Integrated circuit layout design announcement information data processing and indexing method

Abstract

The invention discloses an integrated circuit layout design announcement information data processing and indexing method, which relates to the technical field of integrated circuit layout design announcement and comprises the following steps: acquiring integrated circuit layout design announcement information data; constructing a data model composed of a plurality of sub-data models, wherein the sub-data models are in one-to-one correspondence with announcement types in the integrated circuit layout design announcement information data; analyzing the integrated circuit layout design announcement information data to obtain analysis results in a dictionary list form, and respectively storing the analysis results as sub-announcement information data according to announcement types; processing the sub-announcement information data to obtain structured announcement information data, and storing the structured announcement information data into corresponding sub-data models; and establishing an index for each data model. According to the method, the original disordered unstructured announcement information can be converted into a structured dictionary list form, so that the data processing speed and efficiency are greatly improved, and the data can be accurately classified according to the announcement type.

Description

Technical Field

[0001] The present invention relates to the technical field of integrated circuit layout design announcements, and in particular to a method for processing and indexing integrated circuit layout design announcement information data. Background Art

[0002] Integrated circuit layout design refers to the three-dimensional configuration of two or more components and some or all interconnection lines in an integrated circuit, at least one of which is an active component, or the above three-dimensional configuration prepared for the manufacture of integrated circuits; after the exclusive right of integrated circuit layout design is registered, the relevant information will be announced in the relevant bulletin for public reference. Multiple records are merged and published in each announcement, and various information is mixed and unstructured. If the public needs to query or count, they need to check the content of each announcement one by one, which is inefficient and extremely inconvenient.

[0003] The information on the announcement of the exclusive rights to integrated circuit layout designs can be classified and summarized by manually sorting them one by one, but the efficiency is extremely low. There are already exclusive rights announcement query and retrieval platforms on the market, which cannot be associated and updated with subsequent data of other types of announcement information. There are problems such as low query result accuracy and delayed data update, which cannot meet user needs.

[0004] Other types, such as changes, transfers, corrections, rejections, terminations, etc., are published together, resulting in chaos and irregularity. Traditional data processing methods cannot quickly and effectively achieve data classification and aggregation, and it is difficult to meet indexing requirements. Different types of information cannot achieve data association and synchronous update. Therefore, there is an urgent need for a fast, efficient, and accurate integrated circuit layout design announcement information data processing method to achieve data classification and aggregation to meet user query and statistical needs. Summary of the invention

[0005] The present invention provides a method for processing and indexing integrated circuit layout design announcement information data, which can solve the above-mentioned problem.

[0006] In order to solve the above problems, the technical solution adopted by the present invention is as follows:

[0007] The present invention provides a method for processing and indexing integrated circuit layout design announcement information data, comprising the following steps:

[0008] S1. Obtaining integrated circuit layout design announcement information data;

[0009] S2. constructing a data model composed of a plurality of sub-data models according to the integrated circuit layout-design announcement information data, wherein the sub-data models correspond one to one to the announcement types in the integrated circuit layout-design announcement information data;

[0010] S3, using a data analysis module to analyze the integrated circuit layout design announcement information data of different types, obtain analysis results in the form of a dictionary list, and store each analysis result as sub-announcement information data according to the announcement type;

[0011] S4. Process the sub-announcement information data using a data processing module to obtain structured announcement information data including the announcement registration / application number of the integrated circuit layout design, and store the structured announcement information data into a corresponding sub-data model;

[0012] S5. Create an index for each data model to query integrated circuit layout design announcement information data from the data model.

[0013] As a further description of the above technical solution, in step S1, a txt format text file is used to store the integrated circuit layout design announcement information data, and the status of the txt format text file is updated through a script monitoring interface. When the txt format text file is updated, it needs to be downloaded locally.

[0014] As a further description of the above technical solution, the original integrated circuit layout design announcement information data obtained is a compressed package of a zip format file, and a script program is needed to decompress the zip format file to obtain a txt format text file.

[0015] As a further description of the above technical solution, a scheduled task is set to access the web page where the integrated circuit layout design announcement information data is located, and download the integrated circuit layout design announcement information data.

[0016] As a further description of the above technical solution, in step S2, the announcement types include proprietary rights announcement, transaction announcement, transfer, change, voluntary abandonment, letter return, termination and correction.

[0017] As a further description of the above technical solution, in step S3, the data parsing module parses the integrated circuit layout design announcement information data based on the Python script, performs automatic data classification extraction and processing through regular expression recognition or string segmentation inspection, and organizes the data into a dictionary list.

[0018] As a further description of the above technical solution, in step S4, before storing the structured data into the sub-data model, the data is first cleaned, the information structure is unified, and the update or modification time is recorded.

[0019] As a further description of the above technical solution, after the integrated circuit layout design announcement information data is queried according to the index constructed in step S5, a data analysis module is used to perform a visual analysis on it.

[0020] As a further description of the above technical solution, the data processing module also includes a data preprocessing module and an intelligent data processing model constructed based on a convolutional neural network. The data preprocessing module is used to clean the original text data, remove HTML tags, special characters and redundant blank information, use Chinese word segmentation tools to segment the text, and establish a custom dictionary to ensure accurate word segmentation; the intelligent data processing model is used to automatically identify the announcement type based on the preprocessed data and parse the text data under different types of announcements.

[0021] As a further description of the above technical solution, the intelligent data processing model includes a text classifier for announcement type classification and a sequence tagger for key information extraction; the text classifier and sequence tagger share an embedding layer and a convolution layer, and the generalization ability of the model is improved through multi-task learning.

[0022] Compared with the prior art, the present invention has the following beneficial effects:

[0023] 1) Create a data model based on the announcement type. The data analysis module parses the original unstructured data and stores it as a separate file according to the announcement type. The sub-data processing module further processes the corresponding announcement type file to obtain structured data. After cleaning, it is stored in the sub-data model and imported into the database. Compared with the traditional method, a faster, more efficient and accurate processing method is achieved. Data processing can be automatically completed using automated scripts, which reduces manual operations and reduces labor costs.

[0024] 2) After the data processing is completed, the data of different announcement types are associated in the database, including but not limited to layout-design exclusive rights announcements, business announcements, transfers, changes, voluntary abandonment, letter rejection, termination, correction and other information, or combined with Web applications or other programs to provide a visual interface to realize the service of retrieving integrated circuit layout-design announcement information, and can also realize data classification and aggregation to meet the user's query and statistical needs.

[0025] 3) Convert unstructured announcement information into structured data to facilitate deeper data analysis and processing. By recording data update time, it ensures that the information queried by users is the latest, thereby improving the timeliness of information.

[0026] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the embodiments of the present invention are specifically cited below and described in detail with reference to the attached drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments are briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without creative work.

[0028] Figure 1 is a flow chart of the integrated circuit layout design announcement information data processing and indexing method according to an embodiment of the present invention;

[0029] Figure 2 It is a sub-data model of the announcement type being proprietary announcement;

[0030] Figure 3 It is a sub-data model whose announcement type is transaction announcement;

[0031] Figure 4 Schematic diagram of the analysis process of the data analysis module in the embodiment;

[0032] Figure 5 Schematic diagram of the index structure in the embodiment. DETAILED DESCRIPTION

[0033] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments.

[0034] Example 1

[0035] Please refer to Figure 1 The embodiment of the present invention provides a method for processing and indexing integrated circuit layout design announcement information data, comprising the following steps:

[0036] S1. Obtain integrated circuit layout design announcement information data.

[0037] In this embodiment, the official open interface is used to automatically download files related to the integrated circuit layout design announcement information, and the announcement information is stored in txt format text files. These files will be updated through the script monitoring interface status, and once updated, they will be downloaded to the local processing immediately to ensure the timeliness and accuracy of the data.

[0038] It should be noted that the files downloaded through the official open interface are compressed packages in zip format. You need to use a script program to decompress the zip file first, and then save the TXT file obtained after decompression to the specified location.

[0039] The content of each integrated circuit layout design announcement is displayed on a web page. This embodiment collects and stores it regularly. The web page is accessed through a scheduled task, and the web page update status is monitored through a script to automatically collect integrated circuit announcement information. The content of each announcement is summarized on a specific web page, and the web page content is downloaded and saved as a TXT file. The format of the announcement information stored in the open interface is consistent to facilitate subsequent unified processing.

[0040] S2. Based on the integrated circuit layout-design announcement information data, a data model consisting of a number of sub-data models is constructed, and the sub-data models correspond one by one to the announcement types in the integrated circuit layout-design announcement information data.

[0041] In this embodiment, the announcement types include exclusive rights announcement, transaction announcement, transfer, change, voluntary abandonment, letter return, termination and correction. Figure 2 and Figure 3 They are sub-data models for announcement types of proprietary announcement and transaction announcement respectively.

[0042] Each sub-data model corresponds to a type of announcement, and the sub-data model should include the integrated circuit layout design announcement registration / application number. Such a model design helps to process and query data in a targeted manner, thereby improving system efficiency and data availability.

[0043] S3. Combination Figure 4 As shown, the data parsing module is used to parse the integrated circuit layout design announcement information data of different types to obtain the parsing results in the form of a dictionary list, and each parsing result is stored as sub-announcement information data according to the announcement type.

[0044] In this embodiment, the data parsing module parses the sub-data model based on the Python script, such as the information of "Announcement of Exclusive Rights to Layout-Design", "Changes in Catalog Items" and "Transfer of Exclusive Rights", and performs automatic data classification extraction and processing through regular expression recognition or string segmentation (split) inspection, organizes the information into a dictionary list, and stores the results in CSV files according to the announcement type.

[0045] The parsing method specifically includes: parsing each TXT file, for example, when extracting information about "layout-design exclusive rights announcement", "changes in bibliographic items" and "transfer of exclusive rights", the data is processed separately according to the identified keywords, and each type of information is saved in a separate list. Each line of text is checked to extract data based on whether it contains specific identifiers (such as "layout registration number" and "announcement date"). At the same time, the value of each field is extracted using a delimiter. At the end of the file or when a new keyword is encountered, the currently accumulated data is saved in the corresponding list and a new data record is started. The extracted data is assembled into a Pandas DataFrame, and then stored in CSV files according to the data announcement type (such as announcements of "layout-design exclusive rights announcement", "changes in bibliographic items" and "transfer of exclusive rights").

[0046] It should be noted that:

[0047] In this embodiment, CSV files are used to store relevant information for the convenience of illustration, but this is not the only option. People skilled in the art should know that alternative file formats can also be JSON, XML, Excel, etc. PandasDataFrame is a very powerful data structure, mainly used for data analysis and processing. It is one of the core functions of the Pandas library in the Python programming language.

[0048] S4. Process the sub-announcement information data using a data processing module to obtain structured announcement information data including the announcement registration / application number of the integrated circuit layout design, and store the structured announcement information data into a corresponding sub-data model.

[0049] In this embodiment, the data processing module includes a plurality of sub-data processing modules, each corresponding to each sub-data model.

[0050] The CSV file obtained by the data parsing module is imported into the corresponding sub-data processing module. The sub-data processing module parses the CSV file and uses regular expressions to extract information from the text of each record, such as layout-design registration number, right holder, application date, layout-design name, etc. The extracted information is organized into a dictionary list and saved in a new CSV file. After obtaining the structured data, it is stored in the corresponding sub-data model.

[0051] Furthermore, before storing the structured data in the sub-data model, it is first cleaned to unify the information structure, prevent duplicate data records, and record the time of update or modification. In addition, the announcement information data obtained from different channels can be compared to verify the integrity of the information.

[0052] S5, such as Figure 5As shown, an index is established for each data model to query integrated circuit layout design announcement information data from the data model.

[0053] Choose a suitable data storage system to manage CSV file data. For large-scale data, commonly used relational database management systems (RDBMS) such as MySQL, PostgreSQL, or NoSQL databases such as MongoDB, or even full-text search engines such as Elasticsearch. The present invention prefers MySQL as the data storage system and establishes an index as follows:

[0054] Design MySQL data tables according to the sub-data model fields, and import the data in the CSV file into the corresponding table in the MySQL database. Most databases provide tools or commands to import data from CSV files. For example, in MySQL, you can use the LOAD DATA INFILE command to import data.

[0055] In the database, create indexes to speed up query operations, and use all fields or commonly used fields to establish indexes, such as "layout-design registration number, layout-design name, right holder name" and other fields.

[0056] In order to achieve the integrity of multiple announcement information queries, a multi-table query method can be used. Knowing that the field "layout design registration number" is a unique value, you can use the "layout design registration number" to connect multiple data tables so that data from multiple tables can be combined and compared to achieve multi-table query. Another example is to search in different data tables when constructing a query command in the application. There are many ways to achieve multi-table query, which will not be repeated here.

[0057] After the database is set up, create a web application to provide users with a visual retrieval window.

[0058] In addition to the integrated circuit layout design announcement information data processing and indexing method described above, a data analysis module can also be added. After querying the integrated circuit layout design announcement information data according to the index constructed in step S5, the data analysis module integrates existing data resources, applies modern analysis technology based on structured data, provides intuitive analysis results and a visual interface, and provides data analysis support for users.

[0059] Example 2

[0060] In addition to the above-mentioned embodiment 1, AI or machine learning can also be used, such as based on convolutional neural network (CNN) to realize integrated circuit layout design announcement information data processing. CNN has a strong feature extraction capability in text data processing, and automatically parses, classifies, structures and indexes the integrated circuit layout design announcement information. Through deep learning technology, the accuracy and efficiency of data processing can also be improved to meet the needs of large-scale data processing and real-time query.

[0061] The solution of this embodiment is that: based on the convolutional neural network, it can automatically identify the type of announcement, automatically parse the content of the announcement, and then extract the key information in the announcement according to the type of announcement to generate structured key-value pair data.

[0062] In this embodiment, the data is first preprocessed, and the data text is cleaned by writing a script program to remove HTML tags, special characters and redundant blank information, and then the text is segmented using a Chinese word segmentation tool (such as Jieba). The data obtained by S1 is unstructured, containing raw text data of different types of announcements, and it is necessary to perform preliminary cleaning and sorting on it, removing blank lines, redundant spaces and unified symbols (full-width replaced with half-width), etc., to facilitate the processing of subsequent steps; the data obtained by S1 contains professional terms and words in a specific format, and it is necessary to establish a custom dictionary to ensure accurate segmentation, such as: 'proprietary rights announcement', 'layout design registration number', 'layout design application date', 'announcement date', 'announcement number', 'layout design name', 'layout design category', 'structure', 'technology', 'function', 'layout design right holder', 'layout design right holder nationality', 'layout design right holder address', 'layout design creator', 'layout design creation completion date'.

[0063] Before participle: [Layout design registration number: BS.123456]

[0064] Example after word segmentation: ['Layout design registration number',':','BS','123456']

[0065] Secondly, based on the data obtained from the aforementioned data preprocessing, a CNN model is designed and trained to automatically identify different types of announcements (such as proprietary announcements, change announcements, transfer announcements, etc.), parse the text data under different types of announcements, extract key information from the announcements, and generate structured key-value pair data. The specific implementation process is as follows:

[0066] The overall architecture of the model consists of two main parts: a text classifier for classifying announcement types and a sequence tagger for extracting key information. The two parts share an embedding layer and a convolutional layer, which improves the generalization ability of the model through multi-task learning.

[0067] (1) Design CNN model

[0068] (1.1) Input layer

[0069] In the neural network model, the main task of the input layer is to convert the original text data into a numerical form that the model can process. For text information, word embedding is used to map each word in the text into a dense low-dimensional vector to capture the semantic and grammatical information of the word.

[0070] Since the integrated circuit layout announcement information is text information, word embedding is used to map each word in the text into a dense low-dimensional vector to capture the semantic information of the word. According to the announcement type and based on the word segmentation results, a vocabulary is established to map each word to a unique index ID. For example:

[0071] 'Proprietary Rights Notice': 0

[0072] 'Layout design': 1

[0073] 'Accession number': 2

[0074] 'BS.123:3

[0075] …

[0076] Define unique label names for fields of different announcement types to avoid processing multiple sets of labels in the model. Use word embedding to convert the word index sequence into a word vector sequence, such as Word2Ve c GloV e You can use pre-trained word vectors, or train word vectors on your own defined corpus.

[0077] (1.2) Convolutional layer

[0078] Through convolution operations, local features are extracted from word embedding sequences.

[0079] Convolution kernel r )

[0080] Size: Set to different window sizes, such as 3, 4, and 5, to capture n-gram features of different lengths. Quantity: Multiple convolution kernels of each size can be set, such as 100, to extract rich features.

[0081] Convolution operation input

[0082] One-dimensional convolution (1D Convolution) performs one-dimensional convolution on the embedded sequence, and the convolution kernel slides on the text sequence. For the input sequence, the formula is as follows:

[0083] X∈R L×d

[0084] Where L is the length of the text sequence and d is the word vector dimension.

[0085] The output formula of the convolution operation is as follows:

[0086] C i =f(w·x i：i+k-1 +b)

[0087] Where w is the convolution kernel weight, b is the bias; x i：i+k-1 represents the word vector subsequence from position i to i+k-1, w·x i：i+k-1 Represents the dot product operation between the weight of the convolution kernel and the subsequence, and f is the activation function (such as ReLU).

[0088] When the convolution kernel slides on the sequence, the convolution result calculated at each position i. The convolution kernel will slide step by step on the input sequence and calculate C for each window i , thus obtaining the output sequence.

[0089] (1.3) Activation function

[0090] ReLU (Rectified Linear Unit), increases the nonlinear expression ability of the model, is defined as f(x) = max(0, x), and alleviates the gradient vanishing problem.

[0091] (1.4) Pooling layer

[0092] It is to reduce the dimension of the output of the convolution layer, extract important features, and enhance the generalization ability of the model. Use global maximum pooling to take the maximum value of the features of the entire sequence; the pooling window is usually set to the length of the convolution output sequence to achieve global pooling. The following is the public notice:

[0093]

[0094] in, is the output after pooling, L′ is the sequence length of the convolution output, max{c1, c2, ..., c L′} means taking the maximum value of all convolution outputs.

[0095] Compress multiple feature vectors of a sequence into a single feature value, retaining the maximum activation value in the entire sequence to extract the most significant features.

[0096] (1.5) Fully connected layer

[0097] The output of the pooling layer is mapped to the label space to achieve the classification of announcement types. The specific implementation is as follows: input, feature vector output by the pooling layer; output, probability distribution of categories; activation function, using ReLU.

[0098] (1.6) Output layer

[0099] In the task of implementing the classification of announcement types for integrated circuit layouts, the Softmax layer converts the output of the fully connected layer into a probability distribution to achieve the classification of announcement types. The formula is as follows:

[0100] Assuming the number of announcement types is C, the formula for the Softmax output layer is:

[0101]

[0102] Where P(y=j|x) represents the probability that the announcement belongs to type j given input x, and z j is the score of the jth category output by the fully connected layer, C is the total number of announcement types, and exp(z j ) for the score z j Perform exponential operation so that each output value is positive, is the exponential sum of all categories and is used for normalization so that the sum of the probabilities of all categories is 1.

[0103] Through this formula, the Softmax layer converts the output of the fully connected layer into a probability distribution, so that we can determine which category the announcement belongs to based on the highest probability P(y=j|x).

[0104] (2) Training CNN model

[0105] (2.1) Cross Entropy Loss Function (Cross-Entropy Loss)

[0106] The cross entropy loss calculates the difference between the predicted probability and the true label. The smaller the loss, the closer the model's prediction is to the true label. This formula sums over all categories and takes a negative sign so that the goal is to minimize the loss. The formula is as follows:

[0107]

[0108] Among them, L represents the total loss value, C is the total number of categories, and y i is the indicator value of the actual label, (P(y i |x) is the probability predicted by the model, calculated by the Softmax layer, log(P(y i|x)) is the logarithm of the probability.

[0109] (2.2) Optimizer

[0110] Using Adam optimizer, adaptive learning rate optimization algorithm, with faster convergence speed

[0111] (2.3) Hyperparameter setting

[0112] Learning rate, the initial value is generally set to 0.001, which can be adjusted according to the training effect; batch size, such as 64 or 128, is selected according to the GPU memory size; the number of training rounds (Epochs), is generally between 10-20 rounds, determined in combination with the performance of the validation set.

[0113] (2.4) Regularization and prevention of overfitting

[0114] Dropout, add a Dropout layer after the fully connected layer to randomly discard some neurons to prevent overfitting; L2 regularization, add a weight penalty term to the loss function.

[0115] The data processing method is to extract structured key information from different types of integrated circuit layout design announcement information data. The data processing method can also be constructed based on the above CNN model to obtain structured data by training the model. Use CNN to organize the identified entity information into a dictionary or JSON format, for example: {"application number":"123456","applicant":"a company","publication date":"2021-01-01"}.

[0116] According to the announcement type, the parsing results are stored in the corresponding sub-announcement information data model to establish a clear data table structure.

[0117] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for processing and indexing integrated circuit layout design announcement information data, characterized in that: The following steps are involved: S1. Obtaining integrated circuit layout design announcement information data; S2. constructing a data model composed of a plurality of sub-data models according to the integrated circuit layout-design announcement information data, wherein the sub-data models correspond one to one to the announcement types in the integrated circuit layout-design announcement information data; S3, using a data analysis module to analyze the integrated circuit layout design announcement information data of different types, obtain analysis results in the form of a dictionary list, and store each analysis result as sub-announcement information data according to the announcement type; S4. Process the sub-announcement information data using a data processing module to obtain structured announcement information data including the announcement registration / application number of the integrated circuit layout design, and store the structured announcement information data into a corresponding sub-data model; S5. Create an index for each data model to query integrated circuit layout design announcement information data from the data model.

2. The integrated circuit layout design announcement information data processing and indexing method according to claim 1, characterized in that: In step S1, a txt format text file is used to store integrated circuit layout design announcement information data, and the status of the txt format text file is updated through a script monitoring interface. When the txt format text file is updated, it needs to be downloaded locally.

3. The integrated circuit layout design announcement information data processing and indexing method according to claim 2, characterized in that: The original integrated circuit layout design announcement information data obtained is a compressed package of a zip format file. It is necessary to use a script program to decompress the zip format file to obtain a txt format text file.

4. The integrated circuit layout design announcement information data processing and indexing method according to claim 2, characterized in that: Set a scheduled task to access the web page where the integrated circuit layout design announcement information data is located, and download the integrated circuit layout design announcement information data.

5. The integrated circuit layout design announcement information data processing and indexing method according to claim 1, characterized in that: In step S2, the announcement types include proprietary announcement, transaction announcement, transfer, change, voluntary abandonment, letter return, termination and correction.

6. The integrated circuit layout design announcement information data processing and indexing method according to claim 1, characterized in that: In step S3, the data analysis module analyzes the integrated circuit layout design announcement information data based on the Python script, performs automatic data classification extraction and processing through regular expression recognition or string segmentation inspection, and organizes the data into a dictionary list.

7. The integrated circuit layout design announcement information data processing and indexing method according to claim 1, characterized in that: In step S4, before storing the structured data into the sub-data model, the data is cleaned, the information structure is unified, and the update or modification time is recorded.

8. The integrated circuit layout design announcement information data processing and indexing method according to claim 1, characterized in that: After the integrated circuit layout design announcement information data is queried according to the index constructed in step S5, a data analysis module is used to perform a visual analysis on it.

9. The integrated circuit layout design announcement information data processing and indexing method according to claim 1, characterized in that: The data processing module also includes a data preprocessing module and an intelligent data processing model built based on a convolutional neural network. The data preprocessing module is used to clean the original text data, remove HTML tags, special characters and redundant blank information, use a Chinese word segmentation tool to segment the text, and establish a custom dictionary to ensure accurate word segmentation; the intelligent data processing model is used to automatically identify the announcement type based on the preprocessed data and parse the text data under different types of announcements.

10. The integrated circuit layout design announcement information data processing and indexing method according to claim 9, characterized in that: The intelligent data processing model includes a text classifier for classifying announcement types and a sequence tagger for extracting key information; the text classifier and the sequence tagger share an embedding layer and a convolution layer, and the generalization ability of the model is improved through multi-task learning.

Images