Zero sample image attribute labeling method based on pre-training large model

Abstract

The invention discloses a zero sample image attribute annotation method based on a pre-training large model, and belongs to the field of image annotation, and the method comprises the steps: obtaining a pre-training large language model, interacting with the pre-training large language model, preliminarily defining attributes, repeatedly screening and refining the attributes, and defining values of the attributes, and obtaining a group of attribute values with discrimination for category recognition in a specific domain; the defined attributes and common values thereof are converted into grouped candidate text options, a vision-language model is obtained, the presentation degree of each attribute value in each image is labeled by utilizing the image-text retrieval capability of the vision-language model, and image labeling is converted into category labeling. Based on the ability of the pre-training model, a method for designing and labeling attributes for zero sample learning is provided, and the dependence of a zero sample learning training method on manual labeling data is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of image annotation, and in particular to a zero-sample image attribute annotation method based on a pre-trained large model. Background Art

[0002] Zero-shot learning is an image classification task in which the category intersection of the training data and the test data is empty. Zero-shot learning does not require labeling every image of all categories. Instead, it uses category-level semantic embedding annotations to transfer knowledge learned from labeled category images to unlabeled categories, assisting in the recognition of unseen images. The semantic embedding annotations that connect the training and test categories can be attribute information such as color, texture, and habitat, or word embeddings of category names, knowledge graph embeddings, and other information. Attribute information is a powerful representational information, but it usually requires expert knowledge and human annotation, and is relatively expensive to obtain. However, word embeddings and knowledge graph embeddings, which are relatively cheap to obtain, have poor representational capabilities and cannot achieve good results in zero-shot learning.

[0003] With advances in training data volume, model size, and computing power, pre-trained models have demonstrated remarkable versatility, including conversational models like GPT and visual-language multimodal models like CLIP. These models are beginning to be applied to data generation and annotation, but have yet to be applied to data annotation for zero-shot learning. Summary of the Invention

[0004] The present invention provides a zero-shot image attribute labeling method based on a pre-trained large model. This method can label attributes with high representational capabilities, reducing the labeling cost while bringing significant performance improvement to zero-shot learning, which is of great significance.

[0005] An embodiment of the present invention provides a zero-shot image attribute labeling method based on a pre-trained large model, comprising the following steps:

[0006] S1, obtaining a pre-trained large language model, and obtaining a set of attributes and attribute values for image classification in a specific domain by interacting with the large language model;

[0007] S2: Convert the attributes and attribute values into groups of candidate text options, obtain a pre-trained visual-language model, use the visual-language model to infer the presentation degree of each attribute value in each image, and convert them into image annotations and category annotations.

[0008] Optionally, in one embodiment of the present invention, step S1 specifically includes:

[0009] S101, interacting with the large language model to preliminarily define attributes in a specific domain through question-answering;

[0010] S102, after alternately screening and refining attributes, a set of effective attributes for zero-shot image classification in a specific domain is obtained;

[0011] S103: Continue interacting with the large language model to obtain common values for each attribute.

[0012] Optionally, in one embodiment of the present invention, in step S1, interaction with the large language model is performed through application programming interface access or model access.

[0013] Optionally, in one embodiment of the present invention, in step S102, attributes that do not meet the requirements are eliminated by repeatedly screening and refining the attributes.

[0014] Optionally, in one embodiment of the present invention, step S2 further includes:

[0015] The image-text retrieval capability of the visual-language model is used to calculate the degree of presentation of each attribute value in each image, and the image-text retrieval output is converted into image annotations. The annotations of all images in each category are then averaged to obtain the annotations for each category.

[0016] Optionally, in one embodiment of the present invention, in step S2, two template sentence patterns are used to convert the attributes and attribute values into a group of candidate text options, into which the <field> of the dataset to be annotated, the obtained <attribute>, and the attribute <value> are filled;

[0017] (1) Non-grouping transformation method: Each attribute value is individually transformed into a set of text options, which contains two candidate options, namely the affirmative text statement "In this <domain> image, its <attribute> is <value i>" and the negative text statement "In this <domain> image, its <attribute> is not <value i>";

[0018] (2) Grouping transformation method: All common values of an attribute are taken as a group of candidate text options, and the group contains a number of expressions of values, such as "In this <domain> image, its <attribute> is <value 1>", "In this <domain> image, its <attribute> is <value 2>", ..., "In this <domain> image, its <attribute> is <value n>".

[0019] Optionally, in one embodiment of the present invention, the vision-language model has a model of separate visual and textual unimodal encoders.

[0020] Optionally, in one embodiment of the present invention, in step S2, calculating the presentation degree of each attribute value in each image using the visual-language model includes:

[0021] Use the image encoder and text encoder to process the image to be annotated and a group of candidate texts respectively, obtain the corresponding image embedding and a group of text embeddings, calculate the similarity between the image embedding and each candidate text embedding in the group, pass the similarity sequence through a Softmax layer, and obtain the distribution of the probability of each option in the group with a sum of 1. The probability value is the degree of presentation of the attribute value corresponding to the candidate text option in the image.

[0022] Optionally, in one embodiment of the present invention, in step S2, when converting the image-text retrieval output result into an annotation of the image, the true value distribution or the binary value distribution is retained;

[0023] If the true value distribution is retained: for the method of converting candidate text options without grouping, the probability value of the positive statement is saved; for the method of converting candidate text options by grouping, the complete probability distribution on the group is directly retained;

[0024] If the binary distribution is retained: for the method of converting candidate text options without grouping, if the probability of the positive statement is greater than the probability of the negative statement, save 1, otherwise save 0; for the method of converting candidate text options by grouping, set the attribute value with the highest probability to 1, and set the values of other attributes to 0.

[0025] The zero-shot image attribute labeling method based on a pre-trained large model in an embodiment of the present invention reduces the cost of data labeling. At the same time, compared with various existing weakly supervised labeling data and manually labeled data, the data labeled by the method of the present invention has significant performance improvement in zero-shot learning. By reducing labeling costs and improving performance, it has the potential to promote the application of zero-shot learning in various real-world scenarios.

[0026] Additional aspects and advantages of the present invention will be set forth in part in the description which follows and, in part, will be obvious from the description which follows, or may be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments in conjunction with the accompanying drawings, in which:

[0028] Figure 1 A flowchart of a zero-sample image attribute labeling method based on a pre-trained large model provided according to an embodiment of the present invention;

[0029] Figure 2 and Figure 3 A schematic diagram of annotating an image using a vision-language model according to an embodiment of the present invention. DETAILED DESCRIPTION

[0030] The following describes embodiments of the present invention in detail, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present invention, and are not to be construed as limiting the present invention.

[0031] Figure 1 The present invention provides a flowchart of a zero-sample image attribute labeling method based on a pre-trained large model according to an embodiment of the present invention.

[0032] like Figure 1 As shown in FIG, the zero-shot image attribute labeling method based on the pre-trained large model includes the following steps:

[0033] S1, obtains a pre-trained large language model. For each zero-shot learning dataset in a specific domain, through interaction with the pre-trained large language model, after preliminary definition of attributes, repeated screening and refinement, and definition of attribute values, a set of attributes and attribute values that are discriminative for category recognition in a specific domain is obtained.

[0034] Optionally, in one embodiment of the present invention, step S1 specifically includes:

[0035] S101: Obtain a pre-trained large language model and interact with it to preliminarily define discriminative attributes within a specific domain through question-and-answer format. Specifically, a single prompt interaction with the model asks which attributes contribute to category recognition in a specific zero-shot learning domain. The model will then, based on its knowledge, provide a set of attributes that contribute to category recognition within that domain, thereby providing a preliminary definition of attributes for category recognition within that domain.

[0036] For example, for the CUB dataset on birds, one only needs to give the prompt input "In the fine-grained recognition task of birds, which attributes are helpful for their classification?" The large language model will output a set of preliminary defined bird attribute information, including "feather patterns and colors, body shape and proportions, morphological characteristics, bird behavior and ecology, geographical and ecological distribution, temporal and seasonal changes, and genetic information" and other definitions.

[0037] For large language models, you can choose to use web service interfaces provided by various technology companies, such as OpenAI's ChatGPT API and Anthropic's Claude API. You can also choose to deploy models that have undergone pre-training, supervised training, and alignment with human preferences. The former requires fewer resources and technical difficulty, while the latter allows users more customization.

[0038] S102, based on the attributes preliminarily defined, further screening and refinement of the attributes. Alternately screen and eliminate visually invisible attributes and refine overly generalized attributes to obtain a set of valid attributes for zero-sample image classification in a specific domain. First, screen the preliminary attribute definitions provided by the large language model and eliminate visually invisible attributes. These visually invisible attributes cannot be directly identified from the image and are removed through a small amount of manual intervention. Secondly, the preliminary definitions output by the language model are often relatively broad, so they need to be further refined. For each broad attribute definition, continue to interact with the large language model as part of the prompt, and ask which refined attributes of the broad attribute in the target domain are helpful for image classification. The large language model will gradually provide more specific attribute definitions based on the feedback. Repeat the process of screening and refining attributes to eventually obtain a set of valid attribute definitions for image recognition in a specific domain.

[0039] For example, for the CUB dataset in the field of birds, the large language model initially defines the attributes "feather pattern and color, body shape and proportion, morphological characteristics, bird behavior and ecology, geographical and ecological distribution, time and seasonal changes, and genetic information". The following steps can be performed in sequence: (1) Screening. The time and seasonal changes in the above attributes are dynamic features, and genetic information is information that cannot be directly observed. Both are invisible information in the image, so they should be eliminated. (2) Refinement. The definitions of the remaining attributes are relatively broad and need to be refined. Taking "feather pattern and color" as an example, based on this attribute, it can be further suggested "In the fine-grained recognition task of birds, which feather pattern and color attributes are helpful for their classification?" The large language model will further provide "color contrast and color matching, feather pattern, special markings on wings and tail feathers, color and morphology of head feathers, seasonal changes, gender differences" and other refined feather pattern and color attributes. Repeating the above screening and refinement steps several times will obtain a set of attribute definitions with moderate specificity. Generally speaking, 1-3 refinements are sufficient, and the number of attributes obtained is generally dozens to hundreds. This method does not limit the number of refinements and attributes, and can be adjusted according to the actual situation of the labeled dataset.

[0040] S103: After acquiring the attributes, continue interacting with the large language model to obtain common values for each attribute. Based on the defined attributes, further interact with the large language model, using each specific attribute definition as part of the prompt to obtain common values for each attribute in a specific domain. Given that each attribute may have many values, and the value range of some attributes may be continuous, their values should be integrated and discretized. This method does not require any human intervention and directly utilizes the knowledge of the large model to adaptively acquire common values for each attribute.

[0041] For example, the attribute "wing color" can take any color. By asking the large language model "What are the most common bird feather colors?", we can obtain a moderate number of the most common bird feather colors. For the attribute "wing length," which has a continuous range of values, we can ask the large language model "What are the most common bird wing length intervals?" to obtain its knowledge of bird wing length ranges.

[0042] Optionally, in one embodiment of the present invention, the pre-trained large language model can be used in two ways: application programming interface (API) access or model access. You can interact with a specific pre-trained model through the network interface API provided by the technology company. This method does not require direct operation of the underlying hardware or computing resources of the model. You only need to call the model on the remote server through a simple network request, and the model will return the generated text or the result of performing a specific task. You can also choose to directly access the language model that has undergone three steps of pre-training, supervised training, and alignment with human preferences training, and directly load and run the model in a local or cloud environment. This access method usually requires directly loading and running the model in a local or cloud environment, and API accessibility allows users to have more control over the operation and tuning of the model.

[0043] Optionally, in one embodiment of the present invention, in step S102, the attribute definitions provided by the large language model are first screened, and those visually invisible, low-quality attributes that cannot be discerned from the image are eliminated through manual intervention; then, further interaction is performed with the language model to refine the previously defined, relatively broad attributes. For each broad attribute definition, the attribute to be refined is used as part of a prompt to continue interacting with the large language model to define which refined attributes of the broad attribute in the target domain are helpful for image classification. During the attribute screening and refinement process, these two steps are performed alternately, and the number of repetitions is flexibly adjusted according to the breadth of the attribute definition. Typically, one to three screening and refinement steps are sufficient to complete the process, ultimately obtaining a set of effective attribute definitions for image recognition in a specific domain.

[0044] Optionally, in one embodiment of the present invention, given that each attribute may have many values, and the value ranges of some attributes may be continuous, the knowledge of the large language model is further utilized to define a small number of common values for each attribute in a discretized form. In step S103, each defined attribute is used as part of the prompt separately, and further interacts with the large language model to obtain common values for each attribute in a specific field. When defining the value of each attribute, there is no limit on the number of values for each attribute, and the large language model independently gives a reasonable number and value division. The present invention makes full use of the advantages of the large language model, and the large language model independently gives a reasonable number of values and specific value definitions and divisions. When interacting with the large language model to obtain the value of each attribute, there is no limit on the number and range of values for each attribute, and the knowledge of the large model is used to adaptively obtain common values for each attribute.

[0045] S2, converts attributes and attribute values into groups of candidate text options, obtains a pre-trained visual-language model, uses the image-text retrieval capability of the visual-language model to infer the presentation degree of each attribute value in each image, and converts them into image annotations and category annotations.

[0046] Optionally, in one embodiment of the present invention, step S2 specifically includes the following steps:

[0047] S201: converting defined attributes and common values of each attribute into groups of candidate text options;

[0048] S202: Based on the candidate text sets, the pre-trained visual-language model's image-text retrieval capabilities are used to calculate the degree of presence of each attribute value in each image. Specifically, each set of candidate text sets is treated as an image-text retrieval task. The pre-trained visual-language model is used to perform image-text retrieval on each set of text for each image, thereby calculating the degree of presence of each attribute value in each image.

[0049] S203: Convert the image-text retrieval output result into image annotations, and then calculate the mean of the annotations of all images in each category to obtain the annotations of each category.

[0050] Optionally, in one embodiment of the present invention, in step S201, two methods are designed to convert attributes and their common values into groups of candidate text options. Both methods use template sentences to fill in the <domain> of the dataset to be annotated, the <attribute> that has been obtained, and the attribute <value>. (1) "Non-group conversion method": Each attribute value is converted into a group of text options separately, and the group contains two candidate items, namely, the affirmative text statement "In this <domain> image, its <attribute> is <value i>" and the negative text statement "In this <domain> image, its <attribute> is not <value i>". (2) "Group conversion method": All common values of an attribute are used as a group of candidate text options, and the group contains the number of value statements, such as "In this <domain> image, its <attribute> is <value 1>", "In this <domain> image, its <attribute> is <value 2>", ..., "In this <domain> image, its <attribute> is <value n>". The first method considers the degree of presentation of each attribute value in the image separately, which can retain more detailed information; while the second method refers to all common values of the attribute at the same time, taking into account the dependencies and mutual exclusions between attribute values.

[0051] For example, Figure 2 and Figure 3 The text area in the upper left corner shows the method of converting attributes and their values into candidate text options mentioned in step S201. Both figures show the conversion of the attribute body shape (body shape) and its five common values in the bird field. Figure 2 This is a non-grouping conversion method. Each attribute value is individually converted into a group of candidate texts, each group containing two statements: affirmative and negative. Figure 3 It is a group-based conversion method, where all common values of an attribute are converted into a group of candidate texts, and each common attribute value is a candidate item in the group.

[0052] Optionally, in one embodiment of the present invention, the pre-trained vision-language model selects a model with separate visual and text unimodal encoders. This architecture ensures that all images and texts only need to go through feature extraction once, ensuring efficient image-text retrieval performance.

[0053] Optionally, in one embodiment of the present invention, in step S202, when calculating the degree of presentation of attribute values in the image, each group of candidate text options corresponds to an image-text retrieval task, and first, an image encoder and a text encoder are used to process the image to be annotated and a group of candidate text options respectively to obtain corresponding image embeddings and a group of text embeddings. Subsequently, the similarity between the image embedding and each candidate text embedding in the group is calculated. The similarity is passed through a Softmax layer to obtain a distribution with a probability sum of 1 on each option in the group. Among them, the probability value can be understood as the degree of presentation of the attribute value corresponding to the candidate text option in the image. According to the similarity distribution, the degree of presentation of the attribute value corresponding to the candidate text option in the image is annotated.

[0054] Optionally, in one embodiment of the present invention, in the annotation of step S202, in order to ensure a small amount of video memory usage and avoid waste caused by repeated feature extraction, the following steps are adopted: first, all text options are encoded at one time, and then the text features are continuously stored in the video memory. The number of all candidate texts is generally dozens to hundreds, which will not cause a large video memory burden. Subsequently, the images in the data set are processed in small batches in sequence, and the extracted image features are interacted with all text options. The image annotation is directly obtained by extracting the image features at one time, and only the video memory required for a small batch is required each time.

[0055] To avoid repeated operations and reduce video memory usage, the order of processing images and text is generally as follows: (1) First, use the text encoder to extract the features of all candidate texts in sequence, and continuously save the text features in the video memory. (2) Process the images in the dataset in small batches, and the extracted image features will interact with all group text features. After extracting the image features once, the interaction with all group text options is completed. Among them, for zero-shot learning datasets, the number of all candidate texts is generally dozens to hundreds, which will not cause a large video memory burden. The entire annotation process only extracts the features of the image and text once, and no redundant calculations are generated. Figure 2 Figure 3 What is presented is the process of interactively annotating an image and a set of text options.

[0056] In step S202, the vision-language model can be any model with a unimodal encoder for vision and language. Figure 2 and Figure 3The pre-trained CLIP model with two unimodal encoders is used as the vision-language model. First, the image to be annotated and a set of candidate text options are processed using an image encoder and a text encoder, respectively, to obtain the corresponding image embedding and a set of text embeddings. A dot product is then performed to determine the similarity between the image embedding and each candidate text embedding within the set. This similarity is then passed through a Softmax layer to obtain a distribution with a probability sum of 1 across all options within the set. Based on this similarity distribution, the degree to which the attribute values corresponding to the candidate text options are present in the image is annotated.

[0057] Optionally, in one embodiment of the present invention, in step S203, when converting the image-text search results into annotations for the image, you can choose to retain the true value distribution or the binary distribution. (1) If the true value distribution is retained: for the method of converting candidate text options without grouping, save the probability value of its positive statement; for the method of converting candidate text options by grouping, directly retain its complete probability distribution on the group. (2) If the binary distribution is retained: for the method of converting candidate text options without grouping, if the probability of the positive statement is greater than the probability of the negative statement, save 1, otherwise save 0; for the method of converting candidate text options by grouping, set the attribute value with the highest probability to 1, and set the values of other attributes to 0. Among them, retaining the true value distribution is more susceptible to noise disturbance while retaining more complete detailed information, while the binary distribution is more in line with human's two-category understanding of attributes.

[0058] Table 1 is a comparison of the classification accuracy of the dataset attributes obtained based on the method provided by the above embodiment of the present invention and other semantic embedding annotations on the general zero-shot learning datasets AWA2, CUB, FLO and SUN. The table shows the performance in combination with four classic zero-shot learning methods. The semantic embedding annotations compared in the table include word2vec word embedding, weakly supervised semantic embedding annotations VGSE and I2DEmb designed for zero-shot learning training, and manually annotated attribute annotations. The attribute annotations provided by the present invention, when combined with any method, have surpassed the performance of various compared semantic embedding annotations on all datasets.

[0059] Table 1 Performance comparison of this embodiment with other commonly used semantic embedding annotations

[0060]

[0061] According to the zero-shot image attribute labeling method based on a pre-trained large model proposed in an embodiment of the present invention, a pre-trained large language model is obtained and interacted with, and through the steps of preliminary attribute definition, repeated attribute screening and refinement, and attribute definition and value selection, a set of attribute values that are discriminative for category recognition within a specific domain is obtained; the defined attributes and their common values are converted into groups of candidate text options, and a visual-language model is obtained. Its image-text retrieval capability is used to annotate the degree of presence of each attribute value in each image, and the image annotation is converted into category annotation. Based on the capabilities of the pre-trained model, the present invention proposes a method for designing and labeling attributes for zero-shot learning, reducing the dependence of zero-shot learning training methods on manually labeled data.

[0062] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or N embodiments or examples in a suitable manner. In addition, those skilled in the art can combine and combine different embodiments or examples described in this specification and the features of different embodiments or examples without contradiction.

[0063] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be understood to indicate or imply relative importance or implicitly specify the number of technical features indicated. Thus, a feature specified as "first" or "second" may explicitly or implicitly include at least one such feature. In the description of the present invention, "N" means at least two, such as two, three, etc., unless otherwise specifically defined.

[0064] Any process or method description in a flowchart or otherwise described herein may be understood to represent a module, segment or portion of code comprising one or N executable instructions for implementing a custom logical function or step of a process, and the scope of the preferred embodiments of the present invention includes alternative implementations in which functions may be performed out of the order shown or discussed, including performing functions in a substantially simultaneous manner or in the reverse order depending on the functions involved, which should be understood by those skilled in the art to which the embodiments of the present invention pertain.

Claims

1. A zero-shot image attribute labeling method based on a pre-trained large model, characterized in that: The following steps are involved: S1, obtaining a pre-trained large language model, and obtaining a set of attributes and attribute values for image classification in a specific domain by interacting with the large language model; S2: Convert the attributes and attribute values into groups of candidate text options, obtain a pre-trained visual-language model, use the visual-language model to infer the presentation degree of each attribute value in each image, and convert them into image annotations and category annotations.

2. The method according to claim 1, characterized in that Step S1 specifically includes: S101, interacting with the large language model to preliminarily define attributes in a specific domain through question-answering; S102, after alternately screening and refining attributes, a set of effective attributes for zero-shot image classification in a specific domain is obtained; S103: Continue interacting with the large language model to obtain common values for each attribute.

3. The method according to claim 1, characterized in that In step S1, interaction with the large language model is performed through application programming interface access or model access.

4. The method according to claim 2, characterized in that In step S102, attributes that do not meet the requirements are eliminated through repeated alternating screening and refinement of attributes.

5. The method according to claim 1, wherein The step S2 further includes: The image-text retrieval capability of the visual-language model is used to calculate the degree of presentation of each attribute value in each image, and the image-text retrieval output is converted into image annotations. The annotations of all images in each category are then averaged to obtain the annotations for each category.

6. The method according to claim 5, characterized in that In step S2, two template sentence patterns are used to convert the attributes and attribute values into a group of candidate text options, and the <field> of the dataset to be annotated, the obtained <attribute> and the attribute <value> are filled in. (1) Non-grouping transformation method: Each attribute value is individually transformed into a set of text options, which contains two candidate options, namely the affirmative text statement "In this <domain> image, its <attribute> is <value i>" and the negative text statement "In this <domain> image, its <attribute> is not <value i>"; (2) Grouping conversion method: All common values of an attribute are taken as a group of candidate text options, and the group contains a number of expressions of values, such as "In this <domain> image, its <attribute> is <value 1>", "In this <domain> image, its <attribute> is <value 2>", ..., "In this <domain> image, its <attribute> is <value n>".

7. The method according to claim 1, characterized in that The vision-language model is a model with separate visual and textual unimodal encoders.

8. The method according to claim 1, characterized in that In step S2, calculating the presentation degree of each attribute value in each image using the visual-language model includes: Use the image encoder and text encoder to process the image to be annotated and a group of candidate texts respectively, obtain the corresponding image embedding and a group of text embeddings, calculate the similarity between the image embedding and each candidate text embedding in the group, pass the similarity sequence through a Softmax layer, and obtain the distribution of the probability of each option in the group with a sum of 1. The probability value is the degree of presentation of the attribute value corresponding to the candidate text option in the image.

9. The method according to claim 6, characterized in that In step S2, when converting the image-text retrieval output results into image annotations, the true value distribution or binary value distribution is retained; If the true value distribution is retained: for the method of converting candidate text options without grouping, the probability value of the positive statement is saved; for the method of converting candidate text options by grouping, the complete probability distribution on the group is directly retained; If the binary distribution is retained: for the method of converting candidate text options without grouping, if the probability of the positive statement is greater than the probability of the negative statement, save 1, otherwise save 0; for the method of converting candidate text options by grouping, set the attribute value with the highest probability to 1, and set the values of other attributes to 0.

Images