基于抽取的高考作文生成

Automatic writing is an important research direction in the field of Artificial Intelligence and Natural Language Processing. However, the traditional automatic writing methods mainly focused on generating short text, such as sports news and weather forecast, and lack deep modeling of the automatic generation of discourse-level text. In this paper, we focus on the discourse-level text generation task oriented to the essay generation in College Entrance Examination. In particular, we present an extractive essay generation model for the College Entrance Examination. We formulate the task as essay generation from mind, namely taking the input as many topic words in mind and outputting an organized article (a document) with several paragraphs under the theme of the topic. The task is challenging as it requires the generator to deeply understand the way human beings write articles. In addition, after understanding the meaning of a topic word, the following challenge is how to generate a topic focused article, e.g. how to collect topic-specific "fuel" (e.g. sentences) and how to organize them to form an organized article. This is of great importance as an article is not a set of sentences chaotically. Natural language is structured and the coherence/discourse relationship between sentences is a crucial element to improve the readability of a document and to guarantee the structured nature of a document in terms of lexicalization and semantic. Hopefully, solving this problem contributes to making progress towards Artificial Intelligence. For the issues mentioned above, our proposed model consists of two major modules including sentence extraction module and paragraph generation module. First, in order to generate a high-quality essay, the extractive essay generation model needs to determine the focus of each paragraph. Therefore, we first expand the given topic with more related topic words. Then we cluster them into multiple sets. Each of them represents the focus of each paragraph. Second, the model needs to find candidate sentences that are related to each paragraph's topic. Therefore, it first finds sentences that include the given topic words. Then, we propose two methods to expand the candidate sentences set with more diverse sentences. After obtaining sentences candidate sets for each paragraph, the model needs to choose and arrange sentences to be coherent paragraphs. In this paper, we explore three methods to achieve this. Based on experiments, paragraph generation via pointer network and paragraph generation via pair-wise LSTM network outperforms paragraph generation via learning to rank. In addition, paragraph generation via pointer network achieves best result among those three methods. Through experts' evaluation, the essay our proposed model produced can reach the average scores of level two of Beijing college entrance examination, which indicates that our proposed methods have certain practical application value. For each component in the framework, we explore several strategies and empirically compare between them in terms of qualitative or quantitative analysis. We also analyze the pros and cons of each approach. Although we run experiments on Chinese corpus, the method is language independent and can be easily adapted to other language.