Text Generation from Abstract Semantic Representation for Summarizing Vietnamese Paragraphs Having Co-references

Generating new short text from an abstract semantic representation (ASR) of the input text is a further step in “abstractive” summarization direction. The summary will have high quality in the sense that having the expression structures are similar to human communication in the applied language. Two essential questions should be answered when following this idea: The mechanism for constructing appropriate ASR for source text? The mechanism for generating the summary? This article presents a method to solve these two questions when summarizing short Vietnamese texts with coreferences. Our approach consists of three main phases. The first phase is to build a graph which resolves all co-references and represents the input content. The second phase is to transform the graph into an output ASR. The final stage is to generate new Vietnamese sentences from ASR and complete the summary. The evaluation of summarization is performed automatically with Rouge indices and human assessment. Comparing to two-word graph-based summarization methods, the method in this study generates text relevant to the native speaker and contains rich information.