A method and apparatus for generating text for structured numeric data

Abstract

The application provides a method for generating text for structured numerical data, comprising: obtaining structured numerical data, which is a numerical sequence related to semiconductor manufacturing; inputting the numerical sequence into a trained natural language generation model to obtain a description text corresponding to the numerical sequence, the description text describing abnormal information corresponding to the numerical sequence. The method for generating text for structured numerical data provided by the application converts structured numerical data into text data which is easier to recognize, and better reveals abnormal phenomena in the semiconductor manufacturing process.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of semiconductor manufacturing, and in particular to a method and device for generating text from structured numerical data. BACKGROUND

[0002] Data generated by various machines in the semiconductor manufacturing process is very important for the production and manufacturing of semiconductors. For example, when a wafer has a yield problem, historical data generated by various machines can be analyzed to determine the defect and / or failure characteristics of the wafer, and the root cause corresponding to the defect and / or failure characteristics, and then a solution to improve the yield can be found. However, the data generated by the machines is structured numerical data, which requires high professional knowledge to analyze the structured numerical data, and further experience to determine abnormal information disclosed by the structured numerical data generated by the machines, which undoubtedly increases the workload of the staff and the professional threshold. SUMMARY

[0003] Embodiments of the present application provide a method and device for generating text from structured numerical data, which converts structured numerical data into text data that is easier to identify, and better reveals abnormal phenomena in the semiconductor manufacturing process.

[0004] In a first aspect, the present application provides a method for generating text from structured numerical data, comprising: obtaining structured numerical data, which is a numerical sequence related to semiconductor manufacturing; inputting the numerical sequence into a trained natural language generation model to obtain a description text corresponding to the numerical sequence, the description text describing abnormal information corresponding to the numerical sequence.

[0005] The method for generating text from structured numerical data provided by the present application converts structured numerical data into text data that is easier to identify, and better reveals abnormal phenomena in the semiconductor manufacturing process.

[0006] In one possible implementation, the architecture of the natural language generation model is a Seq2Seq architecture.

[0007] In one possible implementation, the natural language generation model includes an embedding representation layer, a Transformer-based language generation backbone model, and a pointer generation network.

[0008] The inputting of the numerical sequence into the trained natural language generation model to obtain the description text corresponding to the numerical sequence comprises:

[0009] The numerical sequence is inputted into the embedding representation layer to output an embedding vector of the numerical sequence.

[0010] inputting the embedding vector of the numerical sequence into the Transformer-based language generation backbone model, outputting a vocabulary probability distribution and an attention distribution, the attention distribution representing a position probability distribution of each numerical value in the numerical sequence, that is, an interactive attention feature between a context vector of each position of the numerical sequence and text to be generated at a current time step;

[0011] inputting the vocabulary probability distribution and the attention distribution into the pointer generation network, and outputting the description text corresponding to the numerical sequence.

[0012] In another possible implementation, the Transformer-based language generation backbone model comprises an encoder and a decoder.

[0013] An output end of the embedding representation layer is connected to an input end of the encoder, and an output end of the decoder is connected to an input end of the pointer generation network.

[0014] The embedding representation layer is a BiLSTM-based time series numerical embedding model.

[0015] In another possible implementation, the natural language generation model is trained based on a training set, and the training set comprises a plurality of training sample pairs, and each training sample pair comprises a numerical sequence and a description text corresponding to the numerical sequence.

[0016] In another possible implementation, the description text corresponding to the numerical sequence is determined based on an abnormal data segment of the numerical sequence, a rule template, and field information corresponding to the abnormal data segment.

[0017] The rule template is used to match a corresponding text description for the abnormal data segment, and the field information corresponding to the abnormal data segment comprises one or more of time information, wafer batch information, wafer number information, and equipment information corresponding to the abnormal data segment.

[0018] In one example, before the numerical sequence is input into the trained natural language generation model to obtain the description text corresponding to the numerical sequence, the method further comprises:

[0019] preprocessing the numerical sequence.

[0020] In another possible implementation, the preprocessing of the numerical sequence comprises one or more of the following:

[0021] performing cleaning processing on the numerical sequence to remove numerical values greater than or equal to a first preset threshold and numerical values less than or equal to a second preset threshold, wherein the first preset threshold is greater than the second preset threshold.

[0022] The numerical sequence after the cleaning treatment is normalized.

[0023] In a second aspect, the present application provides a device for generating text for structured numerical data, comprising:

[0024] The acquisition module is configured to acquire structured numerical data, the structured numerical data being a numerical sequence related to semiconductor circle manufacturing;

[0025] The language generation module is configured to input the numerical sequence into a trained natural language generation model to obtain a description text corresponding to the numerical sequence, the description text describing abnormal information corresponding to the numerical sequence.

[0026] In a third aspect, the present application provides a computer readable storage medium having a computer program stored thereon, when the computer program is executed in a computer, the computer program causes the computer to execute the method of the first aspect and / or the second aspect.

[0027] In a fourth aspect, the present application further provides a computing device comprising a memory and a processor, the memory having instructions stored therein, when the instructions are executed by the processor, the method of the first aspect and / or the second aspect is implemented.

[0028] In a fifth aspect, the present application provides a computer program or computer program product, the computer program or computer program product comprising instructions, when the instructions are executed, the computer program or computer program product causes the computer to execute the method of the first aspect and / or the second aspect. BRIEF DESCRIPTION OF DRAWINGS

[0029] Figure 1 An application scenario diagram of a method for generating text for structured numerical data provided by an embodiment of the present application is shown;

[0030] Figure 2 A flowchart of a method for generating text for structured numerical data provided by an embodiment of the present application is shown;

[0031] Figure 3 The inference process and the training process of the Data-To-Text model are shown in the schematic diagram;

[0032] Figure 4 A schematic diagram of a preset rule template is shown;

[0033] Figure 5 A structural diagram of the Data-To-Text model is shown;

[0034] Figure 6 A structural diagram of a time series numerical embedding representation layer based on BiLSTM is shown;

[0035] Figure 7 A schematic diagram of the principle of the Masked MultiHead Attention sub-module Mask in the Decoder Block;

[0036] Figure 8 Another flowchart for generating text for structured numerical data provided by the embodiment of the application;

[0037] Figure 9 A structural schematic diagram of a device for generating text for structured numerical data provided by the embodiment of the application;

[0038] Figure 10 A structural schematic diagram of a computing device provided by the embodiment of the application. DETAILED DESCRIPTION

[0039] The technical solutions of the application will be further described in detail below with the aid of the accompanying drawings and embodiments.

[0040] As mentioned earlier, since the data generated by the semiconductor manufacturing equipment is structured numerical data, rather than text data that users are more likely to recognize, professional personnel need to analyze the structured numerical data to find the information contained therein, and then find the abnormal situations existing in the semiconductor manufacturing process, which undoubtedly increases the workload of the staff and the professional threshold.

[0041] Therefore, the application provides a method for generating text for structured numerical data, which inputs the structured numerical data generated by the semiconductor manufacturing equipment into a trained natural language generation model, and then obtains the corresponding description text, so as to better disclose the abnormal situations existing in the semiconductor manufacturing process and better assist the semiconductor yield analysis engineers in yield analysis.

[0042] Wafer manufacturing is a typical scenario in semiconductor manufacturing, and the specific solutions of the application will be introduced below taking wafer manufacturing as an example.

[0043] Figure 1 An application scenario schematic diagram of the method for generating text for structured numerical data provided by the embodiment of the application is shown. Figure 1 As shown, the numerical sequence related to wafer manufacturing is input into the Data-To-Text model (i.e. the natural language generation model), and the model outputs a text sequence, which describes the abnormal information corresponding to the numerical sequence.

[0044] That is, the Data-To-Text model establishes the relationship between the numerical sequence related to wafer manufacturing and the description text, and the numerical sequence related to wafer manufacturing can be converted into the corresponding description text through the Data-To-Text model.

[0045] The Data-To-Text model can be deployed in any computing device with computing capability, for example, various server devices, including special server computers (such as personal computer servers, UNIX servers, terminal servers), blade servers, mainframe computers, server clusters, or any other appropriate arrangement or combination; various terminal devices, including various types of computer devices, such as portable handheld devices, general-purpose computers (such as personal computers or laptop computers), workstation computers, wearable devices, etc. The specific computing device of the device 20 is not limited by the embodiments of the present application.

[0046] Figure 2 A flowchart of a method for generating text for structured numerical data is provided in the embodiments of the present application. The method can be executed by any device, equipment, platform or device cluster with computing capability. The specific computing device for executing the method is not limited by the present application, and a suitable computing device can be selected for execution according to the needs. As shown in the figure, the method for generating text for structured numerical data at least includes steps S201-S202. Figure 2

[0047] In step S201, structured numerical data is obtained.

[0048] The structured numerical data is a numerical sequence related to wafer manufacturing.

[0049] In one example, the numerical sequence related to wafer manufacturing can include a numerical sequence related to wafer manufacturing equipment, such as data obtained by sensors of wafer manufacturing equipment (also referred to as machine) during production, including but not limited to temperature, humidity, pressure, voltage, current, etc.; and usage rate of wafer manufacturing equipment, etc.

[0050] And / or, wafer-related data, such as data obtained by defect detection during the production process of the wafer (such as wafer defect data); data obtained by electrical testing during the production process of the wafer (such as wafer failure type data, including CPU interval failure, GUP interval failure, storage interval failure, etc.); and wafer yield data, etc. pure numerical data.

[0051] ​The acquisition mode can be various, for example, can be directly acquired from the acquisition device of each machine in the semiconductor manufacturing process, that is, the acquisition device directly sends the acquired data to the device 10, or obtains the first type of data from the database, the database stores the structured numerical data generated by each machine in the semiconductor manufacturing process, or is obtained by receiving the input of the user, etc.

[0052] Optionally, the numerical sequence generated by the wafer manufacturing equipment can be SPC (Statistical Process Control) type data, that is, SPC time sequence numerical data.

[0053] Those skilled in the art can understand that SPC is a process control tool by means of mathematical statistics. It analyzes and evaluates the production process, discovers the signs of systematic factors in time according to the feedback information, and takes measures to eliminate their influence, so that the process is maintained in a controlled state only affected by random factors, in order to achieve the purpose of controlling quality. The method of statistics is used to monitor the state of the process, and to determine the production process in the state of control, so as to reduce the variation of product quality. The data generated by the equipment in the semiconductor manufacturing process is SPC type data, and at present, engineers analyze this kind of data by using SPC tool, that is, using statistical method, and then form the report documents of defect analysis and yield improvement according to the analysis results.

[0054] In step S202, the numerical sequence is input into the trained natural language generation model to obtain a description text corresponding to the numerical sequence, and the description text describes the abnormal information corresponding to the numerical sequence.

[0055] The natural language generation model is trained based on a training set, and the training set includes a plurality of training sample pairs, each training sample pair includes a numerical sequence and a description text corresponding to the numerical sequence.

[0056] The training set can be obtained in various ways. For example, it can be obtained by extracting from an engineer experience document, which records historical SPC time series numerical data generated by a wafer manufacturing device and comment texts made on the SPC time series numerical data. A plurality of SPC time series numerical data and comment texts corresponding to the SPC time series numerical data extracted from the engineer document constitute the training set. Alternatively, the historical SPC time series numerical data generated by a plurality of wafer manufacturing devices is obtained, and a preset rule template is used to match the corresponding text for the abnormal data segment in the SPC time series numerical data and the field information corresponding to the abnormal data segment to determine the description text corresponding to the SPC time series numerical data. The preset rule template is used to match the corresponding text description for the abnormal data segment, and the field information corresponding to the abnormal data segment includes one or more of the time information, wafer lot information, wafer number information, and device information corresponding to the abnormal data segment.

[0057] Figure 3 The inference process and the training process of the Data-To-Text model are shown. As shown in Figure 3 The SPC time series numerical data is input into the trained Data-To-Text model to obtain the description text corresponding to the SPC time series numerical data, i.e., a natural language sentence that is easier to recognize.

[0058] The training set of the Data-To-Text model can be based on historical SPC time series numerical data and a preset rule template to match the text description of the abnormal data segment of the SPC time series numerical data. Then, according to the position of the abnormal data segment in the SPC time series numerical data, i.e., the index of the abnormal point in the sequence, the original data is queried according to the index to obtain additional field information of the abnormal data segment, such as wafer manufacturing time information, wafer lot information, wafer number information, wafer test time information, and device information. Then, according to the field information, the index of the abnormal data segment, and the rule, a sentence is spliced according to a specific template, for example, the specific template is: from {} to {} time, lot_id = {}, wafer_id = {} in the device {} when {} is checked, {} phenomenon is observed, which may indicate {} problem. The sentence and the SPC time series numerical data constitute a training sample pair.

[0059] The description text of the abnormal data segment matched by the preset rule template, the field information of the index (wafer manufacturing time information, wafer batch information, wafer number information, wafer test time information, equipment information, etc.), and the abnormal data segment are filled into the {} in the specific template to obtain the sentence corresponding to the SPC time series numerical data.

[0060] Figure 4 A preset rule template is shown. This rule template includes multiple anomaly judgment conditions and descriptions of the anomalies corresponding to each anomaly judgment condition; among the multiple anomaly judgment conditions, the target anomaly judgment condition that the anomaly data sequence meets is determined; and the description of the anomalies corresponding to the target anomaly judgment condition is obtained as the description text.

[0061] In one implementation, multiple anomaly detection conditions and the corresponding anomaly descriptions can be found in Figure 3. When the anomaly sequence of a numerical sequence meets the condition "1 point is outside the control limits," meaning "a point falls outside area A," the anomaly description for that numerical sequence is "A large shift," meaning "there is a large deviation." In other words, the descriptive text for the anomaly sequence of that numerical sequence is "A large shift."

[0062] Rule templates can be implemented programmatically to generate text describing anomalous data. An example programmatic implementation of a rule template is shown below:

[0063] if type == 1:

[0064] phenomenon = "a large shift." # There are N points outside the UCL or LCL.

[0065] elif type==2:

[0066] phenomenon = "a small sustained shift on the upper side." # There are N points on the upper side of the midline.

[0067] elif type==3:

[0068] phenomenon = "a small sustained shift on the lower side." # There are N points below the midline.

[0069] elif type==4:

[0070] phenomenon="a trend of drift up." #N consecutive points rising

[0071] elif type==5:

[0072] phenomenon="a trend of drift down." #N consecutive points falling

[0073] elif type==6:

[0074] phenomenon="a non-random systematic variation." #N points rising and falling between each other

[0075] elif type==7:

[0076] phenomenon="a medium shift on the upper side." #N1 points in Zone A or beyond Zone A on the upper side of the median

[0077] elif type==8:

[0078] phenomenon="a medium shift on the lower side." #N1 points in Zone A or beyond Zone A on the lower side of the median

[0079] elif type==9:

[0080] phenomenon="a small shift on the upper side." #N1 points in Zone B or beyond Zone B on the upper side of the median

[0081] elif type==10:

[0082] phenomenon="a small shift on the lower side." #N1 points in Zone B or beyond Zone B on the lower side of the median

[0083] elif type==11:

[0084] phenomenon="a stratification problem." #N consecutive points in Zone C

[0085] elif type==12:

[0086] phenomenon = "a mixture pattern." # N consecutive points are not in Zone C

[0087] In one example, the Data-To-Text model architecture can be a Seq2Seq architecture. The Seq2Seq architecture includes an encoder and a decoder. The encoder uses a Transformer and employs a single-value MLP mapping to preserve certain numerical features, or other encoding methods to highlight the numerical features of the sequence. The decoder first uses a multi-layer LSTM and introduces a copy mechanism; the output of the decoder is the text corresponding to the numerical sequence.

[0088] In another example, the Data-To-Text model also employs a Seq2Seq architecture, including an embedding representation layer, a Transformer-based language generation backbone model, and a pointer generation network. The numerical sequence is input into the trained natural language generation model to obtain the corresponding descriptive text. This process includes: inputting the numerical sequence into the embedding representation layer, which outputs the embedding vector of the numerical sequence; inputting the embedding vector of the numerical sequence into the Transformer-based language generation backbone model, which outputs a vocabulary probability distribution and an attention distribution. The attention distribution represents the positional probability distribution of each value in the numerical sequence, or, in other words, the interactive attention features between the text to be generated at the current time step and the context vectors of each position in the numerical sequence; and inputting the vocabulary probability distribution and attention distribution into the pointer generation network, which outputs the descriptive text corresponding to the numerical sequence.

[0089] Optionally, the Transformer-based language generation backbone model includes an encoder and a decoder; the output of the embedding representation layer is connected to the input of the encoder, and the output of the decoder is connected to the input of the pointer generation network; the embedding representation layer is a BiLSTM-based temporal numerical embedding model.

[0090] Figure 5 This illustrates a structural diagram of the Data-To-Text model. (For example...) Figure 5 As shown, the Data-To-Text model consists of an Embedding layer, an Encoder layer, a Decoder layer, and a Pointer Generator layer.

[0091] The original input of the model encoding end is a sequence of numerical values, which needs to be vectorized before entering the subsequent process. Therefore, it first passes through the Embedding layer, i.e., the numerical value embedding layer, to form the encoding vector. Then it enters the encoding end composed of multiple encoding layers. These encoding layers learn the mutual attention features of each position in the sequence, and finally output the context encoding vector sequence of the sequence.

[0092] wherein the Embedding layer is a BiLSTM-based time series numerical value embedding representation layer, and the structure is as shown in Figure 6 .

[0093] The original input of the decoding end is the comment text written by the engineer for the numerical sequence. After the text is encoded by word embedding, it enters the decoding end composed of multiple decoding layers. Each decoding layer first learns the semantic features between the words in the text in the decoder self-attention submodule. Then, in the encoder-decoder cross-attention submodule, the output of the encoder is taken as the input to calculate the cross-attention with the output of the self-attention submodule of the decoding layer, to learn the cross-attention features between the text and the context vector of the numerical sequence of the encoding end. This process is repeated until the last layer of the decoder, and the decoding end output is obtained.

[0094] The final output of the decoding end and the encoder-decoder cross-attention of the last layer of the decoding end will be input into the pointer generation network. The network will calculate the generation probability, word distribution and attention distribution, and obtain the final distribution accordingly. The final distribution determines what is generated or copied at this time step.

[0095] The decoder layer is mostly consistent with the structure of the encoder layer, except that the decoder adds a masked multi-head attention layer before the structure of the encoder layer. Unlike the ordinary multi-head attention layer, the calculation of attention at each position is only related to the previous positions, and the subsequent positions are masked and do not participate in the calculation. This makes the decoding step become a kind of cyclic decoding similar to RNN, which can only be decoded at this step after receiving the decoding result of the previous step, as shown in Figure 7 .

[0096] In another example, the method for generating text for structured numerical data provided by the embodiment of the present application further includes a preprocessing step of preprocessing the numerical sequence generated by the wafer manufacturing equipment before step S202, and a step of processing the text output by the Data-To-Text model after step S202 (see Figure 8 ).

[0097] For example, the step of preprocessing the numerical sequence can include cleaning the numerical sequence to remove numerical values greater than or equal to a first preset threshold and numerical values less than or equal to a second preset threshold, wherein the first preset threshold is greater than the second preset threshold; and / or, normalizing the numerical sequence.

[0098] Since there can be some extremely large values in the numerical sequence that are much higher than the normal level or some extremely small values that are much lower than the normal level, these values seriously affect the average value of the wafer, so the average value is calculated after removing these values, that is, the numerical sequence is cleaned, and the preset threshold can be set according to the actual situation, and the value of the preset threshold is not limited in the embodiments of the present application. That is, through data cleaning, some extremely large and small values deviating from the normal level are cleaned to facilitate obtaining the average value of the wafer.

[0099] Since different SPC numerical ranges are different (such as ucl and lcl), the model processing sequence needs the same range, so the different SPC numerical sequences need to be scaled accordingly to be included in the same range, that is, the numerical sequence is normalized, and the method is as follows:

[0100] 1. Calculate the mean μ and standard deviation σ of the same type of SPC data after removing outliers (that is, the numerical sequence after cleaning).

[0101] 2. x'[i] = (x[i] - μ) / σ (z-score normalization).

[0102] 3. ucl = μ + 3σ.

[0103] 4. lcl = μ - 3σ.

[0104] 5. The numerical sequence processed in this way will have values greater than ucl > 3 and values less than lcl <- 3, that is, the range defined by the ucl and lcl of the SPC data will be changed to [-3, 3].

[0105] As understood by those skilled in the art, the meaning of ucl is the upper limit of the specification of the characteristic value, that is, the product characteristic is greater than ucl, which will cause unqualified in engineering; the meaning of lcl is the lower limit of the specification of the characteristic value, that is, the product characteristic is less than lcl, which will cause unqualified in engineering.

[0106] The step of processing the text output by the Data-To-Text model includes further language processing of the text data output by the model to make the finally output text more fluent and easy to read.

[0107] In one example, the method described above can further include, based on the description text, performing knowledge extraction for forming a target knowledge graph related to wafer manufacturing, which can be used for wafer manufacturing related automatic reasoning, such as one or more of inferring a defect, a failure category, a root cause of the defect / failure category, and a decision corresponding to the root cause to improve the yield of the wafer of the wafer according to SPC data of the wafer.

[0108] Based on the same idea as the foregoing method embodiments, the embodiments of the present application also provide a device 900 for generating text for structured numerical data, which comprises units or modules for implementing each step in the method shown in the device 900. Figures 2-8 The device 900 for generating text for structured numerical data comprises units or modules for implementing each step in the method shown in the device 900.

[0109] Figure 9 A structural schematic diagram of a device for generating text for structured numerical data provided by the embodiments of the present application is shown in the device 800. Figure 9 The device 800 for generating text for structured numerical data at least comprises:

[0110] The acquisition module 901 is configured to acquire structured numerical data, which is a numerical sequence related to wafer manufacturing.

[0111] The language generation module 902 is configured to input the numerical sequence into a trained natural language generation model to obtain description text corresponding to the numerical sequence, the description text describing abnormal information corresponding to the numerical sequence.

[0112] The device 900 for generating text for structured numerical data provided by the embodiments of the present application can correspond to performing the method described in the embodiments of the present application, and the above and other operations and / or functions of each module in the device 900 for generating text for structured numerical data are respectively for implementing the corresponding procedures of each method in the device 900 for generating text for structured numerical data. Figures 2-8 The specific implementation can be referred to the description above, and will not be described here for the sake of brevity.

[0113] The embodiments of the present application also provide a computing device comprising at least one processor, a memory and a communication interface, the processor being configured to execute the method described above.The computing device can be a server or a terminal device. Figures 2-8

[0114] Figure 10 A structural schematic diagram of a computing device provided by the embodiments of the present application is shown in the device 800.

[0115] As shown in the device 800. Figure 10As shown, the computing device 1000 includes at least one processor 1001, a memory 1002, a communication interface 1003. Among them, the processor 1001 and the memory 1002 and the communication interface 1003 are in communication connection, and can realize communication through wireless or wired and the like. The communication interface 1003 is used to receive user instructions or information sent by the acquisition device; the memory 1002 stores computer instructions, and the processor 1001 executes the computer instructions to execute the method in the foregoing method embodiments.

[0116] It should be understood that in the embodiments of the present application, the processor 1001 can be a central processing unit CPU, and the processor 1001 can also be other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASIC), field programmable gate arrays (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc.

[0117] The memory 1002 can include read-only memory and random access memory, and provide instructions and data for the processor 1001. The memory 1002 can also include non-volatile random access memory.

[0118] The memory 1002 can be volatile memory or nonvolatile memory, or can include both volatile and nonvolatile memory. By way of illustration, and not limitation, nonvolatile memory can be read-only memory (ROM), programmable ROM (PROM), erasable PROM (EPROM), electrically EPROM (EEPROM), or flash memory. Volatile memory can be random access memory (RAM), which acts as external cache. By way of illustration and not limitation, many forms of RAM are available, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double-data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), Synchlink DRAM (SLDRAM), and direct rambus RAM (DR RAM).

[0119] It should be appreciated that the computing device 1000 according to embodiments of the present application can execute the method shown in the embodiments of the present application, detailed description of which is referred to above, and for brevity, will not be repeated here. Figures 2-8 It should be appreciated that the computing device 1000 according to embodiments of the present application can execute the method shown in the embodiments of the present application, detailed description of which is referred to above, and for brevity, will not be repeated here.

[0120] Embodiments of the present application provide a computer readable storage medium having stored thereon a computer program, which when executed by a processor, causes the above-mentioned method of generating text for structured numerical data to be implemented.

[0121] Embodiments of the present application provide a computer program or computer program product, which includes instructions, which when executed, cause a computer to execute the above-mentioned method of generating text for structured numerical data.

[0122] Those skilled in the art should further understand that the units and algorithm steps of each example described in combination with the embodiments disclosed herein can be realized in electronic hardware, computer software or a combination of both. In order to clearly illustrate the interchangeability of hardware and software, the components and steps of each example have been described in the above description in a general manner. Whether the functions are realized in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to realize the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0123] The steps of the methods or algorithms described in combination with the embodiments disclosed herein can be implemented in hardware, software modules executed by a processor, or a combination of both. The software modules can be placed in random access memory (RAM), internal memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disks, removable disks, CD-ROMs, or any other form of storage medium known in the art.

[0124] The above specific embodiments further illustrate the purpose, technical solutions and beneficial effects of the present application. It should be understood that the above description is only a specific embodiment of the present application and is not intended to limit the protection scope of the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application should be included in the protection scope of the present application.

Claims

1. A method of generating text for structured numeric data type, characterized by, The method comprises: obtaining structured numerical data, the structured numerical data being a numerical sequence related to semiconductor manufacturing; the numerical sequence being statistical process control (SPC) time series numerical data; the SPC time series numerical data comprising time series numerical sequences related to wafer manufacturing equipment and / or wafer-related data; inputting the numerical sequence into a trained natural language generation model to obtain a description text corresponding to the numerical sequence, the description text describing abnormal information corresponding to the numerical sequence; the natural language generation model comprising an embedding representation layer, a Transformer-based language generation backbone model, and a pointer generation network; the embedding representation layer being a BiLSTM-based time series numerical embedding model; the natural language generation model being trained based on a training set, the training set comprising a plurality of training sample pairs, the training sample pairs comprising the numerical sequence and the description text corresponding to the numerical sequence; the description text corresponding to the numerical sequence being determined based on an abnormal data segment of the numerical sequence, a rule template, and field information corresponding to the abnormal data segment; the rule template being used to match a corresponding text description for the abnormal data segment, the field information corresponding to the abnormal data segment comprising one or more of time information, wafer batch information, wafer number information, and equipment information; the rule template comprising a plurality of abnormality judgment conditions and abnormality phenomenon descriptions corresponding to each abnormality judgment condition.

2. The method of claim 1, wherein, The architecture of the natural language generation model is a Seq2Seq architecture.

3. The method of claim 2, wherein, The inputting of the numerical sequence into the trained natural language generation model to obtain the description text corresponding to the numerical sequence comprises: inputting the numerical sequence into the embedding representation layer to output an embedding vector of the numerical sequence; inputting the embedding vector of the numerical sequence into the Transformer-based language generation backbone model to output a vocabulary probability distribution and an attention distribution, the attention distribution representing a position probability distribution of each value in the numerical sequence; inputting the vocabulary probability distribution and the attention distribution into the pointer generation network to output the description text corresponding to the numerical sequence.

4. The method of claim 3, wherein, The Transformer-based language generation backbone model comprises an encoder and a decoder; an output end of the embedding representation layer is connected to an input end of the encoder, and an output end of the decoder is connected to an input end of the pointer generation network.

5. The method according to any one of claims 1 to 3, characterized in that, Before the inputting of the numerical sequence into the trained natural language generation model to obtain the description text corresponding to the numerical sequence, the method further comprises: preprocessing the numerical sequence.

6. The method of claim 5, wherein, The preprocessing of the numerical sequence comprises one or more of the following: performing cleaning processing on the numerical sequence to remove values greater than or equal to a first preset threshold and less than or equal to a second preset threshold, wherein the first preset threshold is greater than the second preset threshold; performing standardization processing on the numerical sequence after the cleaning processing.

7. An apparatus for generating text for structured numeric data, the apparatus comprising: a data source for providing structured numeric data; a data processor for processing the structured numeric data; and a text generator for generating text based on the processed structured numeric data. The method comprises: An acquisition module is configured to acquire structured numerical data, the structured numerical data being a numerical sequence related to semiconductor manufacturing; the numerical sequence being SPC time series numerical data; the SPC time series numerical data including time series numerical sequences related to wafer manufacturing equipment and / or wafer-related data; A language generation module is configured to input the numerical sequence into a trained natural language generation model to obtain a description text corresponding to the numerical sequence, the description text describing abnormal information corresponding to the numerical sequence; the natural language generation model including an embedding representation layer, a language generation backbone model based on a Transformer, and a pointer generation network; the embedding representation layer being a time series numerical embedding model based on BiLSTM; The natural language generation model is trained based on a training set, the training set including a plurality of training sample pairs, the training sample pairs including the numerical sequence and the description text corresponding to the numerical sequence; The description text corresponding to the numerical sequence is determined based on an abnormal data segment of the numerical sequence, a rule template, and field information corresponding to the abnormal data segment; The rule template is used to match a corresponding text description for the abnormal data segment, the field information corresponding to the abnormal data segment including one or more of time information, wafer batch information, wafer number information, and equipment information corresponding to the abnormal data segment; the rule template including a plurality of abnormality judgment conditions and abnormality phenomenon descriptions corresponding to each abnormality judgment condition.

8. A computer-readable storage medium having stored thereon a computer program, characterized in that, When the computer program is executed in the computer, the computer is caused to perform the method of any one of claims 1-6.

Images