Research On Automatic Evaluation Of English Recitation And Retelling Test

With the development of computer technology, CALL (Computer Assisted Language Learning) system has become more and more intelligent. CALL system can give students'pronunciation proficiency evaluation objectively and accurately, thus greatly foster their oral English learning efficiency. It can also help teachers give scores of oral tests much more objectively. Nowadays, technologies of automatic oral language proficiency evaluation for CALL systems are all text-dependent and becoming more and more mature. In the first three questions of PSC test (characters reading, words reading, article reading) CALL system even outperforms human raters who devoted many years in the scoring task.Recitation and retelling are important methods of learning English. They can better reflect students'expressive ability than reading. In recitation task, although students'must recite according to the given texts, they cannot refer to them during reciting process, result in much mismatch between actual oral pronunciation and given texts. From this point of view, we can interpret it as a problem between text-dependent speech evaluation and text-independent speech evaluation. The evaluation of Retelling is text-independent, students'can express texts'main ideas in their own way. The CALL systems nowadays can't perform well on such tasks that do not strictly dependent on text. It greatly hampers the full development of CALL system.Thus based on the work previously done on reading task, we extended our work to text-independent oral speech proficiency evaluation and carried out series pioneering researches on two typical problems: recitation and retelling test. In recitation task, we adopted sentences parallel networks to do the recognition. Such idea brings about little confusing networks during speech recognition and well depicts the mismatch between texts and student's actual pronunciation. The system achieves 90% of expert's scoring level and applied into practical use. In retelling evaluation task, by combining technologies of speech recognition, natural language processing and many improvements based on the characteristic of retelling task, our system achieved 84% of expert's scoring level. It outperforms mass-scoring quality of teachers and can help them to do the evaluation much more objectively.The experimental results of this thesis demonstrated that text-independent speech evaluation is feasible. Our work established the foundation of future work of insight research of text-independent speech evaluation.The whole thesis is organized as follows:Chapter 1 gives a brief introduction on the background and the development of speech evaluation and explains the system structure speech evaluation.Chapter 2 describes the principal, system structure, classic methods of text-dependent speech proficiency evaluation in detail.Chapter 3 mainly focused on reciting proficiency evaluation, and proposed the"sentences parallel networks"which is more agility than its reading task counterpart but still strongly restrictive in recognition. It achieved ideal performance.Chapter 4 mainly focused on retelling task. We firstly constructed retelling proficiency evaluation system. Next we aimed to the characteristic of retelling task: students'needn't retell according to the given text, the data computer received at spot are noisy, we introduced many method to cope with them: using reading data to do speaker adaptation, language model pruning based on the given text, evaluation feature extraction based on word graph and so on. These efforts greatly improved system's performance.The final chapter concludes the thesis. The possible improvements are also discussed here.