Objective : The Chinese-Vietnamese cross-language summarization is the task of generating a summary in Vietnamese for the given Chinese text through a Chinese-Vietnamese cross-language summarization model. Keywords contain the important content of the source text, which is an effective enhancement to the source text. It provides important information guidance for the generation of text summaries, and the bilingual parallel dictionary provides a unified semantic space for bilingual texts. As a continuation of previous research, this paper studies mapping the source language text keywords to the target language based on the Chinese-Vietnamese probability mapping pair, to solve the problem of difficult semantic alignment and poor quality of Chinese-Vietnamese cross-language summaries when the annotateds Chinese-Vietnamese alignment data is scarce.
Methods : Taking Chinese text and Chinese keywords as input, based on the Transformer framework and pointer generation network. The size of the Chinese dictionary is set to 100 000, the size of the Vietnamese dictionary is 10 000, the size of the probability mapping dictionary is set to 39 311 according to the word frequency, and the number of keywords is set to 5 according to the average length of the abstract, and the experiments are carried out on single NVIDIA RTX 2070 SUPER GPU. First, obtain the joint representation of the source language keywords and the source language text. Then we map source language keywords to the target language based on the probability mapping pair. The final distribution is obtained based on the pointer network by combining the word probability generated by the decoder.
Results : In this paper, a Chinese-Vietnamese cross-language summarization model combined with keyword probability mapping is constructed. Then, we comparison with NCLS and other methods on the constructed 100 000 Chinese-Vietnamese cross-language summarization dataset. The results show that better contextual representation can be obtained by obtaining the joint representation of the source keywords and the source text, and the fusion of keyword probability mapping to guide the summary generation can effectively improve the performance of the model, it can also effectively alleviate the problem of poor summary effect caused by poor Vietnamese machine translation performance. In addition, we also explored the influence of the number of keywords and the size of the probability mapping dictionary on the performance of the model. The experimental results show that the guidance of the summary model with prior knowledge such as keywords can effectively improve the performance of the low-resource summary model. The coverage of the keywords in the probability mapping dictionary affects the performance of the model to a certain extent. Small-scale probabilistic mapping dictionaries will affect the performance of the model due to their low coverage and high noise. In addition, we also explored the influence of the fusion method of the keyword probability mapping information on the model performance. The experimental results show that under the same experimental conditions, the fusion method of the pointer network in this paper is more effective than the direct splicing method.
Conclusions : The results of above prove that in the case of low resources, by obtaining the keyword information of the source language text and mapping it to the target language to guide the summary generation. There is a certain improvement in the Chinese-Vietnamese low-resource cross-language summarization task. It can also be proved by experiments that the keyword probability mapping information can provide richer guidance information for the cross-language summarization model, and it is also proved that the method proposed in this paper may be more effective for low-resource cross-language summarization tasks. Multimodality and other multi-source information is a high-level generalize of the text content, which can be a good supplement to the text content. Therefore, how to use multimodality information to guide cross-language summarization is the focus of future studied.