A focus entity attribute prediction technology based on natural language processing

Abstract

The present application belongs to the field of natural text analysis and prediction, and is used for focus type entity attribute prediction. The present application provides a focus entity attribute prediction technology based on natural language processing, which can not only output the objective prediction value commonly seen in the current attribute prediction model, but also can induce and output different expected values or evaluations of different attributes of the entity in the text data. According to the entity set and attribute set provided by the user, the present application constructs a larger related entity set or an iterative set and crawls a large amount of text data for basic attribute prediction and attribute development trend prediction. The expected value and evaluation output by the present application can be used to guide the optimization of different entities or predict the development direction thereof. In view of the problems of low utilization rate of current text data and small amount of hidden information mining, the present application improves the ability of accurate acquisition of focus entity information and provides more abundant and valuable data for the downstream application of attribute prediction.

Description

Technical Field

[0001] This invention belongs to the field of natural text analysis and prediction, and is particularly suitable for predicting the attributes of focus-type entities. Its purpose is to address the current problems of low utilization and poor analytical performance of internet text data, and to improve the ability to accurately acquire focus entity information in specific domains and perform attribute analysis and prediction. This invention combines natural language understanding and analysis to infer the attribute characteristics of entities, the attention others pay to them, and their expected development. Background Technology

[0002] Predictive models are applied in various fields, with attribute prediction being a popular application. Attribute prediction typically relies on data sources of text and images, while the focus of the prediction is usually non-biological entities such as products and items, and biological entities such as people and animals. Taking user attribute prediction as an example, text-based prediction sets the target entity as an internet user, collecting their behavioral trajectory information on social networks, social forums, or news websites, such as comments, favorites, and likes. This behavioral information is extracted and its features are analyzed and fed into the predictive model to predict the user's age, occupation, interests, and other attribute characteristics. Image-based user attribute prediction, on the other hand, uses a given image of a person, analyzing and summarizing details in the image to output predefined target attribute prediction values, such as age, height, and gender. Video image-based prediction can further analyze behavioral habits, physical condition, and other attribute characteristics over a specific period.

[0003] The characteristics of the above attribute prediction models are: (1) the information sources are all generated or provided by the behavior of the focus entity, and the data are objective text or images describing subjective behavior; (2) the prediction results are descriptive text, aiming to realistically and objectively describe the behavior and characteristics of the focus entity. Further analysis of the attributes is still needed when downstream applications use the prediction results.

[0004] This invention proposes an entity attribute prediction technology based on natural language processing. It provides a complete workflow template for entity attribute prediction in a focused domain, including data collection, analysis, and prediction. It focuses more on textual subjective evaluations of focus entities, generating emotionally charged opinion assessments of these entities, and ultimately summarizing and predicting certain attributes of entities from a network perspective using statistical methods. This invention further improves the scope of attribute prediction based on existing attribute prediction models, extracting valuable attribute features from existing data to support more intelligent data analysis techniques. Summary of the Invention

[0005] This invention provides a focus entity attribute prediction technology based on natural language processing to effectively solve the problems of low utilization rate and poor analysis performance of existing text data. This improves the ability to accurately acquire focus entity information in specific domains and perform attribute analysis and prediction, providing valuable potential attribute information in addition to objective descriptions for downstream data applications.

[0006] To achieve the above objectives, a focus entity attribute prediction technique based on natural language processing is proposed. Its key feature is that it does not specify a particular person or object, but requires a list of focus entities and focus attributes within a specified domain. It analyzes sentiment and opinions by collecting news and commentary texts from the internet, and extracts predicted values ​​of focus attributes and information such as online users' attention, opinions, and expectations regarding the entity. Specifically, the method includes the following steps:

[0007] 1) Entity Data Collection. Users provide a list of entities and key attributes specific to a particular domain, guiding the collection of relevant data and identifying key descriptive attributes for analysis during the prediction process. For entities in the list, their full names are first obtained. Then, highly relevant related entities are retrieved using a search engine. For entities with iterative relationships, their predecessor / level or successor / level entities are selected, such as iterations of the Huawei mobile phone series. The entity data to be collected is reconstructed based on the iterative and related relationships. Entity names are retrieved via web crawling, and information such as data source, time, material title, and material text is collected.

[0008] 2) Data Summary Extraction. For entity E, let its intelligence news data be D: {d1, d2, d3, d4, d5…}. First, extract the text summary A {a1, a2, a3…} of its news content D. Based on the final data display material support requirements, because the obtained equipment data text materials have varying lengths, and some text materials are too long, which may cause confusion in the final display, the first step is to extract the summary of the equipment data text materials. The text summary extraction is based on the TextRank algorithm, which originates from the browser content recommendation algorithm PageRank. First, identify the text title and text content. By segmenting the text title and calculating its frequency, the weight is recorded, resulting in the title segmentation weight table T: {w1: r1, w2: r2… w n :r n}, where w n r represents each word in the word segmentation set. n Assign corresponding weights. Obtain the word segmentation weight table T. Then, segment the text content into sentences, divide the content, and calculate the weights of the words in the word segmentation set contained in each sentence. Finally, the weight set of each sentence is obtained, and they are sorted. Then, based on the required number of words for the summary, the sentences with the highest weights are selected from the weight sort to construct the text summary.

[0009] 3) Data Attribute Prediction. This involves extracting key information from the text, including the equipment's strengths, weaknesses, advantages, disadvantages, and other deficiencies, as well as the emotional orientation and viewpoints of the text's author or initiator. In practice, adjustments are necessary because the original material for key information extraction differs from traditional commentary in both volume and purpose. Since the text is too large, directly using it as the extraction material would fail to capture key information; therefore, it is sliced. For text d1, it is segmented into sentence pieces, constructing a sentence set S: {s1, s2, s3…s}. n For a user-provided list of key attributes, keywords from different slices are directly extracted for routine attribute value prediction. Opinion extraction treats each slice's statement as a separate comment, implemented using the PaddleNLP comment opinion extraction tool. The basic principle is to extract opinions using sequence labeling; specifically, attributes and corresponding opinions are extracted from the comments. After extracting the opinion, targeted comments can be made on each attribute. The comment text string is fed into the SKEP model, which performs semantic encoding and then predicts the corresponding label based on the output at each position. The extracted comment attributes and opinion are concatenated and then combined with the original statement to form an independent training statement. Comment attributes and opinion words are concatenated, and then the concatenated words are combined with the original text and fed into the SKEP model. The vector at the "CLS" position is used for fine-grained sentiment analysis. Finally, this method yields the opinion extraction results for each sentence slice of each news text. The extraction results are then filtered to remove useless information caused by errors in the original material, ultimately constructing the advantages, disadvantages, characteristics, strengths, weaknesses, and other defects of each piece of equipment. For each predicted value and opinion attribute, a quantitative test is performed on the final result. This means that all text materials with the same root and the same text are considered as the same material. If the frequency F of the predicted attribute value is greater than 75% of the total number of materials among all non-repeating materials, the prediction result is considered reliable.

[0010] 4) Data Attribute Analysis. After data attribute prediction, this invention should obtain two types of results: the attribute values ​​(predicted results) of the key attributes given by the user and the author's opinion on the data attributes. Both attribute values ​​and evaluation opinions are assessed and verified based on related or iterative entities in this step. For entity E0, entity E1 is a secondary or highly related entity. The key values ​​v0 and v1 of the same key attribute A are compared. Based on data analysis, key attribute values ​​with significant differences are identified, and the author's opinion on that key attribute is also compared. The deviation threshold is denoted as |δ|. If |v0-v1|≥|δ|, it is determined whether the author's opinion makes |v0-v1| tend to decrease. If so, the author's opinion is accepted, considered credible, and represents a possible development or optimization direction.

[0011] Furthermore, the specific steps of step 1) above are as follows:

[0012] Step 1)-1: Initial state;

[0013] Step 1)-2: Input the set of focus entities and the set of key attributes V = {V1, V2, V3, ...};

[0014] Steps 1)-3: Obtain highly relevant entities, pre-type / level or sub-type / level entities through a search engine, and update the entity set;

[0015] Steps 1)-4: For entities with iterative relationships, construct the iterative path;

[0016] Steps 1)-5: Use a web crawler to search for data on the entities in the entity set;

[0017] Steps 1)-6: End state;

[0018] Furthermore, the specific steps of step 2) above are as follows:

[0019] Step 2)-1: Initial state;

[0020] Step 2)-2: Input entity E and intelligence news data D: {d1, d2, d3, d4, d5…};

[0021] Steps 2)-3: Segment the title and record the frequency and weight table T;

[0022] Steps 2)-4: Based on the weight table T, record the frequency weights of sentence segmentation in the text content;

[0023] Steps 2)-5: Sort the sentences by weight and select the top n sentences to construct a summary;

[0024] Steps 2)-6: Output and display the summary;

[0025] Steps 2)-7: End state;

[0026] Furthermore, the specific steps of step 3) above are as follows:

[0027] Step 3)-1: Initial state;

[0028] Step 3)-2: Input text data d1;

[0029] Step 3)-3: Slice the data into sentences to obtain the sentence set S: {s1, s2, s3…s} n};

[0030] Steps 3)-4: Extract statement s n Predicted values ​​of key attributes in the data;

[0031] Steps 3)-5: N is the total number of non-repeating materials. If the frequency F of a certain attribute prediction value is less than N, the guess value of that attribute is eliminated.

[0032] Steps 3)-6: Extract statements using PaddleNLP n The viewpoints in the text consist of keywords related to equipment attributes and their evaluations.

[0033] Steps 3)-7: Filter out useless information that appears due to material errors;

[0034] Steps 3)-8: Output the key attribute V n and predicted value v n Output evaluation opinions;

[0035] Step 3)-9: End state;

[0036] Furthermore, the specific steps of step 4) above are as follows:

[0037] Step 4)-1: Initial state;

[0038] Step 4)-2: Input the set of predicted key attributes and output the viewpoint;

[0039] Step 4)-3: Based on the entity set, traverse and retrieve entities E with high relevance or iterative relationships. i E j ;

[0040] Step 4)-4: Iterate through the common attributes of the two entities. If the attribute |E i :v k -E j :v k |>|δ|, and Ei for attribute V k Emotional perspectives make |E i :v k -E j :vk If the value decreases, the viewpoint is considered reliable;

[0041] Steps 4)-5: Output reliable sentiment perspectives as predictions of development or optimization directions;

[0042] Steps 4)-6: End state. Attached Figure Description

[0043] Figure 1 This is a flowchart illustrating a focus entity attribute prediction method based on natural language processing in an embodiment of the present invention.

[0044] Figure 2 for Figure 1 A flowchart for collecting entity data.

[0045] Figure 3 for Figure 1 A flowchart for extracting data summaries from Chinese.

[0046] Figure 4 for Figure 1 A flowchart for predicting data attributes in a dataset.

[0047] Figure 5 for Figure 1 A flowchart for data attribute analysis. Detailed Implementation

[0048] To better understand the technical content of the present invention, specific embodiments are described below in conjunction with the accompanying drawings.

[0049] Figure 1 This is a flowchart of a focus entity attribute prediction technology based on natural language processing in an embodiment of the present invention.

[0050] The focus entity attribute prediction technology based on natural language processing is characterized by the following steps.

[0051] S1 Entity Data Acquisition: Given an initial entity set and a set of key attributes. For existing entities, retrieve highly relevant entities, or pre-type / level or sub-type / level entities using a search crawler, and construct an iterative update path.

[0052] S2 data summarization extraction takes entity and text data from the dataset prepared in the previous step as an example. Generally, the collected data includes two main parts: titles and text content. Titles are segmented into words, and the weights of different nouns are calculated based on their frequency. The text content is then segmented sentence by sentence, and the weights of different sentences are calculated based on the weights of nouns appearing in the titles. The sentences are sorted in descending order of their weight sums, and the top n sentences with the largest weight sums are selected to construct a summary, facilitating the presentation of the main content of the data.

[0053] S3 data attribute prediction involves segmenting textual material for a given entity into sentences to construct a set of sentences. Each sentence is treated as a separate comment, and PaddleNLP is used to extract viewpoints to obtain predicted values ​​for the equipment-specific attributes involved in the sentence, as well as evaluations of those attributes. Finally, the obtained attribute prediction values ​​are filtered based on frequency to obtain reliable attribute prediction values ​​for each entity.

[0054] S4 data attribute analysis processes the opinions and evaluations of key attributes obtained from attribute prediction in the text. It determines the feasibility and credibility of the text's opinions and evaluations based on the differences in predicted values ​​between entities with high correlation or iterative relationships.

[0055] Figure 2 This is a flowchart for entity data collection. For an entity set, a web crawler is used to obtain highly related entities, their predecessors or successors, and the entity set is updated. After obtaining the complete entity iteration path, data is crawled from all entities along the path. The specific steps are as follows:

[0056] Step 1: Initial state; Step 2: Input the set of focus entities and the set of key attributes V = {V1, V2, V3, ...}; Step 3: Obtain highly related entities, pre-type / level or sub-type / level entities through a search engine, and update the entity set; Step 4: For entities with iterative relationships, construct iterative paths; Step 5: Search for data on the entities in the entity set using a web crawler; Step 6: End state;

[0057] Figure 3 This is a flowchart for data summarization. This step is used to extract the central idea of ​​long news texts and standardize the length of intelligence texts of varying lengths. The method focuses on analyzing keywords in the text titles, assigning different weights to these keywords, and analyzing the relevance of different sentences in the body text, which is represented by weighted sums. Finally, the sentences with the highest weighted sums are selected and semantically reorganized to form sums of uniform length.

[0058] The specific steps are as follows:

[0059] Step 1: Initial state; Step 2: Input entity E and intelligence news data D: {d1, d2, d3, d4, d5…}; Step 3: Segment the title and record the frequency weight table T; Step 4: Based on the weight table T, segment the sentences in the text content and record the frequency weight sum; Step 5: Sort the weight sums of the sentences and select the top n sentences to construct a summary; Step 6: Output and display the summary; Step 7: End state;

[0060] Figure 4This is a flowchart for data attribute prediction. The text is segmented sentence by sentence. Sentences are then segmented into words, and equipment-related professional attribute nouns and sentiment tendencies are extracted to determine the viewpoint of each sentence. The sentiment-oriented viewpoint and professional attributes are combined to obtain the text's opinion evaluation, which is used to predict the entity's development trend and improvement direction. Simultaneously, basic key attribute value predictions are obtained based on traditional attribute prediction methods such as attribute keywords and matching principles. The specific steps are as follows:

[0061] Step 1: Initial state; Step 2: Input text data d1; Step 3: Slice the data into sentences to obtain the sentence set S: {s1, s2, s3…s n Step 4: Extract statement s n Step 5: N is the total number of unique materials. If the frequency F of a certain attribute prediction value is less than N, the guessed value of that attribute is discarded. Step 6: Use PaddleNLP extraction statements s n The viewpoint consists of equipment attribute keywords and their evaluations; Step 7: Filter out useless information caused by material errors; Step 8: Output key attribute V n and predicted value v n Step 9: Output evaluation opinions; End status;

[0062] Figure 5 This is a flowchart for data attribute analysis. The obtained evaluation opinions are validated by examining the attribute gaps between highly correlated entities. If an evaluation opinion containing a key attribute gap points to reducing the gap between correlated or iterative entities, then the evaluation opinion is considered effective. The specific steps are as follows:

[0063] Step 1: Initial state; Step 2: Input key attribute prediction value set and output viewpoint; Step 3: Based on the entity set, traverse and select entities E with high relevance or iterative relationship. i E j Step 4: Iterate through the common attributes of the two entities. If the attribute |E i :v k -E j :v k |>|δ|, and Ei for attribute V k Emotional perspectives make |E i :v k -E j :v k If the value decreases, the viewpoint is considered reliable; Step 5: Output reliable sentiment viewpoints as predictions of development or optimization directions; Step 6: End state;

[0064] In summary, this invention addresses the problem of low utilization of text data in attribute prediction. By performing semantic-level analysis and understanding on existing limited text data, and based on the prediction of key attribute values, it adds subjective viewpoint extraction to predict the development direction and improvement direction of entities, thereby providing more hidden information and offering more valuable predictive information for downstream applications.

Claims

1. A method for predicting the attributes of a focal entity based on natural language processing, characterized in that, This method analyzes sentiment orientation in retrieved equipment text data based on natural language understanding, and associates text content with key attributes of specific entities by identifying positive and negative sentiment trends to predict entity attribute values. Simultaneously, for a single-focus entity, it constructs a set of entities with high correlation or iterative relationships, and infers the rationality of existing natural language text evaluations of the attributes based on the associated entity attribute values, thereby determining the entity's potential development and optimization directions. Specifically, this method includes the following steps: 1) Entity Data Acquisition: Users provide a list of entities and key attributes in a specific domain to guide the focus entity attribute prediction method on which relevant data to collect and which basic descriptive attributes need to be analyzed during the prediction process. For entities in the entity list, their full names are first obtained, and then highly relevant related entities are retrieved through a search engine. For entities with iterative relationships, their predecessor / level or subordinate / level entities are selected. For iterations of the same Huawei mobile phone series, the entity data to be collected is reconstructed based on the iterative and related relationships. Entity names are retrieved through web crawling, and data sources, time, material titles, and material text information are collected. 2) Data Summary Extraction; For entity E, let its intelligence news data be D: {d1, d2, d3, d4, d5…}; First, extract the text summary A{a1, a2, a3…} of its news content D. Based on the final data display material support requirements, because the obtained equipment data text materials have varying lengths, and some text materials are too long, causing confusion in the final display, the first step is to extract summaries from the equipment data text materials. The text summary extraction is based on the TextRank algorithm, which originates from the browser content recommendation algorithm PageRank. First, confirm the text title and text… This section describes a process where the text title is segmented into words, and the frequency of each word is calculated and recorded as its weight. This results in a word segmentation weight table T: {w1:r1, w2:r2…wn:rn}, where wn represents each word in the segmentation set and rn represents its corresponding weight. After obtaining the word segmentation weight table T, the text content is further segmented into sentences. The weights of the words in each sentence's segmentation set are calculated, resulting in a weight set for each sentence. These weights are then sorted, and finally, based on the required word count for the summary, the sentences with the highest weights are selected from the weighted list to construct the text summary. 3) Data Attribute Prediction; This involves extracting key information from the text, including the advantages, disadvantages, strengths, weaknesses, and other defects of equipment, as well as the emotional orientation and viewpoints of the author or initiator. In practice, adjustments are needed because the original material for key information extraction differs from traditional commentary in both volume and purpose. Due to the large amount of text content, direct extraction would fail to capture key information; therefore, it is sliced. For text material d1, sentence segments are created, constructing a sentence set S: {s1, s2, s3…sn}. For the user-provided list of key attributes, keywords from different segments are directly extracted for conventional attribute value prediction. Viewpoint extraction treats each segment's sentences as a separate comment, using the PaddleNLP comment viewpoint extraction tool. The basic principle is to extract viewpoints using sequence labeling. Specifically, attributes and corresponding viewpoints are extracted from the comment. After extracting the comment viewpoints, targeted analysis is then performed on each segment. The process involves commenting on attributes; specifically, the comment text string is fed into the SKEP model, which performs semantic encoding on the text string and then predicts the corresponding label based on the output of each position; the extracted comment attributes and comment opinions are concatenated and then concatenated with the original sentence as an independent training sentence; the comment attributes and opinion words are concatenated and then concatenated with the original text and fed into the SKEP model, and the vector at the "CLS" position is used for fine-grained sentiment analysis; finally, the opinion extraction results for each sentence slice of each news text are obtained through this method, and the extraction results are then filtered to remove some useless information caused by errors in the original material, and finally, the advantages, disadvantages, characteristics, strengths, weaknesses, and other defects of each piece of equipment are constructed; for each predicted value and the final result of the opinion attribute, a quantity test is performed, that is, all text materials with the same root and the same text are regarded as the same material, and among all non-repeating materials, if the frequency F of the predicted attribute value is greater than 75% of the total number of materials, the prediction result is considered reliable; 4) Data Attribute Analysis; After data attribute prediction, two types of results should be obtained: the attribute values ​​of the key attributes given by the user and the text writer's opinion on the data attributes. In this step, both attribute values ​​and evaluation opinions are evaluated and verified based on related entities or iterative entities. For entity E0, entity E1 is a secondary or highly related entity. Compare the key values ​​v0 and v1 of the same key attribute A. Based on data analysis, identify the key attribute values ​​with large differences. At the same time, compare the author's opinion on the key attribute. Let the deviation threshold be |δ|. If |v0-v1|≥|δ|, then determine whether the author's opinion makes |v0-v1| tend to decrease. If it is true, the author's opinion is accepted and considered to have credibility, which is a possible development direction or optimization direction.

2. The method for predicting the focus entity attribute based on natural language processing according to claim 1, characterized in that, In step 1), not only are individual isolated entities initialized, but also a complete iterative update path or a highly related relation table is constructed to identify their evolving or degenerating professional attributes in order to verify the reliability of subsequent opinion evaluations.

3. The method for predicting the focus entity attribute based on natural language processing according to claim 1, characterized in that, In step 2), based on the frequency weight of the title segmentation, the keywords appearing in each sentence of the text are weighted and finally reorganized to obtain a new summary.

4. The method for predicting the focus entity attribute based on natural language processing according to claim 1, characterized in that, In step 4), after the viewpoint is extracted in step 3), it is evaluated based on the objective predicted value of the attribute to verify its credibility and feasibility, thereby ensuring that the viewpoint output by the present invention conforms to the development direction or expected development direction of the focus entity.

Images