generate text headlines.Unfortunately,the headline generation model based on popular encoder-decoder framework is more effective only when handling short texts with less characteristics,resulting in poor performance in generating headline for long source texts.Furthermore,previous researchers on text headline generation technology mainly focused on generating text headlines based on grammatical characteristics of the source text such as word frequency and part-of-speech,while ignoring the basic human behavior data such as eye-tracking data,which results in poor quality of the text headline.To address these problems and generate a high-quality text headline,this study explores the influence of user's dwell time on different parts of the source text during reading on the quality of the generated headline,and construct an eye-tracking attention mechanism based on the correlation between the reading time and the importance of words.Then,we propose an effective neural text headline generation model based on this mechanism.Besides,we compare our proposed model with some classical text headline generation models.Finally,the experimental results demonstrate that the performance of our model is significantly superior to that of the classical headline generation models.The main contributions of this study are as follows:·This study establishes an eye-tracking attention mechanism to capture the significant information of the source text.First,we construct a reading time prediction model based on three eye-tracking datasets to obtain the mapping relationship between text words and the reading time.Then,the eye-tracking attention weight of the words is calculated according to the reading time of different words,and the words with higher attention weight obtain more attention.Finally,the eyetracking attention mechanism identifies valued parts of the source text that contain crucial information.·This study proposes an eye-tracking attention based model for abstractive text headline to generate a high-quality headline.First,in order to capture the important contents of the source text.We construct a bidirectional encoder based on gated recurrent unit to generate the context semantic vector.Next,the model integrates the eye-tracking attention weight with the context semantic vector.Finally,this integration results input into our attention decoder to generate a text headline.·This study conducts experimental to analyze the performance of reading time prediction model based on different neuron and word embedding methods.Moreover,we conduct several comparative experiments to verify the effectiveness of our proposed model.Finally,the experimental results validity of our proposed model can generate a text headline with high-quality.