Aspect-level sentiment analysis method and equipment based on large language model

Abstract

The invention provides an aspect-level sentiment analysis method and device based on a large language model, and the method comprises the steps: firstly extracting syntactic structure information of a to-be-predicted statement, and then fusing the syntactic structure information of the statement into the large language model to achieve aspect-level sentiment classification. In the fusion process, parsing of syntactic structure information, extraction of viewpoint information and analysis of sentiment polarity are completed in sequence through a large language model, and a final aspect-level sentiment classification result is obtained. According to the method, syntactic information is effectively fused into the aspect-level sentiment classification process of the large language model by utilizing a method of multiple understanding and reasoning of the large language model, and the accuracy of the large language model in aspect-level sentiment classification can be effectively improved through multi-step understanding and reasoning. And specific prompt statements are formulated for executing different steps for the large language model, so that the performance of the large language model on aspect-level sentiment classification tasks is effectively improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of text sentiment analysis, and in particular relates to an aspect-level sentiment analysis technical solution based on a large language model. Background Art

[0002] With the popularization of the Internet and the booming development of social media, people's lifestyles and ways of obtaining information have changed significantly, and human life has become more and more dependent on the Internet. The huge user group has generated a large amount of data on various social media, news websites and other platforms, recording people's daily behavior trajectories and reflecting people's views and attitudes on various events, products and topics. Using automated sentiment analysis technology to efficiently and accurately extract sentiment information can better understand human emotional expressions and attitudes.

[0003] Sentiment Analysis refers to the process of identifying and extracting people's opinions, attitudes and emotional tendencies from text data through natural language processing technology. It aims to establish an effective analysis method and system to analyze the unique emotional information of certain objects in the input information. It can not only help companies understand consumer needs and optimize product services, but also provide strong support for government decision-making, public opinion monitoring, intelligent education and other fields. Aspect-based Sentiment Analysis is a more fine-grained sentiment analysis task. Its purpose is to identify the emotional tendencies in text for specific aspects, so it can provide users with more detailed and specific sentiment analysis results. Aspect-based sentiment analysis generally includes two elements: the target of analysis and the corresponding emotion. The target is usually an entity, and the emotion is usually divided into positive, negative, or neutral. A mainstream method of text sentiment analysis is based on deep learning technology. First, training data with sentiment polarity annotations is obtained, then a deep learning model is built and trained, and finally the model is used to predict the emotion of the test document. For example, the Chinese invention patent - Subjective text sentiment analysis method based on deep learning (CN106776581B) builds a deep learning model to calculate the sentiment tendency of comment sentences. However, in order to train deep learning models for specific tasks, this type of method requires a large amount of data to be annotated in advance, which consumes a lot of manpower and material resources. Summary of the invention

[0004] In view of the deficiencies in the prior art, the present invention provides a technical solution for aspect-level sentiment analysis based on a large language model, which integrates the syntactic information of the sentence to be analyzed into the large language model, and sequentially uses the large language model to complete the three tasks of syntactic parsing, opinion extraction, and sentiment analysis, effectively improving the ability of the large language model in aspect-level sentiment classification tasks without providing labeled data.

[0005] The present invention proposes an aspect-level sentiment analysis method based on a large language model, comprising the following processing: firstly, extracting the syntactic structure information of the sentence to be predicted, and then integrating the syntactic information of the sentence into the large language model to realize aspect-level sentiment classification; during the integration process, the large language model is used to sequentially complete the parsing of syntactic structure information, the extraction of opinion information and the analysis of sentiment polarity to obtain the final aspect-level sentiment classification result.

[0006] Moreover, the syntactic structure information is represented in the format of word line structure, including taking words as processing objects, splitting sentences into multiple word lines, and each word line contains multiple key information of the word in the sentence.

[0007] Moreover, the key information in the word line includes the position index of the word, the actual text of the word, the part of speech, the position index of the words that each word depends on, and the relationship between the words that each word depends on.

[0008] Moreover, the parsing of the syntactic structure information includes, after receiving the syntactic structure parsing task, based on the syntactic structure parsing task template P1, combining sentence X, aspect word A and syntactic structure information S to construct a syntactic parsing task prompt sentence Q1, inputting Q1 into the large language model, and the large language model generates syntactic parsing information about the aspect word A based on the syntactic structure information S and sentence X, which is recorded as reply R1.

[0009] Moreover, the extraction of the opinion information includes, after receiving the opinion information extraction task, based on the opinion information extraction task template P2, combining sentence X, aspect word A and reply R1 to construct the opinion extraction task prompt sentence Q2, inputting Q2 into the large language model, and the large language model extracting the opinion information on the aspect word A in sentence X based on the reply R1, which is recorded as reply R2.

[0010] Moreover, after receiving the sentiment polarity analysis task, the sentiment polarity analysis is performed based on the sentiment polarity analysis task template P3, combining sentence X, aspect word A and reply R2 to construct the sentiment analysis task prompt sentence Q3, and Q3 is input into the large language model. The large language model analyzes the sentiment polarity of aspect word A in sentence X based on R2, which is recorded as reply R3.

[0011] Moreover, it is used for hot public opinion analysis, product quality analysis or service experience analysis.

[0012] On the other hand, the present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the aspect-level sentiment analysis method based on the large language model as described above is implemented.

[0013] On the other hand, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the aspect-level sentiment analysis method based on a large language model as described above.

[0014] On the other hand, the present invention also provides a computer program product, including a computer program, which, when executed by a processor, implements the aspect-level sentiment analysis method based on a large language model as described above.

[0015] According to the aspect-level sentiment analysis solution based on the large language model provided by the present invention, syntactic information is integrated into the large language model, and the large language model is used to sequentially complete syntactic information analysis, opinion information extraction, and sentiment polarity analysis. Through multi-step understanding and reasoning, the accuracy of the large language model in aspect-level sentiment classification can be effectively improved. The present invention can effectively and automatically identify the sentiment polarity of a specific target in a text. This information can not only monitor social public opinion, but also can be used for product quality analysis, and has great market value and research significance. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 This is a flow chart of the aspect-level sentiment analysis method based on a large language model in an embodiment of the present invention;

[0017] Figure 2 It is a schematic diagram of the structure of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0018] The technical scheme of the present invention is described in detail below in conjunction with the accompanying drawings and embodiments. The described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0019] The present invention will be further described below in conjunction with specific embodiments, but the present invention is not limited thereto.

[0020] The present invention believes that large language models (Large Language Models) demonstrate powerful text understanding and generation capabilities, and can complete various specified natural language processing tasks without the need for any training data. Large language models are pre-trained on a large amount of unlabeled data to form a language model that conforms to the distribution of human language habits. Without providing any training data, they can generate outputs that meet the requirements by simply receiving specific task prompts, thereby completing the specified task, which can effectively solve the shortcomings of previous deep learning models that require a large amount of labeled data training. Therefore, the present invention proposes an aspect-level sentiment analysis solution based on a large language model.

[0021] A large language model generally refers to a deep learning model that is pre-trained using a large amount of text data, which can understand natural language text and generate corresponding natural language responses. In specific implementation, the large language model can refer to the prior art: Ouyang, Long, et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems 35 (2022): 27730-27744. This invention will not be repeated.

[0022] Aspect-level sentiment analysis requires accurate understanding of the sentiment towards a specific target in a sentence. Due to its sophisticated task requirements, aspect-level sentiment analysis methods based on large language models place higher demands on the accurate text sentiment understanding capabilities of large language models. Considering that syntactic information can provide relationships and connections between words in a sentence, previous aspect-level sentiment analysis methods would use graph structures to model the dependencies between syntactic information, thereby capturing opinion information related to the target word, thereby enhancing the sentiment understanding of the target word. However, large language models usually process serialized text information. How to utilize structured syntactic dependency information and thereby enhance the aspect-level sentiment analysis capabilities of large language models is the key technical problem to be solved by the present invention.

[0023] like Figure 1 As shown, an embodiment of the present invention discloses an aspect-level sentiment analysis method based on a large language model, which first obtains the syntactic structure information of the sentence to be predicted, and then integrates the syntactic information of the sentence into the large language model to realize aspect-level sentiment classification, including using the large language model to sequentially complete the parsing of syntactic structure information, the extraction of opinion information and the analysis of sentiment polarity to realize the final aspect-level sentiment classification. Furthermore, the embodiment of the present invention effectively integrates syntactic information into the aspect-level sentiment classification process of the large language model by using the method of multiple understanding and reasoning of the large language model, formulates unique prompt sentences for the large language model to execute different steps, and effectively improves the performance of the large language model in the aspect-level sentiment classification task. The specific implementation process of the aspect-level sentiment analysis method based on the large language model provided in the embodiment includes the following steps:

[0024] Step 1, obtaining the syntactic structure information S of the sentence to be analyzed X;

[0025] The present invention further proposes that the syntactic structure information S is represented by a word line structure format. The word line structure format takes words as processing objects and splits sentences into multiple word lines. Each word line contains multiple key information of the word in the sentence.

[0026] For example, the sentence X to be analyzed is "the food is fresh!", the aspect word A is "food", and the syntactic structure information S of the sentence X="the food is fresh!" is obtained.

[0027] In the embodiment, this step involves using syntactic parsing technology, taking the words in the sentence as processing units, dividing the sentence into word line structures according to the words, and each word line contains a variety of key information of the word in the sentence, including: the position index (ID) of the word, the actual text (Text) of the word, the part of speech (POS), the position index (Head) of the words on which each word depends, and the relationship (DelRel) between the words on which each word depends. An example is as follows:

[0028]

[0029] Where ID is the identifier of each word, which is used to indicate the position of the word in the sentence. In the example, ID is an integer from 1 to 4. POS is the part-of-speech tag of each word, which is used to indicate whether each word is a noun, verb, adjective, etc. NOUN, AUX, ADJ, and PUNCT in the example represent nouns, auxiliary verbs, adjectives, and punctuation marks, respectively. Head represents the ID of the dependent word. In the syntactic structure, each word depends directly or indirectly on a word. DepRel is the relationship label between each word and its dependent words. These relationship labels provide detailed information about the sentence structure. In the example, the relationship between "food" and "is" is nsubj (noun subject), "is" as root (root node) indicates that it is the core of the sentence, and the relationship between "fresh" and "is" is acomp (adjective complement), and "!" is punct (punctuation mark) indicates that it is the punctuation mark of the sentence.

[0030] Step 2: After receiving the syntactic structure parsing task, the large language model constructs the syntactic parsing task prompt sentence Q1 based on the syntactic structure parsing task template P1, combined with the sentence X, the aspect word A and the syntactic structure information S, and inputs Q1 into the large language model. The large language model generates syntactic parsing information about the aspect word A based on the syntactic structure information S and the sentence X, which is recorded as the response R1.

[0031] The present invention proposes a method for constructing the prompt sentence Q1: firstly, the requirements and description of the syntactic structure parsing task are given in the prompt sentence, and then X, A and S are spliced ​​into the prompt sentence to obtain the prompt sentence Q1.

[0032] In the embodiment, a syntactic structure analysis task template P1 is further set to construct a syntactic analysis task prompt sentence Q1. The P1 template consists of three parts: task description, syntactic structure information and the sentence to be analyzed. The specific contents are as follows:

[0033] Task description: Based on the syntactic dependency information of the sentence, analyze the information related to {aspect word A} in the sentence.

[0034] Syntactic information: {Syntactic information S}.

[0035] Sentence to be analyzed: {sentence X to be analyzed}.

[0036] The content in {·} is what needs to be filled in according to the specific case. Input sentence X, aspect word A and syntactic structure information S into template P1 to construct syntactic parsing task prompt sentence Q1. Then input Q1 into the large language model. The large language model generates syntactic parsing information with aspect word A based on syntactic structure information S and sentence X, which is recorded as response R1.

[0037] In the embodiment, sentence X, aspect word A and syntactic structure information S are input into template P1, and a syntactic parsing task input prompt sentence Q1 is constructed. The goal of the task is to analyze the relevant information of the aspect word "food" based on the syntactic structure information. Q1 is input into the large language model, and the large language model completes the syntactic parsing task and generates a response R1: "food" (ID 1) is the subject of the verb "is" (ID 2), which means it is the subject of the sentence, and the adjective "fresh" describes its state.

[0038] Step 3: After receiving the opinion information extraction task, the large language model constructs the opinion extraction task prompt sentence Q2 based on the opinion information extraction task template P2, combined with sentence X, aspect word A and reply R1, and inputs Q2 into the large language model. The large language model extracts the opinion information of aspect word A in sentence X based on R1, which is recorded as reply R2.

[0039] The present invention proposes a method for constructing the prompt statement Q2: first, the requirements and description of the viewpoint information extraction task are given in the prompt statement, and then X, A and R1 are spliced ​​into the prompt statement to obtain the prompt statement Q2.

[0040] In the embodiment, a viewpoint information extraction task template P2 is further set to construct a viewpoint extraction task prompt sentence Q2. The P2 template consists of three parts: task description, the reply to the syntactic parsing task in step 2, and the sentence to be analyzed. The specific contents are as follows:

[0041] Task Description: Considering the contextual information related to {aspect word A}, what is the user's opinion on {aspect word A}?

[0042] Syntactic parsing task reply: {Reply R1}.

[0043] Sentence to be analyzed: {sentence to be analyzed X}

[0044] The content in {·} is what needs to be filled in according to the specific case. Sentence X, aspect word A and reply R1 are input into template P2 to construct the prompt sentence Q2 for the opinion extraction task. Then Q2 is input into the large language model. The large language model extracts the opinion information about aspect word A in sentence X based on R1, which is recorded as reply R2.

[0045] In the embodiment, sentence X, aspect word A and reply R1 are input into template P2 to construct the input prompt sentence Q2 of the opinion extraction task. The goal of the task is to analyze the user's opinion on the aspect word "food" based on context information. Q2 is input into the large language model, which completes the opinion extraction task and generates reply R2: The adjective "fresh" is usually associated with a positive dining experience.

[0046] Step 4: After receiving the sentiment polarity analysis task, the large language model constructs the sentiment analysis task prompt sentence Q3 based on the sentiment polarity analysis task template P3, combined with sentence X, aspect word A and reply R2, and inputs Q3 into the large language model. The large language model analyzes the sentiment polarity of aspect word A in sentence X based on R2, which is recorded as reply R3.

[0047] The present invention proposes a method for constructing the prompt sentence Q3: first, the requirements and description of the sentiment analysis task are given in the prompt sentence, and then X, A and R2 are spliced ​​into the prompt sentence to obtain the prompt sentence Q3.

[0048] In the embodiment, a sentiment polarity analysis task template P3 is further set to construct a sentiment analysis task prompt sentence Q3. The P3 template consists of three parts: task description, the reply to the opinion information extraction task in step 3, and the sentence to be analyzed. The specific contents are as follows:

[0049] Task Description: Based on common sense and user opinions, what is the sentiment polarity of {aspect word A}?

[0050] Opinion extraction task response: {Reply R2}.

[0051] Sentence to be analyzed: {sentence to be analyzed X}

[0052] The content in {·} is what needs to be filled in according to the specific case. Sentence X, aspect word A and reply R2 are input into template P3 to construct sentiment classification task prompt sentence Q3. Then Q3 is input into the large language model, which analyzes the sentiment polarity of aspect word A in sentence X based on R2 and records it as reply R3.

[0053] In the embodiment, sentence X, aspect word A and reply R2 are input into template P3 to construct prompt sentence Q3 of sentiment analysis task. The goal of the task is to analyze the sentiment polarity of aspect word "food" based on common sense knowledge and opinion information. Q3 is input into the large language model, and the large language model completes the sentiment classification task and generates reply R3: In the sentence to be analyzed, the sentiment polarity of aspect word "food" is positive.

[0054] Step 5, perform sentiment polarity keyword matching on the reply R3, and finally parse and obtain the sentiment polarity category of the aspect word "food" as positive.

[0055] This embodiment uses syntactic structure information to add to the prompt template and input it into the understanding process of the large language model, and utilizes the multi-step understanding and reasoning capabilities of the large language model to enable the large language model to better utilize syntactic structure information and improve the performance of the large language model in aspect-level sentiment classification tasks.

[0056] In some embodiments, the above process is used for hot public opinion analysis. By analyzing the public's emotional tendencies in the occurrence, process, and results of events, analysts can understand the public's emotional distribution in these specific aspects and provide decision makers with more detailed and accurate public opinion intelligence. For example, aspect-level sentiment analysis of texts in channels such as social media and forums can obtain the public's emotions on hot events, which can be used for hot public opinion analysis.

[0057] In some embodiments, the above process is used for product quality analysis and service experience analysis. By analyzing user comments on product appearance, performance, price, durability, etc., consumers' satisfaction and emotional tendencies in these aspects can be understood. Enterprises can improve and optimize products based on these feedbacks to improve product quality and user experience. For example, aspect-level sentiment analysis of feedback information such as user comments and product evaluations can identify user satisfaction with products or services, which can be used for product quality improvement and analysis.

[0058] In order to verify the effect of the method of the present invention, comparative experiments were conducted on two standard datasets of aspect-level sentiment analysis, Rest14 and Lap14. The comparative experiments were divided into two groups, both of which used large language models to solve aspect-level sentiment analysis. One group used the multi-step reasoning method proposed by the present invention that incorporates syntactic information, and the other group used a method that did not provide syntactic information and used a large language model to directly analyze sentiment polarity.

[0059] This experiment uses two large language models, GPT-3.5 and GPT-4o. GPT-3.5 and GPT-4o are large language models developed by OpenAI and represent the most advanced large language model technology. The Rest14 and Lap14 datasets are user review data from the catering and notebook shopping fields, respectively. The specific experimental results are as follows:

[0060] Table 1 Performance comparison of two methods of using the large language model on the aspect-level sentiment analysis dataset

[0061]

[0062] It can be seen from the results in Table 1 that the method of the present invention effectively improves the accuracy of the large language model in aspect-level sentiment analysis by incorporating syntactic information, indicating that the technical solution of the present invention is effective.

[0063] In specific implementation, those skilled in the art can use software technology to realize the automatic operation of the above process. Accordingly, if an aspect-level sentiment analysis solution based on a large language model is provided, including a computer or a server, and the above process is executed on the computer or server to fuse syntactic information into the large language model to realize aspect-level sentiment analysis, it should also be within the scope of protection of the present invention.

[0064] During specific implementation, the above process can be automatically executed using computer software technology.

[0065] In addition, the present application also relates to an electronic device, comprising at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the above-mentioned aspect-level sentiment analysis method based on a large language model.

[0066] The present application also relates to a non-transitory computer-readable storage medium storing computer instructions, wherein the computer instructions are used to enable the computer to execute the above-mentioned aspect-level sentiment analysis method based on a large language model.

[0067] The present application also relates to a computer program product, including a computer program, characterized in that when the computer program is executed by a processor, the above-mentioned aspect-level sentiment analysis method based on a large language model is implemented.

[0068] Figure 2 An example of a physical structure diagram of an electronic device is shown in FIG. Figure 2As shown, the electronic device may include: a processor, a communications interface, a memory, and a communications bus, wherein the processor, the communications interface, and the memory communicate with each other via the communications bus. The processor may call logic instructions in the memory to execute the method for reconstructing the timing of effective angular momentum of earth fluids.

[0069] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0070] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0071] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for aspect-level sentiment analysis based on a large language model, comprising the following processing: First, the syntactic structure information of the sentence to be predicted is extracted, and then the syntactic information of the sentence is integrated into the large language model to realize aspect-level sentiment classification. During the integration process, the large language model is used to successively complete the parsing of syntactic structure information, the extraction of opinion information and the analysis of sentiment polarity to obtain the final aspect-level sentiment classification result.

2. The aspect-level sentiment analysis method based on a large language model according to claim 1, characterized in that: The syntactic structure information is represented in the format of a word line structure, including taking a word as a processing object, splitting a sentence into multiple word lines, and each word line contains multiple key information of the word in the sentence.

3. The aspect-level sentiment analysis method based on a large language model according to claim 2 is characterized in that: The key information in a word line includes the word's position index, the word's actual text, the word's part of speech, the position index of each word's dependent words, and the relationship between each word's dependent words.

4. The aspect-level sentiment analysis method based on a large language model according to claim 1, characterized in that: The parsing of the syntactic structure information includes, after receiving the syntactic structure parsing task, constructing a syntactic parsing task prompt sentence Q1 based on the syntactic structure parsing task template P1, combining sentence X, aspect word A and syntactic structure information S, and inputting Q1 into the large language model. The large language model generates syntactic parsing information about the aspect word A based on the syntactic structure information S and sentence X, which is recorded as reply R1.

5. The aspect-level sentiment analysis method based on a large language model according to claim 4 is characterized in that: The extraction of the opinion information includes, after receiving the opinion information extraction task, constructing the opinion extraction task prompt sentence Q2 based on the opinion information extraction task template P2, combining sentence X, aspect word A and reply R1, inputting Q2 into the large language model, and the large language model extracting the opinion information on the aspect word A in sentence X based on the reply R1, which is recorded as reply R2.

6. The aspect-level sentiment analysis method based on a large language model according to claim 5 is characterized by: The analysis of the sentiment polarity, after receiving the sentiment polarity analysis task, is based on the sentiment polarity analysis task template P3, combined with sentence X, aspect word A and reply R2 to construct the sentiment analysis task prompt sentence Q3, and Q3 is input into the large language model. The large language model analyzes the sentiment polarity of the aspect word A in sentence X based on R2, which is recorded as reply R3.

7. The aspect-level sentiment analysis method based on a large language model according to any one of claims 1 or 2 or 3 or 4 or 5 or 6, characterized in that: Used for hot public opinion analysis, product quality analysis or service experience analysis.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the aspect-level sentiment analysis method based on the large language model is implemented as described in any one of claims 1 to 7.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the aspect-level sentiment analysis method based on a large language model is implemented as described in any one of claims 1 to 7.

10. A computer program product, comprising a computer program, characterized in that: When the computer program is executed by a processor, the aspect-level sentiment analysis method based on a large language model is implemented as described in any one of claims 1 to 7.

Images