Natural language processing method

Abstract

The invention discloses a natural language processing method which comprises the steps of inputting, text data preprocessing, NLP task, answer generation, answer decompilation, comparison of decompilation generation and initial output, question correction and outputting, and inputting a natural language into the method through the input. Performing text data preprocessing on the input natural language information through a natural language processing system; the NPL task performs recognition and matching generation on the processed text data; answer generation: sorting and fusing to generate answers; the decompilation answer forms a reverse question for the generated answer, and a natural language of the decompilation answer is generated through processing of the steps; comparing decompilation generation with initial input, and comparing a natural language of a formed decompilation answer with an input natural language; in question correction, an answer which is not generated through decompilation and is compared with the initial input is corrected, and natural language processing is performed again; and outputting the generated answer which is compared with the initial input and generated through decompilation.

Description

Technical Field

[0001] The present invention relates to the technical field of natural language processing, and in particular to a natural language processing method. Background Art

[0002] In the fields of computer science and artificial intelligence, the study of natural language processing is a very important direction, mainly to study various methods and theories of communication between people and computer systems through natural language. Natural language processing requires the integration of linguistics, computer science and mathematics, and natural language processing is to achieve smooth communication between people and computer systems through natural language and enrich human life, rather than general research on natural language. In terms of human-computer interaction, natural language processing has broad development prospects. Therefore, research in this field will involve natural language, that is, the language used by people in daily life, so it is closely related to the study of linguistics, but there are important differences. Natural language processing is not a general study of natural language, but to develop computer systems that can effectively realize natural language communication. Natural language processing is mainly used in machine translation, public opinion monitoring, automatic summarization, opinion extraction, text classification, question answering, text semantic comparison, speech recognition, Chinese OCR and other aspects;

[0003] However, the existing natural language processing has low accuracy during human-computer interaction, resulting in a poor human-computer interaction experience and is insufficient to meet the smooth communication requirements of human-computer interaction. To address the above problems, a natural language processing method is proposed. Summary of the invention

[0004] The purpose of the present invention is to provide a natural language processing method to solve the problems raised in the above background technology.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] A natural language processing method, comprising: input, text data preprocessing, NLP task, answer generation, decompiled answer, decompiled answer comparison with initial output, question correction and output,

[0007] Feed it with natural language via input;

[0008] Performing text data preprocessing on the input natural language information through a natural language processing system;

[0009] The NPL task identifies, matches and generates processed text data;

[0010] Answer generation is to organize and integrate answers;

[0011] The decompiled answer forms a counter-question to the generated answer, and is processed through the above steps to generate the natural language of the decompiled answer;

[0012] Decompilation generation and initial input comparison: the natural language of the decompiled answer is compared with the natural language of the input;

[0013] Question correction: Correct the generated answers that were not compared with the initial input through decompilation and re-perform natural language processing;

[0014] The output is the answer generated by decompilation and compared with the original input.

[0015] As a further solution of the present invention: the input determines whether the collected natural language is voice input, the voice input collects the input voiceprint, performs noise reduction processing on the collected voice to obtain clear natural language information, and performs emotion recognition through the intonation and tone of the input sound.

[0016] As a further solution of the present invention: the text data preprocessing performs word segmentation to obtain clear natural language information, performs text error correction after word segmentation, and recognizes the named entities after word segmentation through named entity recognition, vectorizes the natural language after word segmentation through feature representation, obtains high-order features through part-of-speech tagging, and performs reference resolution to convert the text into a processable form.

[0017] As a further solution of the present invention: the NLP task performs intent recognition on the input natural language, performs question-answer matching, and generates natural language.

[0018] As a further solution of the present invention: the answer generation combines the natural language generated by question and answer matching with the emotional information obtained by emotion recognition under multiple information fusion, sorts the answers according to the template through intent annotation and slot relationship under the separation of dependencies, optimizes the user experience through familiarity calculation, and forms the answer to be output by answer representation, answer retrieval and automatic annotation.

[0019] As a further solution of the present invention: the output answer is fed back to the natural language processing system to achieve feedback learning.

[0020] Compared with the prior art, the present invention has the following beneficial effects:

[0021] 1. The present invention decompiles the generated answer by decompiling the answer, and processes the decompiled text through text data preprocessing, NLP tasks and answer generation, to achieve decompilation and answer generation, and compares it with the initial input. If it meets the requirements, the generated answer is output, and if it does not meet the requirements, it is adjusted through question correction combined with training corpus and question knowledge base, so as to improve the accuracy of natural language processing results, improve the experience of human-computer interaction, and increase the fluency of human-computer interaction.

[0022] 2. The answers output by the present invention are fed back to the model in the natural language processing system (through traditional machine learning (naive Bayes, logistic regression, random forest), word vector models (such as word2vec), deep learning neural networks (RNN, CNN, LSTM, BERT, etc.) and knowledge mining combined with model fusion, model distillation, incremental learning, transfer learning, sentiment analysis, user experience design, knowledge base operation tools, model iteration mechanism and other engineering methods), and through big data training, it also includes a training corpus and a question knowledge base. The present invention feeds back the output answers to the natural language processing system, and through feedback learning, it increases the accuracy of natural language processing under long-term use. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 A block diagram of a natural language processing method.

[0024] Figure 2 A flowchart of a natural language processing method. DETAILED DESCRIPTION

[0025] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0026] See also Figures 1-2 In an embodiment of the present invention, a natural language processing method includes: input, text data preprocessing, NLP task, answer generation, decompiled answer, decompiled answer and initial output comparison, question correction and output,

[0027] Feed it with natural language via input;

[0028] The input determines whether the collected natural language is voice input. The voice input collects the voiceprint of the waker, recognizes the input voice through the voiceprint, and obtains clear natural language information by performing noise reduction processing on the collected voice. At the same time, emotion recognition is performed through the intonation and tone of the input sound to facilitate the accurate production of answers.

[0029] Performing text data preprocessing on the input natural language information through a natural language processing system;

[0030] Text data preprocessing is carried out to obtain clear natural language through word segmentation based on dictionary matching, word segmentation algorithm based on semantic analysis and probability statistical model. After word segmentation, the text information after word segmentation is corrected by text error correction algorithm to reduce the interference of erroneous words, and named entity recognition is used to identify the named entities after word segmentation (identification of names of people, places, institutions, proper nouns, etc.). Feature representation uses TF-IDF (term frequency-inverse document frequency), LSI (latent semantic indexing), word embedding and other methods to extract key information from the natural language after word segmentation and convert it into vector form. High-order features are obtained through part-of-speech tagging, and reference resolution is performed to convert the text into a processable form.

[0031] The NPL task identifies, matches and generates processed text data;

[0032] The NLP task uses the user's intention of the input natural language of the human-computer interaction system, and learns the deep semantic features in the sentence in a combined recurrent neural network (RNN) and convolutional neural network (CNN) model, and calculates the similarity distance of the feature vector. On the basis of this model, two different attention mechanisms are added to construct the feature representation of the answer according to the question to learn the detailed semantic matching relationship in the question and answer pairs, perform question and answer matching, and generate natural language.

[0033] Answer generation is to organize and integrate answers;

[0034] Answer generation combines the natural language generated by question-answer matching with the emotional information obtained by emotion recognition, combines multiple types of information from different sources, and sorts the answers according to preset templates through the intent annotation and slot relationship of text data under the dependency analysis of the syntactic relationship between natural language words, optimizes the user experience through familiarity calculation, retrieves the output answers, and automatically annotates the text data through computer algorithms to identify and classify the meaning and usage of words or phrases in the text to form answers to be output that can be expressed in multiple ways.

[0035] The decompiled answer forms a counter-question to the generated answer, and is processed through the above steps to generate the natural language of the decompiled answer;

[0036] By decompiling the answer, the generated answer is preset as a rhetorical question, such as "Free installation for orders over 1,500 yuan, only in urban areas" after decompilation becomes "Why free installation for orders over 1,500 yuan".

[0037] Decompilation generation and initial input comparison: the natural language of the decompiled answer is compared with the natural language of the input;

[0038] The text information "Why is installation free for purchases over 1,500 yuan?" is processed through text data preprocessing, NLP tasks, and answer generation to generate "door purchase installation conditions". The original input natural language information is "Does installation come with purchases?"

[0039] Question correction: Correct the generated answers that were not compared with the initial input through decompilation and re-perform natural language processing;

[0040] If the input is "door installation conditions", the generated answer is "the door opening needs to be cleaned and leveled", the decompiled answer is "why the door opening needs to be cleaned and leveled", the decompiled answer is "what needs to be done when installing the door", and the question is corrected by combining the input context, the training corpus and the question knowledge base to obtain "door installation conditions", that is, "buy door installation conditions".

[0041] The output is the answer generated by decompilation and compared with the original input.

[0042] The output answer is fed back to the model in the natural language processing system (through traditional machine learning (naive Bayes, logistic regression, random forest), word vector models (such as word2vec), deep learning neural networks (RNN, CNN, LSTM, BERT, etc.) and knowledge mining combined with model fusion, model distillation, incremental learning, transfer learning, sentiment analysis, user experience design, knowledge base operation tools, model iteration mechanism and other engineering methods), and through big data training, it also includes a training corpus and a question knowledge base to achieve feedback learning, and increase the accuracy of natural language processing under long-term use.

[0043] Although the present invention has been described in detail with reference to the aforementioned embodiments, it is still possible for those skilled in the art to modify the technical solutions described in the aforementioned embodiments, or to make equivalent substitutions for some of the technical features therein. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A natural language processing method, comprising: Input, text data preprocessing, NLP task, answer generation, decompiled answer, decompiled generation and initial output comparison, question correction and output, characterized by: Feed it with natural language via input; Performing text data preprocessing on the input natural language information through a natural language processing system; The NPL task identifies, matches and generates processed text data; Answer generation is to organize and integrate answers; The decompiled answer forms a counter-question to the generated answer, and is processed through the above steps to generate the natural language of the decompiled answer; Decompilation generation and initial input comparison: the natural language of the decompiled answer is compared with the natural language of the input; Question correction: Correct the generated answers that were not compared with the initial input through decompilation and re-perform natural language processing; The output is the answer generated by decompilation and compared with the original input.

2. A natural language processing method according to claim 1, characterized in that: The input determines whether the collected natural language is voice input, the voice input collects the input voiceprint, performs noise reduction processing on the collected voice to obtain clear natural language information, and performs emotion recognition through the intonation and tone of the input sound.

3. A natural language processing method according to claim 1, characterized in that: The text data preprocessing performs word segmentation on the acquired clear natural language information, performs text error correction after word segmentation, and recognizes the named entities after word segmentation through named entity recognition. The feature representation vectorizes the natural language after word segmentation, obtains high-order features through part-of-speech tagging, and performs reference resolution to convert the text into a processable form.

4. A natural language processing method according to claim 1, characterized in that: The NLP task performs intent recognition on the input natural language, performs question-answer matching, and generates natural language.

5. A natural language processing method according to claim 1, characterized in that: The answer generation combines the natural language generated by question-answer matching with the emotional information obtained by emotion recognition under multiple information fusion, sorts the answers according to the template through intent annotation and slot relationship under the separation of dependencies, optimizes the user experience through familiarity calculation, and forms the answer to be output through answer representation, answer retrieval and automatic annotation.

6. A natural language processing method according to claim 1, characterized in that: The output answer is fed back to the natural language processing system to achieve feedback learning.

Images