Auto-Evaluation Study On The Speakability Of English Spoken Text

Speaking is a crucial part in second language learning.With the rapid development in internet technology and educational informationization,spoken language materials become more available and plentiful.Therefor how to provide learners with personalized materials to serve their dynamic ability has been a more importance topic in educational technology.In order to improve the adaptive learning in speaking field,this paper proceeds from the factors influencing spoken items difficulty.Difficulty of reading items has been an object of research since the 20 century,and listening items are also brought into researchers' focus in recent days.However,a few researches refers to spoken items.According to the results of existing literature,the purpose of the research reported in this paper is to build an automated process for measure the speakability of spoken text.The primary aims are:1.To assess the influencing factors of English spoken texts' speakability;2.To develop a solution for quantifying large-scale spoken text.The main contents of paper include:First of all,the study collects a bunch of logging data from a spoken language assessment system.Due to the environment noise and improper operation,measurement error is increased.Therefor,symbolic regression is adopted to calibrate this score,taking the mean score from expert ratings as dependent variable and score from audio analyzer as independent variables.This adjustment aims to make research more accurate.Next,the paper introduces algorithms in educational measurement field,including Classical Testing Theory(as know as CTT)and Item Response Theory(as known as IRT).Research generalizes models and use cases from IRT family and discussed different estimating methods of difficulty.In the experiment part,it selects Partial Credit Model as the base model,estimates item difficulty by mini-batch and anchor items.Finally,the paper lists some influencing factors for the speakability of spoken text,trying to define and extract.The experiment part takes item difficulty as dependent variable and influencing factors as independent variables to select features,build model and fine-tuning.Results shows phonetic features have a greater impact on the difficulty of spoken items,such as phonemes,syllables and accentsThis paper elaborates thinking,design and practice in the research of spoken item difficulty,provides a complete analysis process,including data calibration,extraction,selection and model construction,validation,tuning and explaination,achieving the goal for evaluating the factors and quantifying spoken text.