Chart English abstract generation method based on fusion spatial position attention mechanism

Abstract

The invention discloses a chart English abstract generation method based on a fusion spatial position attention mechanism. The method comprises the following steps: 1) creating a chart English abstract description data set; 2) replacing a chart data value with a data variable; 3) performing word vector position coding based on a spatial relationship; and (4) searching a word vector result by adopting Diverse Beam Search. The method comprises replacing a chart data value with a data variable based on a fusion spatial position attention mechanism; learning a relationship between words by adopting a spatial attention mechanism; enhancing the spatial position relationship between the word vectors and correct word position sorting; and using a Diverse Beam Search to search for a better word vector result, so that the quality of generating the chart English abstract can be improved.

Description

technical field [0001] The invention relates to a computer natural language generation technology, in particular to a method for generating English summaries of graphs based on a fusion spatial position attention mechanism. Background technique [0002] Data visualization presents high-dimensional and complex data in intuitive forms such as bar charts and line charts. However, according to relevant research, there are certain difficulties in the analysis and utilization of charts in practice and are not fully utilized. Research on the graph corpus shows that 35% of the descriptions cannot express the key information conveyed by the text through the traditional visual way of observing graphs, and 26% of the descriptions can only express a small part of the expected information of the graph. However, using English abstracts to analyze and describe the content of the chart can achieve the purpose of reducing the difficulty of chart analysis and making the chart more intuitive a...

AI Technical Summary

Problems solved by technology

Among them, the traditional English abstract generation method uses the corpus to focus on the data table "what is said" and "in what way". They use the method of statistical learning to search for the abstract describing the data table. Once the style of the data table changes, the The method will cause its description to be distorted, so this type of method lacks versatility; the time-series English summary generation method can analyze the key information expressed in the data table through the time in the data table and the causal relationship between the data, although the time-series English summary generation method method can also generate a descriptive summary of the data, but there are the following problems:

[0004] (1) This method is different from other NLG tasks. For example, sentences and words in machine translation appear in pairs, and training data is easier to generate. However, data visualization tasks require structured data, and there are problems in the process of obtaining data. certain difficulties;

[0005] (2) This method equivalently marks the instance data (name, value, etc.) in the abstract, which may lead to incorrect abstracts predicted by the model;

[0006] (3) In the process of generating data description summaries, each group of source sentences and target sentences are equivalent to each other, without considering the possible spatial position embedding relationship between each group of sentences, and the word vector search algorithm is not used in the generation process, May generate illogical English summaries

Although statistically-learned models can also generate data descriptions, they use pre-defined templates to generate summaries, resulting in a lack of versatility in these methods and fewer choices in terms of grammatical style and vocabulary

In the model based on the encoder-decoder framework, the role of the encoder is to identify the data of the input table, and the role of the decoder is to use the long short-term memory network to create a description based on the table data. However, this method does not perform well in content selection, and lacks coherence between sentences

The model for generating text summarization based on structured data first encodes the summarization text into a record table, and then combines the content selection and planning mechanism into the neural network model for description. However, this method not only does not encode the potential space between words The positional relationship cannot cover the positional relationship between various sentences, and does not classify and mark different types of data, resulting in illogical situations in the generated summary

[0008] Traditional natural language algorithms RNN and LSTM can only extract features sequentially from left to right or from right to left, which may lead to two problems: (1) The feature calculated at time t depends on the value of the feature at time t-1, which is extremely large. This limits the parallel capability of the model; (2) In the long-term dependence process between eigenvalues, there may be a loss of characteristic information

[0009] Once the style of the data table of the traditional English summary generation method changes, it will lead to distortion of the English summary description and lack of versatility

The generation method of English summary based on time series also has the problems of difficulty in obtaining structured data, incorrect model prediction, and illogical English summary generation.

Images