A multi-label text interpretation method based on thought chain prompt engineering

Abstract

This application belongs to the field of natural language processing and discloses a multi-label text interpretation method based on thought chain prompt engineering, including: dividing the documents in the multi-label text according to sentence-level segments, constructing a set of sentences corresponding to each label, and obtaining the vector representation of each label by averaging the representation of the pre-trained model of the sentences in each set; visualizing the obtained topic vectors based on the cosine similarity between the vector representation of each word in the sentence and the topic vector representation; and mining topics that are consistent with the label semantics to provide fine-grained topic interpretation. The present invention takes advantage of the advantages of pre-trained language models and large language models to provide multi-granular multi-label text interpretation.

Description

Technical Field

[0001] This application belongs to the field of natural language processing, and specifically relates to a multi-label text interpretation method based on thought chain prompt engineering. Background Art

[0002] Topic modeling is an effective text mining method in the field of natural language processing, and is often used in an unsupervised manner to mine latent semantic patterns in text corpora. Although topic models have been widely explored, from traditional probabilistic topic models to neural topic models and existing methods based on pre-training and large language models, unsupervised topic modeling methods are unable to effectively incorporate document tags into the topic modeling process, resulting in their inability to effectively adapt to scenarios with multi-label text corpora. Existing topic modeling methods for multi-label text only use word co-occurrence information and are unable to incorporate external semantic knowledge such as pre-trained language models and large language models. At the same time, they incorporate irrelevant fragments in the text into topic inference, resulting in insufficient text interpretation capabilities.

[0003] Multi-label text interpretation aims to solve the credit attribution problem in multi-label text, that is, different fragments in the text correspond to different labels, different words in the text also correspond to different labels, and the text also contains some meaningless fragments. Multi-label text interpretation needs to provide three levels of interpretation: (1) Provide sentence-level interpretation by mapping the fragments in the text to their semantically related labels. (2) For sentences that have been assigned labels, map the words they contain to semantically related labels. (3) For each label in the text corpus, mine the topics that are semantically consistent with it and provide explanations for the label.

[0004] However, existing multi-label topic models rely solely on word co-occurrence information and are unable to incorporate external semantic knowledge into their modeling methods. They are unable to accurately provide effective explanations and are unable to identify meaningless fragments in the text, resulting in insufficient explanation capabilities. Summary of the Invention

[0005] In order to address the defects of the existing technology, this application provides a multi-label text interpretation method based on thought chain prompt engineering. This application provides an effective interpretation for multi-label text corpus by integrating pre-trained language models and large language models.

[0006] In order to achieve the above objectives, this application is implemented through the following technical solutions:

[0007] This application is a multi-label text interpretation method based on thought chain prompt engineering, which includes the following steps:

[0008] Step 1: Obtain the corpus on the multi-label text platform, clean the corpus, and obtain a multi-label text corpus collection;

[0009] Step 2: For each label in the multi-label text corpus, use the large language model to generate label description information, and use the thinking chain prompting project based on the large language model to align the text fragments in each document in the multi-label text corpus to the labels that are semantically related to the text fragments, that is, assign a label to each text fragment, and use a text fragment that is not related to any label to align the text fragments in each document in the multi-label text corpus to align the text fragments in each document to the labels that are semantically related to the text fragments. Labels are aligned and the topic vector representation matrix is obtained based on the pre-trained language model ;

[0010] Step 3: Based on the labels assigned to the text fragment by the large language model, the words in the sentence are visualized by calculating the correlation between the word vector of each word in the text fragment and the corresponding vector of the label assigned to the text fragment;

[0011] Step 4: Obtain the topic vector representation matrix based on the label corresponding to the text segment fed back by the large language model and step 2 , extracting topics that are semantically consistent with the labels in a multi-label text corpus and providing explanations for the labels. Unlike existing multi-label topic modeling methods, this application's method utilizes a large language model for sentence-level topic alignment, while preventing irrelevant text fragments from participating in topic inference. This ensures the accuracy of topic vector representation and facilitates subsequent word visualization and topic mining.

[0012] A further improvement of the present application is that: the step 1 obtains the corpus on the multi-label text platform, and cleans the corpus specifically by filtering the text with less than three text fragments in the corpus, and deleting irrelevant content in the corpus text such as emoticons and website links.

[0013] A further improvement of the present application is that the step 2 specifically includes the following steps:

[0014] Step 2.1: For each label in the multi-label text corpus, use the large language model to generate a label description, use the label description to supplement the label's semantics, and check the generated label description to avoid the introduction of erroneous information that affects model performance;

[0015] Step 2.2, each document The corresponding tag set is:

[0016] ,

[0017] in, Represents a document The number of sentences in Indicates the number of tags corresponding to the document, Represents a document The corresponding tag set, Represents a document Corresponding to the last tag in the tag set, Represents a document The number of tags in the corresponding tag set.

[0018] For the label corresponding to each document in the multi-label text corpus, first The tag is added to the multi-tag corpus corresponding to the document to obtain the document Corresponding enhanced tag set ;

[0019] Step 2.3: Documentation Each text fragment in , first let the large language model summarize each text segment Extract key information from the keywords, combine the context of the keywords, purify the extracted keywords, and combine the tag description information generated in step 2.1 to create a tag for each text segment. Assign at least one semantically relevant label;

[0020] Step 2.4: For all texts in the corpus, construct labels A collection of sentences , use the pre-trained language model to obtain a sentence set No. text snippets The word vector of each word in , average the word vector representation of each word to obtain the text fragment Vector representation of , the above transformation is expressed as follows:

[0021]

[0022]

[0023] in, is a tag collection; Represents a text fragment The number of words in Represents a text fragment The The context word vector of each word, For a set of sentences The number of sentences in Represents a text fragment The last word in ;

[0024] Based on the above transformation, we get the theme Vector representation of :

[0025]

[0026] in, Represents a pre-trained language model, and obtains a topic vector representation matrix based on the pre-trained language model ;

[0027]

[0028] in, represents the hidden dimension of the pre-trained language model, Indicates the number of labels in the label set corresponding to the multi-label corpus.

[0029] A further improvement of the present application is that step 3 specifically includes the following steps:

[0030] Step 3.1, context word vector representation acquisition: given a document The text snippets , text snippet The word sequence contained is , and text snippets The set of assigned tags , through pre-trained language model Get the context representation of each word, i.e. the word vector:

[0031]

[0032] in, Representing word sequences The words, , Represents a text fragment The number of words contained in for The word vector of Represents the assignment of a large language model to a text fragment The number of labels, express The words in

[0033] Step 3.2: Calculate the label-word correlation matrix: Based on the word vector calculated in step 3.1 , word sequence With label collection Middle Tags Semantic relevance Calculated by the following formula:

[0034]

[0035] in, Representation and Labeling The corresponding topic vector representation; For The topic vector representation of the corresponding label.

[0036] Step 3.2: Based on word vector Vector representation of labels Constructing a label-word correlation matrix Capture the semantic correlation between words and labels and visualize the words in the sentence. The larger the correlation weight, the more it indicates that the current word can indicate the theme of the sentence, making the word-level interpretation meaningful.

[0037] A further improvement of the present application is that step 4 specifically includes the following steps:

[0038] Step 4.1: For the first word in the vocabulary containing different words in the multi-label text corpus words , , Indicates the number of different words in the multi-label corpus, word Every occurrence in the multi-label corpus , then for the word No. times, expressed as , , obtain words through pre-trained language model Contextual representation of :

[0039]

[0040] in, Indicates the The context of the occurrence of express The number of times it appears in the multi-label corpus;

[0041] word Local word-topic distribution of The calculation method is:

[0042]

[0043] in, represents the cosine similarity, for Normalization operation, Indicates the The topic vector corresponding to the label is represented by ,word Corresponding corpus-level word-topic distribution It is obtained by the following formula:

[0044]

[0045] Calculate the corpus-level word-topic distribution of each word in the vocabulary, and then get the global word-topic distribution matrix by taking the average ;

[0046] Step 4.2: Based on the obtained global word-topic distribution matrix , through column normalization operation, we get the final topic-word distribution matrix , specifically:

[0047]

[0048] in, Represents a column normalization operation. For each label's corresponding topic-word distribution matrix, obtain the top 10 words with the highest probability in the topic-word distribution matrix to represent the label, providing a fine-grained explanation of the label.

[0049] The beneficial effect of this application is that it provides an effective interpretation for multi-label text corpus by integrating pre-trained language models and large language models.

[0050] This application uses a thought chain prompt project based on a large language model to provide effective sentence-level interpretation for multi-label text, which helps to quickly locate text fragments related to the label.

[0051] This application uses a label-related word visualization mechanism based on a pre-trained language model to effectively extract the correlation between words and labels in text fragments and provide fine-grained word explanations.

[0052] This application incorporates the knowledge of pre-trained language models into the topic extraction process to improve the quality and interpretability of the extracted topics. At the same time, the topic alignment module ensures the semantic consistency between labels and topics and provides corpus-level label explanations. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 This is a flowchart of the multi-label text interpretation method of this application.

[0054] Figure 2 It is a schematic diagram of the alignment and representation of the main body of this application.

[0055] Figure 3It is a diagram that visualizes the words related to basic application tags.

[0056] Figure 4 This is a schematic diagram of extracting topics related to the application tags. DETAILED DESCRIPTION

[0057] The following diagrams illustrate embodiments of the present invention. For clarity, many practical details are included in the following description. However, it should be understood that these practical details are not intended to limit the present invention. In other words, in some embodiments of this application, these practical details are not essential.

[0058] To provide an effective tool for multi-label text interpretation, this paper proposes a multi-label topic modeling method (CPTM) based on the integration of multi-source semantic knowledge, based on the current research hotspot of large language models and effective pre-trained models for text representation tools. This method utilizes a thought chain prompting project based on large models, incorporates keyword information, context, and label description information, aligns text fragments to their semantically related labels, and provides sentence-level interpretation, ensuring the accuracy of subsequent word visualization and the semantic consistency of mined topics and labels. For word visualization and topic mining, this application incorporates the contextual knowledge of the pre-trained language model to ensure the effectiveness of word interpretation and label interpretation.

[0059] like Figure 1 As shown, the present application is a multi-label text interpretation method based on thought chain prompt engineering, and the multi-label text interpretation method includes the following steps:

[0060] Step 1: Use crawler technology to obtain corpus from the multi-label text platform, clean the corpus, filter out texts with less than three text fragments, and delete irrelevant content in the text such as emoticons and URLs to obtain a multi-label text corpus collection;

[0061] Step 2: For each label in the multi-label text corpus, use the large language model to generate label description information, and use the thinking chain prompting project based on the large language model to align the text fragments in each document in the multi-label text corpus to the labels that are semantically related to the text fragments, that is, assign a label to each text fragment, and use a text fragment that is not related to any label to align the text fragments in each document in the multi-label text corpus to align the text fragments in each document to the labels that are semantically related to the text fragments. Labels are aligned and the topic vector representation matrix is obtained based on the pre-trained language model , as attached Figure 2 As shown, the specific steps include:

[0062] Step 2.1: For each tag in the tag set corresponding to the multi-label corpus, such as "mystery," "romance," "war," and "sports," we first use the large language model to generate corresponding description information for each tag. This step avoids the problem of insufficient semantic information provided by the tags during the topic alignment phase. It ensures the accuracy of the subsequent topic representation matrix.

[0063] Step 2.2: Then, for each text in the multi-label text corpus , first of all, 'Add label to text Corresponding tag set Get .

[0064] Step 2.3, then use the thought chain prompt project to align the text fragments to Go to the label in the Figure 2 As shown, for the document , the corresponding label set of the three text fragments is {mysterious, romantic}, and the final topic alignment result is: fragment 1 corresponds to 'romance' and 'mysterious', fragment 2 corresponds to 'romance', and fragment 3 is not related to any label, which is ' '.

[0065] Step 2.4: Based on the above steps, for each label in the label set corresponding to the multi-label text corpus, this application can obtain a set of text segments corresponding to each label. Then, based on the pre-trained model, the vector representation of each label is obtained by averaging the representation of the pre-trained language model of the text segment corresponding to each label, and the topic representation matrix is obtained. Used for subsequent tag-related word visualization and tag-related topic extraction.

[0066] The topic vector representation matrix is used to subsequently visualize tag-related words and mine tag-related topics. Existing methods often assume that text snippets are a mixture of document labels, which results in each word being associated with a label, leading to inaccurate word visualization. This application, however, ensures accurate word-label correlation by assigning relevant labels to sentences.

[0067] Step 3: Based on the labels assigned to the text fragment by the large language model, the words in the sentence are visualized by calculating the correlation between the word vector of each word in the text fragment and the corresponding vector of the label assigned to the text fragment; the correlation is calculated by the cosine similarity between the vector representation of the word and the vector representation of the label. Figure 3 In the example shown, step 3 specifically includes the following steps:

[0068] Step 3.1, Attachment Figure 3 An example of text The visualization results for the first text snippet in

[15] are shown. This text snippet was assigned the labels "romantic" and "mysterious" by the large language model through the thought chain prompting project. First, this application obtains vector representations for each word in the snippet using the pre-trained language model.

[0069] Step 3.2. Then, calculate the cosine similarity between the vector representation of each word and the vector representation of the assigned labels 'romantic' and 'mysterious' to obtain the label-word correlation matrix .

[0070] Step 3.3: Based on the label-word correlation matrix The relevance score between the label and each word is used to visualize the word.

[0071] Existing multi-label topic models rely solely on word co-occurrence information when mining topics, without incorporating external semantic knowledge. This results in low quality and interpretability of mined topics, as well as low semantic consistency between labels and topics. To address these shortcomings, this application incorporates knowledge from pre-trained language models for topic mining.

[0072] Step 4: Obtain the topic vector representation matrix based on the label corresponding to the text segment fed back by the large language model and step 2 , extract topics that are semantically consistent with the labels in the multi-label text corpus and provide explanations for the labels. Different from the existing multi-label topic modeling methods, the method of this application uses a large language model to perform sentence-level topic alignment, while avoiding irrelevant text fragments from participating in topic inference, ensuring the accuracy of topic vector representation, and facilitating subsequent word visualization and topic mining, such as Figure 4 As shown, the specific steps include:

[0073] Step 4.1, word-topic distribution: to mine the tag set With each label Semantically consistent topics provide fine-grained label explanations. This application designs a pre-trained language model based on Specifically, for the first word in the vocabulary containing different words in the multi-label text corpus, words , , Indicates the number of different words in the multi-label corpus. Get word Every occurrence in the multi-label corpus , then for the word No. times, expressed as , , obtain words through pre-trained language model Contextual representation of :

[0074]

[0075] in, Indicates the The context of the occurrence of express The number of times it appears in the multi-label corpus;

[0076] word Local word-topic distribution of The calculation method is:

[0077]

[0078] in, represents the cosine similarity, for Normalization operation, Indicates the The topic vector corresponding to the label is represented by ,word Corresponding corpus-level word-topic distribution It is obtained by the following formula:

[0079]

[0080] Calculate the corpus-level word-topic distribution of each word in the vocabulary, and then get the global word-topic distribution matrix by taking the average ;

[0081] Step 4.2: Based on the obtained global word-topic distribution matrix , through column normalization operation, we get the final topic-word distribution matrix , specifically:

[0082]

[0083] in, Represents the column normalization operation. For each label’s corresponding topic-word distribution matrix, the top 10 words with the highest probability in the topic-word distribution matrix are obtained to represent the label, providing a fine-grained explanation of the label. Figure 4 As shown, for each tag in the tag set, such as 'sports', 'mystery', 'romance', 'war', etc., the topics related to each tag are extracted.

[0084] This application aims to investigate the effectiveness of multi-label text interpretation and conducts experiments on plot text data of IMDB movies. For the interpretation of labels, we use the indicators (CP, CA, NPMI, UCI) and the topic diversity index (UTR) commonly used in topic models to measure topic coherence. At the same time, we design two indicators to measure the semantic consistency between the extracted topics and labels, namely Label Discovered Rate (LDR) and Correlation Score (CS). For sentence-level interpretation, this application uses manually annotated datasets to measure using MicroF1 and MacroF1 classification indicators. For comparison, this application uses unsupervised topic model comparison examples 1 and 2, and for multi-label topic models we use comparison examples 3 and 4. For word-level interpretation, we use manual evaluation methods to calculate the top 5 words most relevant to the label obtained by the model in each sentence, and then manually re-rank the 5 words corresponding to each shuffled sentence according to their relevance to the label, and then calculate the difference between the two sorted sequences. Two evaluation indicators for measuring sequence differences are used here, the Kendall correlation coefficient and Spearman coefficient Table 1 shows the numerical results of label explanations, i.e., topic quality. Table 2 shows the numerical results of semantic consistency between labels and topics. Table 3 shows the numerical results of sentence-level explanation accuracy. Table 4 shows the results of manual evaluation of word-level explanations.

[0085] Table 1

[0086]

[0087] Table 2

[0088]

[0089] Table 3

[0090]

[0091] Table 4

[0092]

[0093] The experimental results are as follows: the average results of the tag-level explanation and the topic quality metrics are shown in Table 1. The LDR and CS results are shown in Table 2. The sentence-level explanation MicroF1 and MacroF1 are shown in Table 3. The word-level explanation, Kendall correlation coefficient and Spearman coefficient As shown in Table 4, it can be seen that the evaluation results of this application are higher than those of the comparison method.

[0094] This experiment uses comparative models including unsupervised topic model comparison example 1, comparison example 2, and multi-label topic model comparison example 3, comparison example 4.

[0095] The comparison results of this application include the evaluation of labels, text fragments and word visualization. First, the explanation at the label level, as shown in Table 1, is a comparison between Comparative Examples 1-4 and this application. The results show that the method of this application outperforms Comparative Examples 1-4 in terms of topic quality indicators (CP, CA, NPMI, UCI) and topic diversity indicators (UTR). At the same time, as shown in Table 2, the performance of the label and topic semantic consistency indicators is also better than Comparative Examples 1-4. For the interpretation of text fragments, this application is compared with Comparative Examples 3 and 4. The comparison results are shown in Table 3. It can be seen that the method of this application is better than Comparative Examples 3 and 4. Finally, for the explanation at the word level, this application uses a manual evaluation method. The results are shown in Table 4. It can be seen that the method of this application can provide effective word-level explanations. In summary, this application outperforms Comparative Examples 1-4 in all evaluation indicators.

[0096] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A multi-label text interpretation method based on thought chain prompt engineering, characterized by: The multi-label text interpretation method comprises the following steps: Step 1: Obtain the corpus on the multi-label text platform, clean the corpus, and obtain a multi-label text corpus collection; Step 2: For each label in the multi-label text corpus, use the large language model to generate label description information, and use the thought chain prompting project based on the large language model to align the text fragments in each document in the multi-label text corpus to the labels that are semantically related to the text fragments, that is, assign a label to each text fragment. At the same time, align the text fragments that are not related to any label with an 'irrelevant' label, and obtain the topic vector representation matrix E based on the pre-trained language model. t ; Step 3: Based on the labels assigned to the text fragment by the large language model, the words in the sentence are visualized by calculating the correlation between the word vector of each word in the text fragment and the corresponding vector of the label assigned to the text fragment; Step 4: Get the topic vector representation matrix E based on each label in the multi-label text corpus and step 2 t , extract topics that are semantically consistent with the labels in the multi-label text corpus and provide explanations for the labels.

2. The multi-label text interpretation method based on thought chain prompting engineering according to claim 1 is characterized by: The step 1 obtains corpus on a multi-label text platform and cleans the corpus by filtering texts with less than three sentences in the corpus and deleting emoticons and website links in the corpus text.

3. The multi-label text interpretation method based on thought chain prompting engineering according to claim 1 is characterized by: The step 2 specifically includes the following steps: Step 2.1: For each tag in the multi-tag text corpus, generate a tag description using the large language model, use the tag description to supplement the semantics of the tag, and check the generated tag description; Step 2.2, each document d={s1,s2,...,s M The corresponding tag set is: Where M represents the number of sentences in document d, L d Represents the tag set corresponding to document d, K represents the last tag in the tag set corresponding to document d. d Indicates the number of tags in the tag set corresponding to document d; For each label corresponding to a document in a multi-label text corpus, first add the 'irrelevant' label to In the multi-label corpus set corresponding to the document, the enhanced label set corresponding to the document d is obtained Step 2.3: For each text segment s in document d i , i∈{1,2,...,M}, first let the large language model summarize each text segment s i The keywords in the text are combined with the context of the keywords to purify the extracted keywords. According to the purified keywords, combined with the label description information generated in step 2.1, a tag is generated for each text segment s. i Assign at least one semantically relevant label; Step 2.4: For all texts in the corpus, construct labels l k The set D of sentences ∈L k , use the pre-trained language model to obtain the sentence set D k The jth text segment The word vector of each word in , average the word vector representation of each word to obtain the text fragment s j Vector representation of The above transformation is expressed as follows: Among them, L is the label set, Represents a text fragment s j The number of words in Represents a text fragment s j The last word in Represents a text fragment s j The context word vector of the nth word in ; Based on the above transformation, we get the vector representation of topic k in, represents the pre-trained language model, N k For the sentence set D k The number of sentences in the topic vector representation matrix E is obtained based on the pre-trained language model t ; in, Represents the hidden dimension of the pre-trained language model, and K represents the number of labels in the label set corresponding to the multi-label corpus.

4. The multi-label text interpretation method based on thought chain prompting engineering according to claim 3 is characterized by: Step 3 specifically includes the following steps: Step 3.1, context word vector representation acquisition: given the i-th text segment s in document d i , text fragments i The word sequence contained is and text snippets i The set of assigned tags Through pre-trained language model Get the context representation of each word, i.e. the word vector: Among them, w' n Representing word sequences The nth word in Represents a text fragment s i The number of words contained in for w' n The word vector of represents the assignment of a large language model to a text segment s i The number of labels, Representing word sequences The words in Step 3.2: Calculate the label-word correlation matrix: Based on the word vector calculated in step 3.1 word sequence The nth word w' in n With label collection The k'th label in Semantic relevance Calculated by the following formula: in, Representation and Labeling The corresponding topic vector representation; Step 3.3: Based on word vector Vector representation of labels Constructing a label-word correlation matrix Capture the semantic correlation between words and labels and visualize the words in a sentence.

5. The multi-label text interpretation method based on thought chain prompting engineering according to claim 1 is characterized by: The step 4 specifically includes the following steps: Step 4.1: For the vth word w in the vocabulary containing different words in the multi-label text corpus v , v∈{1,2,...,V}, V represents the number of different words in the multi-label corpus, get word w v Every occurrence in the multi-label corpus word w v The nth occurrence of n∈{1,2,...,N v }, get word w through pre-trained language model v Contextual representation of in, Indicates the context representation of the nth occurrence, N v Representation word w v The number of times it appears in the multi-label corpus; word w v Local word-topic distribution of The calculation method is: Among them, s(·) represents cosine similarity, smax(·) is the softmax normalization operation, Represents the topic vector corresponding to the kth tag, k∈{1,2,...,K}, word w v Corresponding corpus-level word-topic distribution It is obtained by the following formula: Calculate the corpus-level word-topic distribution of each word in the vocabulary and obtain the global word-topic distribution matrix Step 4.2: Based on the obtained global word-topic distribution matrix Through column normalization operation, the final topic-word distribution matrix is obtained Specifically: Φ=norm c (C') Among them, norm c () represents the column normalization operation. For each label’s corresponding topic-word distribution matrix, the top 10 words with the highest probability in the topic-word distribution matrix are obtained to provide a fine-grained explanation of the label.

Images