3.2.1.

Speech recognition and machine learning

As a speech extension, speech evaluation is one of the downstream tasks of speech recognition, which automatically evaluates the pronunciation level through intelligent speech technology, which requires local explanation, that is, the comparison of departure sound errors, defect location and problem-solving analysis.

In 2016, Ribeiro of the University of Washington (University of Washington) gave the definition of black box-based interpretation (Model-Agnostic Interpretability) in machine learning. Further discussion on interpretability is becoming more and more important in machine learning. The paper “Why I believe you” by Ribeiro et al¹⁰. explains the prediction of the classifier “why should I trust you?”: Explaining the predictions of any classifier. A local interpretation framework called LIME (Local interpretable model-agnostic explanations, LIME) is proposed. The algorithm uses an explanations alternative model to make local prediction. These alternative models do not necessarily have better accuracy but have better transparency. The better the transparency of the predicted results of the classifier or regression quantity, the more human beings can understand the explanation given by the model in their own way. For the black box model, as long as we input a training data, the black box model can output a prediction.

LIME first selects one of many models with good interpretability, such as Lasso or decision tree or linear regression model, and then locally perturbs (or permutes) the data points of the interested sample instances to get a new data set composed of disturbed instances. Finally, a better interpretable model is trained for this new data set LIME, which is a local substitution model (Local surrogate models). Suppose x is the data point of the sample instance of interest, G is the set of all alternative transparent models, the specific transparent model is g, and the loss function L is selected, then theoretically, Ribeiro et al give a specific single sample instance x and its corresponding prediction model g mathematical formula as follows. As a function of transparent model g, the number of participating features is required to be as small as possible, which is a sparse representations measure. The disturbance measure represents the size of the new data set. From the traditional neural network to the current deep learning technology, what is the key to distinguish them? Literally, it is the word “structure”, because the traditional neural network is also called shallow learning technology. But now more and more people think that deep learning is essentially a new way of programming. In 2015, the DeepMind team published a research paper entitled “Neural Program interpreter” (Neural Programmer-Interpreters) in ICLR’16, which was one of the best papers of the year. As a recursive synthetic neural network, the core module of NPI is a sequence model based on LSTM. NPI is a program-memory neuro-Turing machine (Neural Turing Machines), which has three learning capabilities: a recursive kernel with unknown tasks, an embedded continuous key program memory for learnable programs, and a domain-specific encoder that enables a single NPI to provide completely different functions in multiple perceptual environments. The program is generated by the controller RNN one operation at a time, and each operation is constructed on the basis of the existing program to generate an operation sequence. Because attention is introduced on the basis of probability distribution, the output of the program is differentiable in probability. So NPI learned to generate programs without the correct program paradigm. When using LIME software package, similar to the neural program interpreter proposed by Scott Reed, the corresponding interpreter should be built first when the program is implemented.

Explainer=lime.lime_tabular.LimeTabularExplainer (class_names=y_train.unique.

Values. Values.tolist (), values.tolist ().

Categorical_features, categoriesical_names, Kernel width).

The parameters used by this function are:

X_train = training set. Connection list of all feature_names= feature names. Class_names= target value. List of classified columns in the categorical_features= dataset. List of categoriesical_names= classification column names. Kernel width= parameters control the linearity of the induced model, the larger the width of the model, the greater the linearity. The new data set obtained from the disturbance example is a data enhancement technique. Basic data enhancement techniques include a series of fixed transformations (such as horizontal flipping, filling, and cropping). Specific to the speech evaluation here, the key is the combination, that is, to properly replace some speech segments and test their scores. Combined with the neural program interpreter, the computer can learn the program, so that the machine has the ability of self-evolution. Then the locally interpretable alternative model specified in the LIME package should be self-constructed.

3.2.2.

Questionnaire

The freshmen and sophomores who are non- English majors in a certain university are selected as the experimental subjects. The entrance scores of these students are in the middle level in the whole school and are representative to a certain extent. Among them, there are 65 boys and 100 girls, aged 18 or 19 years old. 130 students are from the city where the school is located, and the others are from the province outside the city. A total of 160 students participated in this study, but only 110 students participated in the whole experiment and all the corpus could be analyzed. Based on the fact that the high dropout rate may not be random, resulting in that the final sample can not objectively represent the initial sample and the research object population ¹². At the end of the experiment, the independent sample T test of the students who dropped out and the remaining students was done, and it was found that there was no significant difference between the two groups. The purpose of this study is to help researchers draw the usefulness and limitations of EAP Talk through the evaluation of the program after the periodic use of EAP Talk. In the course of the experiment, a mixed method of questionnaire survey and semi-structured interview is used, and the qualitative results of semi-structured interview are helpful to analyze and explain the quantitative results of the questionnaire. The interview questions are designed around the perception of students’ oral skills and software restrictions after using EAP TALK, and then qualitative data are collected and analyzed to explain the results of the questionnaire, so as to collect students’ rich cognitive feelings after using EAP TALK, so as to obtain information about the practicability of EAP Talk in improving oral English skills and the limitations of its use in practice. The experimental data were collected by a specially designed questionnaire through the “questionnaire Star” program, which aims to further elicit the relevant research objectives for appropriate answers. There are 18 specific questions in the questionnaire, which are carried out by the following three dimensions:

Dimension I: perception of the improvement of oral English by EAP TALK.

Dimension II: perception of the relationship between EAP TALK and offline Teaching.

Dimension III: perception of the functions of EAP TALK software.

Participants’ views were graded on a Likert five-point scale¹¹with questions ranging from “Strongly Disagree” to “Strongly Agree”.

After the questionnaire was collected, the data were statistically analyzed.

The five-point score of the scale is as follows:

Strongly disagree = 1.

Disagree = 2.

Neutral = 3.

Agree = 4.

Strongly agree = 5.

3.2.3.

Interview

Subsequently, we randomly selected six students in the sample to conduct a semi-structured interview to explore the results of the questionnaire and to understand the students’ views on the limitations and follow-up development of the software after using EAP Talk.

The details of the interview are as follows:

The contents of the interview were transcribed, and the answers of all participants were initially coded in the same order as the interview questions, and then analyzed. The student interviewee number is S1 (student 1, 2, etc.) and the gender is Whip M (female / male).