Research on the Generation of Patented Technology Points in New Energy Based on Deep Learning

Introduction

New energy refers to various forms of energy other than traditional energy that is renewable energy developed and utilized on the basis of new technologies. In recent years, new energy technology has achieved great progress and all countries take to develop new energy as an important measure for industrial adjustment. Researches show that the number of new energy patent has increased rapidly, the annual patent applications up to 100,000 (Zhao. 2019). Although the total number of new energy patents has reached one million, the uneven quality of patents isn’t conducive to technological innovation and the transformation and application of achievements. Therefore, it is of great practical significance to study the effective acquisition of patent technology points. The field of new energy is now an emerging industry. The technical point of the patent represents the core technical value in the patent. Therefore, the research on the generation of patent technology points has great practical significance for the transformation of patent achievements and the promotion of technological innovation in the field of new energy.

Patent technology points can cover concisely the core technology of the patent. Compared with traditional patent abstract information, it can present patent technical points more clearly and brightly. Generally, text generation technology can effectively obtain patented technology points, including extraction method and generation method. The extractive methods can ensure the grammatical structure of the text and not generate new words. But, the coherence of semantics in the generated text cannot be guaranteed. The generative methods can effectively maintain semantic coherence and readability. However, it can't guarantee the grammatical structure of the generated text. Therefore, to ensure the semantic coherence and the correctness of the grammatical structure in the text concurrently, it is practicable to combine the advantages of the two methods to complete the text generation task.

The patent data of new energy owns the characteristics of different information lengths, too long terms and words, more new words, and different text structure and syntax. Although the existing methods have well performance in the text generation task, it has performed poorly in the task of generating new energy patent technology points. To address the above problems, we propose a two-stage patent technology points generation method named IGPTP. It fuses the edge of extraction method and generation method. Meanwhile, it combines external knowledge and copy mechanism to guide the generation of patent technology points and improve the quality. The contributions made in this study are summarized as follows:

• 1.

  Authors build a corpus of patent technology points (about 5200) and a corpus of domain terms (about 3200) in the field of new energy, which bridges the gap in information on patented technologies in the field of new energy.

• 2.

  Authors propose a two-stage patented technology point generation method named IGPTP. It combines the advantages of extractive and generative method to generate the patent technology points.

• 3.

  IGPTP uses a multi-strategies ensemble method to improve the quality of patent technology points. In model training, it introduces adversarial training to promote the robustness of data. Simultaneously, the external knowledge is used to guide word segmentation and vectorization of input information. Finally, in the model decoding stage, the copy mechanism is introduced to embellish the text. The accuracy and readability of patent technology points have been promoted.

• 4.

  Substantial experiments are conducted to verify the proposed IGPTP can generate more accurate and coherent text that owns the performance edge over classical text generation methods.

The remaining parts of this paper are organized as follows. Section 2 briefly reviews some related work on text generation tasks. Section 3 elaborates on the design and implementation details of IGPTP. The experimental evaluation in which IGPTP is compared against the other classic text generation methods can be found in Section 4. Section 4 also discusses the wherefore about a few experimental results. Finally, we conclude this paper with future work in Section 5.