Research And Design Of The Automated Translation Scoring For Chinese College Students

The research in the field of human translation is not deep enough, even though there are relatively mature reseach in the area of machine translation evaluation(MTE) and the automated essay scoring. The essay scoring, another type of subjective ques-tion, used multiple linear regression method to establish the equation between the text feature and the score. But the relation between the text feature in human scoring is more complicated. If we build the model in the same way as essay writing does, we may not get the best results. The neural network,as an adaptive learning model, have an advan-tage in dealing with the relationship between the complex variables. The dissertation has tried to bulid a Chinese-English(C-E) translation scoring model, on the basis, the automated translation scoring system of College Student will be built.Firstly, text features will be extracted from translation passages in the direction of language base, semantic, consistency and test points. There are several design and improvement of the algorithm such as:(1) A new algorithm to extract the latent semantic based on the set of similar translation passages.(2)A new way to calculate the value of consistency with the considering of the conjunction weight, solving the unreasonable phenomenon of the negative correlation between the number of conjunction and the score. Secondly, the dissertation choosed the BP neural network to build our model after a comparison between the neural network and linear regression. As the pure BP neural network have some shortcomings such as slow convergence and not easily access to the global optimization, we used the PSO algorithm to optimize it. Finally, on the base of feature extraction and model building, the writer has bulid a translation scoring system for Chinese college student,the system can score the translation passage and generate evaluation.The algorithm based on the set of similar translation passages have achieved the positive correlation with the scoring, and the correlation was promoted by1percent further when we take consideration of the conjunction weight. There are significant effects when we used the latent semantic ratio algorithm based on the similar trans-lation passages. The correlation was premoted by24percent singly in the similarity range of0.9to1.0. In the model construction, the dissertation verified the result of the multiple linear regression and PSO-BP neural network. The result shows that PSO-BP neural network can premoted by almost6percent in comparison of the multiple linear regression.