Text automatic summary generation

According to the output type, automatic text summarization can be divided into extractive summarization and generative summarization.

According to different classification standards, text abstracts can be divided into many types. According to whether the data is marked or not, it can be divided into supervised and unsupervised.

According to the input type, the text summary can be divided into single document summary and multi-document summary.

Advantages: it has certain guarantee in grammar and syntax;

Disadvantages: wrong content selection, poor consistency and poor flexibility.

Extraction method selects keywords or key sentences from the original text to form a summary. This method has a low error rate in grammar and syntax, which ensures a certain effect.

Traditional method: using graph method and clustering to complete unsupervised summarization. It mainly includes Lead-3, text ranking and clustering.

At present, the popular method is to model the problem into two tasks: sequence labeling and sentence sorting.

Advantages: It allows the abstract to contain new words or phrases, which is very flexible.

Disadvantages: The generation process often lacks the control and guidance of key information.

At present, the sequence-to-sequence (Seq2Seq) model is widely used in the task of generating abstracts, and some achievements have been made.

Considering that the model based on Seq2Seq is often unfriendly to the generation of long texts, we can use real abstracts to guide the generation of text abstracts. The core idea is that the abstracts of similar sentences also have certain similarities, which are used as soft templates and assisted by external knowledge.

The task can be roughly divided into two steps: first, select the important content, and then rewrite the content.

The basic structure of generating neural network model is mainly composed of encoder and decoder, both of which are realized by neural network.

Greedy search algorithm

Classic text abstract baseline model: the text abstract model of Seq2Seq combines attention mechanism and pointer to generate network model.