Text encoding device, system and method for network planning
Through the neural network model, text data is segmented and encoded, and low-dimensional network business influence data is generated, which solves the problem of low text data extraction efficiency in the existing technology and improves the automation and accuracy of network planning.
Patent Information
- Application Number
- CN202010831082.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2020-08-18
- Publication Date
- 2024-11-26
- Estimated Expiration
- 2040-08-18
AI Technical Summary
The prior art is difficult to effectively extract network service-related information from a large number of natural language text data, resulting in inefficient network planning and lack of unified standards and automated processing methods.
The text encoding system based on the neural network model is adopted to convert text data into low-dimensional word vectors through word segmentation processing and single hot encoding, and network business influence data is generated through training models, which is used for resource allocation decisions in the network planning system.
It realizes efficient and automated text data encoding, generates low-dimensional and high-information network business influence data, significantly improving the operating efficiency and accuracy of the network planning system.
Smart Images

Figure CN114077820B_ABST
Abstract
Description
Technical Field
[0001] The present disclosure relates to a text encoding device, system and method for network planning, and in particular to a text encoding device, system and method for network planning based on a neural network model. Background Art
[0002] In the process of network service construction of communication networks and other networks, it is necessary to plan the allocation of various resources in advance. For example, it is necessary to plan the equipment, personnel, funds, etc. allocated to each network node, or the distance between each network point and the corresponding cost. There are many factors that affect such resource allocation. As one of the factors, we can cite the information related to network services contained in various text data obtained from the Internet, paper media, etc. The impact of these text data on network development is very obvious. However, since text descriptions belong to natural language, their impact can often only be judged by practitioners based on their own experience, and there is no unified standard. Moreover, the amount of these text data is very large, resulting in very low efficiency in obtaining information from these text data.
[0003] In the network planning oriented to refinement, accuracy and automation, a method is needed to obtain information from relevant text data and generate coded information containing the effect of text data, which is used to provide to the network planning system as a parameter affecting network planning. And the coded information should be as low-dimensional and high-information as possible. Therefore, it is urgent to propose a text coding system for network planning. Summary of the invention
[0004] One purpose of the embodiments of the present disclosure is to provide a text encoding system and device for network planning, which can automatically obtain information from relevant text data and encode it in a low-dimensional manner. The encoded information contains the impact capability on the corresponding network service and can be used to provide it to the network planning system for analysis.
[0005] According to one aspect of the present disclosure, a text encoding device for network planning is provided, including: an input unit, which inputs text data to be encoded; a word segmentation unit, which performs word segmentation on the text data using a word segmentation vocabulary that stores words used for word segmentation, and converts the result of the word segmentation into a word vector group including multiple W-dimensional word vectors through one-hot encoding; and a text encoding neural network model, which inputs the word vector group and outputs an N-dimensional vector as network business influence data, wherein W and N are both natural numbers greater than 1, and W>N.
[0006] According to another aspect of the present disclosure, a text encoding device for network planning is provided, comprising: an input unit, which inputs text data and a network service change amount corresponding to the text data; a word segmentation unit, which performs word segmentation on the text data using a word segmentation vocabulary storing words used for word segmentation, and converts the result of the word segmentation into a word vector group including multiple W-dimensional word vectors through one-hot encoding; and a text encoding neural network model, which inputs the word vector group and outputs an N-dimensional vector, wherein W and N are both natural numbers greater than 1, and W>N, and a training unit, which stores an N-dimensional training vector, multiplies the N-dimensional vector output from the text encoding neural network model by the training vector to obtain a training scalar, and trains the text encoding neural network model in a manner such that the value of the training scalar is equal to the value of the network service change amount.
[0007] According to another aspect of the present disclosure, a text encoding system for network planning is provided, comprising: a training data storage unit storing training text data and training network service variation corresponding to the training text data; a document receiving unit receiving text data to be encoded; a word segmentation unit performing word segmentation on the training text data or the text data to be encoded using a word segmentation word library storing words for word segmentation, and converting the result of the word segmentation into a word vector group including a plurality of W-dimensional word vectors through single-hot encoding; a text encoding neural network model inputting the word vector group and outputting an N-dimensional vector, wherein W and N are both natural numbers greater than 1, and W>N; a text encoding neural network training unit storing an N-dimensional training vector, wherein When the word segmentation unit performs word segmentation processing on the training text data and outputs the corresponding word vector group to the text encoding neural network model, the N-dimensional vector output from the text encoding neural network model and the training vector are multiplied to obtain a training scalar, so that the value of the training scalar is equal to the value of the training network business variation, and the text encoding neural network model is trained; the text encoding unit, when the word segmentation unit performs word segmentation processing on the text data to be encoded and outputs the corresponding word vector group to the text encoding neural network model, outputs the N-dimensional vector output from the text encoding neural network model as network business influence data; and an output unit outputs the network business influence data output from the text encoding unit to the outside.
[0008] According to another aspect of the present disclosure, a network planning system is provided, comprising: an input unit, which inputs network business influence data from the network planning-oriented text encoding device or the network planning-oriented text encoding system; a network planning unit, which plans resource allocation of each node on the network according to the network business influence data from the input unit; and an output unit, which outputs the planning results obtained by the network planning unit to the outside.
[0009] According to another aspect of the present disclosure, a text encoding method for network planning is provided, comprising: an input step of inputting text data to be encoded; a word segmentation step of performing word segmentation on the text data using a word segmentation vocabulary storing words for word segmentation, and converting the result of the word segmentation into a word vector group including multiple W-dimensional word vectors through one-hot encoding; and an encoding step of running a text encoding neural network model that inputs the word vector group and outputs an N-dimensional vector, and outputting the N-dimensional vector as network business influence data, wherein W and N are both natural numbers greater than 1, and W>N.
[0010] According to another aspect of the present disclosure, a training method for a text encoding device for network planning is provided, the training method comprising: an input step of inputting text data and a network service change amount corresponding to the text data; a word segmentation step of performing word segmentation on the text data using a word segmentation vocabulary storing words for word segmentation, and converting the result of the word segmentation into a word vector group including a plurality of W-dimensional word vectors through one-hot encoding; and a training step of training a text encoding neural network model that inputs the word vector group and outputs an N-dimensional vector, wherein W and N are both natural numbers greater than 1, and W>N, and in the training step, the N-dimensional vector output from the text encoding neural network model and the N-dimensional training vector are multiplied to obtain a training scalar, and the text encoding neural network model is trained in such a way that the value of the training scalar is equal to the value of the network service change amount.
[0011] Through the implementation methods of the present disclosure, efficient text encoding for network planning can be achieved, and the impact of text data on network services can be automatically encoded and directly used for network planning predictions. In particular, the encoded information based on this implementation method includes the impact effect on network services, and is low-dimensional and high-density. Therefore, the impact effect on network services provided to the network planning system is no longer an abstract concept, but is directly adopted by the network planning system as a certain encoded data, which greatly improves the operating efficiency of the entire system and has a very broad application prospect.
[0012] Other features and advantages of the present disclosure will become apparent from the following detailed description of exemplary embodiments of the present disclosure with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS
[0013] The accompanying drawings, which constitute a part of the specification, illustrate embodiments of the present disclosure and, together with the description, serve to explain the principles of the present disclosure.
[0014] The present disclosure can be more clearly understood from the following detailed description with reference to the accompanying drawings.
[0015] in:
[0016] Figure 1 It is a block diagram illustrating the structure of a text encoding device 100 for network planning.
[0017] Figure 2 is a schematic diagram illustrating a text encoding method for network planning.
[0018] Figure 3 It is a block diagram illustrating the structure of a text encoding device 200 for network planning.
[0019] Figure 4 This is a diagram illustrating a training method for a text encoding device for network planning.
[0020] Figure 5 It is a block diagram illustrating the structure of the text encoding system 300 for network planning.
[0021] Figure 6 is a schematic diagram illustrating the internal structure of the neural network 400.
[0022] Figure 7 is a schematic diagram illustrating the internal structure of the neural network 500.
[0023] Figure 8 It is a block diagram illustrating the structure of the network planning system 600.
[0024] Fig. 9 is a diagram showing an exemplary configuration of a computing device 1000 that can implement an embodiment according to the present disclosure. DETAILED DESCRIPTION
[0025] Hereinafter, preferred embodiments of the present disclosure will be described in detail with reference to the accompanying drawings. Note that, in this specification and the drawings, structural elements that have substantially the same function and structure are denoted with the same reference numerals, and repeated explanation of these structural elements is omitted.
[0026] In the present disclosure, the terms “first”, “second”, etc. are used merely to distinguish between elements or steps, but are not intended to indicate a time sequence, priority, or importance.
[0027] Implementation Method 1
[0028] Figure 1 It is a block diagram illustrating the structure of a text encoding device 100 for network planning.
[0029] like Figure 1 As shown, the text encoding device 100 for network planning of this embodiment includes an input unit 11, a word segmentation unit 12, a neural network 13 and a word segmentation vocabulary 14.
[0030] The input unit 11 inputs text data to be encoded. The text data may be data of various texts collected from the Internet, paper media, etc. For example, these texts may include "vigorously promote the construction of 5G networks", "gradually carry out the withdrawal of 2G networks", etc.
[0031] The word segmentation unit 12 uses the word segmentation vocabulary 14 to perform word segmentation processing on the text data input by the input unit 11, and converts the word segmentation data into multiple W (W is a natural number greater than 1) dimensional word vectors through one-hot encoding.
[0032] The word segmentation word library 14 is a database storing words used for word segmentation processing. For example, the words stored in the word segmentation word library may be words such as "vigorously", "promote", "gradually", "carry out", "5G", "2G" and so on that can reflect trends, degrees and related network services.
[0033] In addition, Figure 1 , the word segmentation thesaurus 14 is shown to be located inside the text encoding device 100 , but the word segmentation thesaurus 14 may also be located outside the text encoding device 100 , or may also be located inside the word segmentation unit 12 .
[0034] One-hot encoding can be performed in a manner known in the art. One-hot encoding is a widely used encoding method in the field of machine learning, and has the advantage of being easy to optimize and learn in the model. However, the dimension of the vector data after one-hot encoding is very high, which increases the burden of data processing and affects the operating efficiency of the entire network planning system. Therefore, in this embodiment, the dimension of the vector data after one-hot encoding is reduced by the neural network 13 and the high density of information is maintained.
[0035] The neural network 13 is a neural network model that inputs multiple W-dimensional vectors and outputs corresponding N (N is a natural number greater than 1, W>N)-dimensional vectors. The output N-dimensional vector is the network business influence data used for planning of the network planning system. The network business influence data represents the influence of the content of the input text data on the network business. Preferably, W is more than 100 times N. In this way, the dimension of high-dimensional vector data can be reduced to a low dimension far lower than the original dimension, greatly reducing the data processing load.
[0036] Figure 2 is a schematic diagram illustrating a text encoding method for network planning. Figure 1 and Figure 2 This encoding method is described.
[0037] First, the input unit 11 inputs text data X to be encoded.
[0038] Then, the word segmentation unit 12 performs word segmentation processing on the text data X using the word segmentation vocabulary 14, and converts the word segmentation processed data into a word vector group including multiple W-dimensional word vectors through one-hot encoding.
[0039] For example, the text data X is processed to generate Figure 2 The word vector group x1, x2, ..., x1 includes n W-dimensional word vectors. n Each word vector is obtained by one-hot encoding, so most of its elements are zero, for example, x n =(0, 0, 0,…, 1, 0,…, 0).
[0040] Then, the n W-dimensional word vectors are input into the neural network 13, and after being processed by the neural network 13, an N-dimensional vector y is obtained, and the N-dimensional vector y is the network business influence data used for planning of the network planning system. The network business influence data represents the influence of the content of the input text data on the network business. Preferably, W is more than 100 times N.
[0041] The following describes the structure of training the text encoding device 100 for network planning.
[0042] Figure 3 2 is a block diagram illustrating the structure of a text encoding device 200 for network planning. The text encoding device 200 corresponds to the structure of the text encoding device 100 when performing training.
[0043] like Figure 3 As shown, the text encoding device 200 for network planning includes an input unit 21, a word segmentation unit 12, a neural network 13, a training unit 25, and a word segmentation vocabulary 14. Compared with the text encoding device 100, the difference lies in the input unit 21 and the training unit 25. Therefore, the following mainly describes these two units, and the repeated description is omitted.
[0044] The input unit 21 inputs text data and network traffic variation corresponding to the text data. Similar to the text data input by the input unit 11, the text data input by the input unit 21 may be various text data collected from the Internet, paper media, etc.
[0045] Network services can be services of networks such as communication networks. The change in network services represents the change in network service volume within a specified period. For example, when the number of 4G users is used to represent the network service volume, assuming that there were 20,000 4G users in District D of City A at the end of 2014, and 120,000 4G users in the district at the end of 2019, the change in network services in the five years from 2015 to 2019 is 12-2=10 (ten thousand). For example, when the number of users who apply for home broadband is used to represent the network service volume, assuming that there were 20,000 home broadband users in District G of City E at the end of 2017, and 40,000 home broadband users in the district at the end of 2019, the change in network services in the two years from 2018 to 2019 is 4-2=2 (ten thousand).
[0046] The training unit 25 is used for training the neural network 13, and stores an N-dimensional training vector. The N-dimensional vector output from the neural network 13 is multiplied by the training vector to obtain a training scalar L. The vector obtained after the word segmentation unit 12 performs single-hot encoding can be either a row vector or a column vector. If the word segmentation unit 12 obtains a row vector, the training vector is a column vector. If the word segmentation unit 12 obtains a column vector, the training vector is a row vector. As long as a scalar can be obtained by multiplying the N-dimensional vector output from the neural network 13 and the training vector, it will be fine.
[0047] The training unit 25 uses the training scalar L and the text data and network traffic variation input by the input unit 11 to train the neural network 13 in such a way that the value of the training scalar L is equal to the value of the network traffic variation.
[0048] Figure 4 This is a diagram illustrating a training method for a text encoding device for network planning according to Embodiment 1. Figure 3 and Figure 4 Describe the training method.
[0049] First, the input unit 11 inputs the kth text data X (M and k are natural numbers greater than or equal to 1, k≤M) among the M text data X. k Assume that the content of text data X affects the network traffic change ΔL within Δt time.
[0050] Then, the word segmentation unit 12 uses the word segmentation word library 14 to segment the text data X k Word segmentation is performed, and the data after word segmentation is converted into a word vector group including multiple W-dimensional word vectors through one-hot encoding.
[0051] For example, for the text data X k , generated as Figure 4 The word vector group shown includes n W-dimensional word vectors Each word vector is obtained by one-hot encoding, so most of its elements are zero, for example
[0052] Then, the n W-dimensional word vectors are input into the neural network 13 to obtain the N-dimensional vector y k , y k Multiply it with the N-dimensional training vector to get the training scalar L k . Where n, W and N are all natural numbers greater than 1, and W>N.
[0053] As mentioned above, the word vector group input by the neural network 13 The dimensions of each word vector are much larger than the output N-dimensional vector y k dimension, preferably, W is more than 100 times of N.
[0054] Finally, using the text data X k , network traffic change ΔL and training scalar L k , train the neural network 13 so that L k Equal to ΔL.
[0055] In addition, preferably, when training the neural network 13, according to the loss function loss=(L-ΔL) 2 , using stochastic gradient descent for training, where L is the training scalar L k .
[0056] Figure 5 is a block diagram illustrating the structure of a text encoding system 300 for network planning. The text encoding system 300 includes a structure for encoding and a structure for training.
[0057] Specifically, the text encoding system 300 includes a training data storage unit 31, a document receiving unit 32, a word segmentation unit 33, a neural network 34, a training unit 35, an encoding unit 36 and an output unit 37.
[0058] The training data storage unit 31 stores training text data and training network traffic variations corresponding to the training text data. The training text data and training network traffic variations correspond to the text data input by the input unit 21 of the text encoding device 200 and the corresponding network traffic variations.
[0059] The document receiving unit 32 receives text data to be encoded, which corresponds to the text data to be encoded input by the input unit 11 of the text encoding device 100 .
[0060] The word segmentation unit 33 corresponds to the word segmentation unit 12 of the text encoding device 100 or the text encoding device 200, and uses a word segmentation vocabulary that stores words used for word segmentation processing to perform word segmentation processing on the training text data from the training data storage unit 31 or the text data to be encoded from the document receiving unit 32, and converts the data after word segmentation processing into a word vector group including multiple W-dimensional word vectors through one-hot encoding.
[0061] The word segmentation thesaurus here corresponds to the word segmentation thesaurus 14 of the text encoding device 100 or the text encoding device 200 , and can be located inside or outside the text encoding system 300 , or can also be located inside the word segmentation unit 33 .
[0062] The neural network 34 corresponds to the neural network 13 of the text encoding device 100 or the text encoding device 200, and is a neural network model that inputs a plurality of W-dimensional vectors and outputs corresponding N-dimensional vectors (W and N are natural numbers greater than 1, W>N). Preferably, W is greater than 100 times N.
[0063] The training unit 35 is a unit for training the neural network 34, and corresponds to the training unit 25 of the text encoding device 200. When the word segmentation unit 33 processes the data from the training data storage unit 31 (in the training mode), the training unit 35 multiplies the N-dimensional vector output from the neural network 34 and the stored N-dimensional training vector to obtain a training scalar L, and uses the training scalar L and the training text data from the training data storage unit 31 and the training network business variation to train the neural network 34 in such a way that the value of the training scalar L is equal to the value of the training network business variation.
[0064] Preferably, when training the neural network 34, according to the loss function loss = (L-ΔL) 2 , the stochastic gradient descent method is used for training, where ΔL is the change in network traffic used for training.
[0065] When the word segmentation unit 33 processes the data from the document receiving unit 32 (in encoding mode), the encoding unit 36 outputs the N-dimensional vector output from the neural network 34 as network business influence data.
[0066] The output unit 37 is the output interface of the entire text encoding system 300, and outputs the network business influence data output from the encoding unit 36 to the outside.
[0067] Through the technical solution of this embodiment, useful information can be automatically obtained from relevant text data and low-dimensional coding information containing the impact effects of the text data can be generated, reducing the operating load of the entire network planning system.
[0068] Implementation Method 2
[0069] In this embodiment, the internal structure of the neural network 400 in the text encoding device for network planning is described.
[0070] Figure 6 4 is a schematic diagram illustrating the internal structure of the neural network 400. The neural network 400 here corresponds to the neural network 13 of the text encoding device 100 and the text encoding device 200 and the neural network 34 of the text encoding system 300.
[0071] In addition, as mentioned above, the vector obtained after the word segmentation unit 12 performs one-hot encoding can be either a row vector or a column vector. Here, the word segmentation unit 12 is used as an example to illustrate that the vector is a row vector. Figure 6 The n word vectors x1, x2, ..., x are input to the neural network m , …, x n are all W-dimensional row vectors.
[0072] like Figure 6 As shown, the neural network 400 includes a plurality of first matrices A1, A2, ..., A m , …, A n and a second matrix B with H rows and N columns, wherein m and n are natural numbers greater than 1. The elements of each first matrix and second matrix may be random values in the initial state. However, the initial value of each matrix element may also be specified. For example, the initial value of each matrix element may also be specified to be 0 or 1.
[0073] Multiple word vectors x1, x2, ..., x are input to the neural network 400. m , …, x n are multiplied by a plurality of first matrices A1, A2, ..., A m , …, A n To become multiple H-dimensional row vectors, add the multiple H-dimensional row vectors to get the sum S, which is also an H-dimensional row vector, and multiply the sum S by the second matrix B to get an N-dimensional row vector y.
[0074] When the text data X to be encoded is processed (in encoding mode), the N-dimensional row vector y is the network business influence data as the output of the neural network 400.
[0075] In the text data X used for training k (k is a natural number greater than 1) (in training mode), the N-dimensional row vector y is equivalent to Figure 4 The y in k , for this kFurther multiply it by the N-dimensional training vector (the training vector is a column vector at this time) to obtain the training scalar L k Then, the neural network 400 is trained according to the method of implementation mode 1.
[0076] Here, the description is made by taking the case where the word segmentation unit 12 obtains a row vector as an example. However, the technical solution of this embodiment can also be applied when the word segmentation unit 12 obtains a column vector.
[0077] In this case, the n word vectors x1, x2, ..., x m , …, x n Each of them is a W-dimensional row vector (m and n are natural numbers greater than 1), and the neural network 400 includes a plurality of first matrices with H rows and W columns and a second matrix with N rows and H columns. The plurality of first matrices with H rows and W columns are multiplied by the plurality of word vectors x1, x2, ..., x1 input to the neural network 400. m , …, x n To obtain multiple H-dimensional column vectors, add the multiple H-dimensional column vectors to obtain a sum S, which is also an H-dimensional column vector, and multiply the second matrix of N rows and H columns by the sum S to obtain an N-dimensional column vector y.
[0078] When the text data X to be encoded is processed (in encoding mode), the N-dimensional column vector y is the network business influence data as the output of the neural network 400.
[0079] In the text data X used for training k (k is a natural number greater than 1) (in training mode), the N-dimensional column vector y is equivalent to Figure 4 The y in k , for this k Further multiply it by the N-dimensional training vector (the training vector is a row vector at this time) to get the training scalar L k Then, the neural network 400 is trained according to the method of implementation mode 1.
[0080] Implementation 3
[0081] In this embodiment, a preferred structure of a neural network in a text encoding device for network planning is described.
[0082] Figure 7 Schematic diagram illustrating the internal structure of the neural network 500 of this embodiment. Figure 7 As shown, the neural network 500 of this embodiment adds a nonlinear transformation module based on the neural network of embodiments 1 and 2.
[0083] The nonlinear transformation module uses the Softmax function to map the value of each element of the N-dimensional vector obtained after being processed by the neural network 13, the neural network 34 or the neural network 400 in embodiments 1 and 2 to a value greater than 0 and less than 1.
[0084] Therefore, the value of each element of the N-dimensional vector y obtained by the neural network 500 is greater than 0 and less than 1. This is beneficial to improving the training efficiency of the neural network and can provide easy-to-process data to the network planning system.
[0085] Implementation 4
[0086] Figure 8 It is a block diagram illustrating the structure of the network planning system 600 according to the present embodiment.
[0087] exist Figure 8 In the embodiment, the network planning system 600 includes an input unit 61 , a network planning unit 62 and an output unit 63 .
[0088] The input unit 61 inputs the network business influence data from the text encoding device or text encoding system for network planning described in the above-mentioned embodiments 1 to 3.
[0089] The network planning unit 62 plans the resource allocation of each node on the network according to the network business influence data from the input unit 61. For example, according to the network business influence data, an evaluation is made on how much equipment, personnel, or funds are needed to build a regional communication business point, and the total amount of equipment, personnel, or funds allocated to the communication business point is estimated accordingly.
[0090] The output unit 63 outputs the planning result obtained by the network planning unit 62 to the outside.
[0091] Fig. 9 An exemplary configuration of a computing device 1000 that can implement embodiments according to the present disclosure is shown. The computing device 1000 is an example of a hardware device to which the above aspects of the present disclosure can be applied. The computing device 1000 can be any machine configured to perform processing and / or calculations. The computing device 1000 can be, but is not limited to, a workstation, a server, a desktop computer, a laptop computer, a tablet computer, a personal data assistant (PDA), a smart phone, a vehicle-mounted computer, or a combination thereof.
[0092] like Fig. 9As shown, computing device 1000 may include one or more elements that may be connected or communicated with bus 1002 via one or more interfaces. Bus 1002 may include, but is not limited to, Industry Standard Architecture (ISA) bus, Micro Channel Architecture (MCA) bus, Enhanced ISA (EISA) bus, Video Electronics Standards Association (VESA) local bus, and Peripheral Component Interconnect (PCI) bus, etc. Computing device 1000 may include, for example, one or more processors 1004, one or more input devices 1006, and one or more output devices 1008. One or more processors 1004 may be any kind of processor, and may include, but is not limited to, one or more general-purpose processors or special-purpose processors (such as special-purpose processing chips). Input device 1006 may be any type of input device capable of inputting information to a computing device, and may include, but is not limited to, a mouse, a keyboard, a touch screen, a microphone, and / or a remote controller. Output device 1008 may be any type of device capable of presenting information, and may include, but is not limited to, a display, a speaker, a video / audio output terminal, a vibrator, and / or a printer.
[0093] The computing device 1000 may also include or be connected to a non-transitory storage device 1014, which may be any non-transitory storage device that can implement data storage, and may include, but is not limited to, a disk drive, an optical storage device, a solid-state memory, a floppy disk, a flexible disk, a hard disk, a tape, or any other magnetic medium, a compressed disk or any other optical medium, a cache memory, and / or any other storage chip or module, and / or any other medium from which a computer can read data, instructions, and / or code. The computing device 1000 may also include a random access memory (RAM) 1010 and a read-only memory (ROM) 1012. The ROM 1012 may store programs, utilities, or processes to be executed in a non-volatile manner. The RAM 1010 may provide volatile data storage and store instructions related to the operation of the computing device 1000. The computing device 1000 may also include a network / bus interface 1016 coupled to a data link 1018. The network / bus interface 1016 can be any type of device or system capable of enabling communication with an external device and / or a network, and can include but is not limited to a modem, a network card, an infrared communication device, a wireless communication device and / or a chipset (such as a Bluetooth™ device, a 1302.11 device, a WiFi device, a WiMax device, a cellular communication facility, etc.).
[0094] Various aspects, embodiments, implementations or features of the foregoing embodiments may be used alone or in any combination. Various aspects of the foregoing embodiments may be implemented by software, hardware or a combination of hardware and software.
[0095] For example, the foregoing embodiments may be embodied as computer readable code on a computer readable medium. A computer readable medium is any data storage device that can store data that can thereafter be read by a computer system. Examples of computer readable media include read-only memory, random access memory, CD-ROM, DVD, magnetic tape, hard drive, solid state drive, and optical data storage device. The computer readable medium may also be distributed in a network coupled computer system so that the computer readable code is stored and executed in a distributed manner.
[0096] For example, the aforementioned embodiments may be in the form of hardware circuits. The hardware circuits may include any combination of combinational logic circuits, clock storage devices (such as floppy disks, flip-flops, latches, etc.), finite state machines, memories such as static random access memories or embedded dynamic random access memories, custom designed circuits, programmable logic arrays, etc.
[0097] In one embodiment, a hardware circuit according to the present disclosure may be implemented by coding a circuit description in a hardware description language (HDL) such as Verilog or VHDL. The HDL description may be synthesized from a library of cells designed for a given integrated circuit manufacturing technology and may be modified for timing, power, and other reasons to obtain a final design database that may be transmitted to a factory to produce integrated circuits through a semiconductor manufacturing system. The semiconductor manufacturing system may produce integrated circuits by depositing semiconductor materials (e.g., on a wafer that may include a mask), removing materials, changing the shape of the deposited materials, modifying the materials (e.g., by doping materials or modifying dielectric constants using ultraviolet treatment), and the like. The integrated circuit may include transistors and may also include other circuit elements (e.g., passive elements such as capacitors, resistors, inductors, etc.) and interconnections between transistors and circuit elements. Some embodiments may implement multiple integrated circuits coupled together to implement the hardware circuit, and / or discrete elements may be used in some embodiments.
[0098] Although some specific embodiments of the present invention have been shown in detail by example, it will be appreciated by those skilled in the art that the above examples are intended to be illustrative only and do not limit the scope of the present invention. It will be appreciated by those skilled in the art that the above embodiments can be modified without departing from the scope and spirit of the present invention. The scope of the present invention is defined by the appended claims.
Claims
1. A text encoding device for network planning, comprising: Input unit, input text data to be encoded; A word segmentation unit, using a word segmentation word library storing words for word segmentation processing, performs word segmentation processing on the text data, and converts the result of the word segmentation processing into a word vector group including a plurality of W-dimensional word vectors through one-hot encoding; as well as A text encoding neural network model inputs the word vector group and outputs an N-dimensional vector as network business influence data, where W and N are both natural numbers greater than 1, and W>N; Wherein, the word vector is a W-dimensional row vector, the text encoding neural network model outputs an N-dimensional row vector, the text encoding neural network model includes a plurality of first matrices with W rows and H columns and a second matrix with H rows and N columns, the text encoding neural network model multiplies a plurality of the word vectors by a plurality of the first matrices respectively to obtain a plurality of H-dimensional row vectors, and multiplies a sum obtained by adding the plurality of H-dimensional row vectors by the second matrix to obtain the N-dimensional row vector; or The word vector is a W-dimensional column vector, and the text encoding neural network model outputs an N-dimensional column vector. The text encoding neural network model includes a plurality of first matrices with H rows and W columns and a second matrix with N rows and H columns. The text encoding neural network model multiplies the plurality of first matrices by a plurality of the word vectors to obtain a plurality of H-dimensional column vectors, and multiplies the second matrix by the sum of the plurality of H-dimensional column vectors to obtain the N-dimensional column vector.
2. The text encoding device for network planning according to claim 1, wherein: The text encoding neural network model also includes a nonlinear transformation component, which uses a Softmax function to map the elements of the N-dimensional vector to values greater than 0 and less than 1.
3. A text encoding device for network planning, comprising: An input unit for inputting text data and a network service change amount corresponding to the text data; A word segmentation unit, using a word segmentation word library storing words for word segmentation processing, performs word segmentation processing on the text data, and converts the result of the word segmentation processing into a word vector group including a plurality of W-dimensional word vectors through one-hot encoding; as well as The text encoding neural network model inputs the word vector group and outputs an N-dimensional vector, where W and N are both natural numbers greater than 1, and W>N, A training unit storing an N-dimensional training vector, multiplying the N-dimensional vector output from the text encoding neural network model by the training vector to obtain a training scalar, and training the text encoding neural network model in a manner such that the value of the training scalar is equal to the value of the network service variation; Wherein, the word vector is a W-dimensional row vector, the text encoding neural network model outputs an N-dimensional row vector, the training vector is an N-dimensional column vector, the text encoding neural network model includes a plurality of first matrices with W rows and H columns and a second matrix with H rows and N columns, the text encoding neural network model multiplies a plurality of the word vectors by a plurality of the first matrices respectively to obtain a plurality of H-dimensional row vectors, and multiplies a sum obtained by adding the plurality of H-dimensional row vectors by the second matrix to obtain the N-dimensional row vector; or The word vector is a W-dimensional column vector, the text encoding neural network model outputs an N-dimensional column vector, the training vector is an N-dimensional row vector, the text encoding neural network model includes a plurality of first matrices with H rows and W columns and a second matrix with N rows and H columns, the text encoding neural network model multiplies the plurality of first matrices by a plurality of the word vectors to obtain a plurality of H-dimensional column vectors, and multiplies the second matrix by the sum of the plurality of H-dimensional column vectors to obtain the N-dimensional column vector.
4. The network planning-oriented text encoding device according to claim 3, wherein: The text encoding neural network model also includes a nonlinear transformation component, which uses a Softmax function to map the elements of the N-dimensional vector to values greater than 0 and less than 1.
5. The network planning-oriented text encoding device according to claim 3, wherein: When training the text encoding neural network model, according to the loss function , the stochastic gradient descent method is used for training, where L is the training scalar and ΔL is the network service change.
6. A text encoding system for network planning, comprising: A training data storage unit storing training text data and a training network service variation corresponding to the training text data; A document receiving unit receives text data to be encoded; A word segmentation unit, using a word segmentation word library storing words for word segmentation processing, performs word segmentation processing on the training text data or the text data to be encoded, and converts the result of the word segmentation processing into a word vector group including a plurality of W-dimensional word vectors through one-hot encoding; A text encoding neural network model inputs the word vector group and outputs an N-dimensional vector, where W and N are both natural numbers greater than 1, and W>N; A text encoding neural network training unit stores an N-dimensional training vector, and when the word segmentation unit performs word segmentation processing on the training text data and outputs a corresponding word vector group to the text encoding neural network model, the N-dimensional vector output from the text encoding neural network model and the training vector are multiplied to obtain a training scalar, so that the value of the training scalar is equal to the value of the training network service variation, and the text encoding neural network model is trained; A text encoding unit, when the word segmentation unit performs word segmentation processing on the text data to be encoded and outputs a corresponding word vector group to the text encoding neural network model, outputs the N-dimensional vector output from the text encoding neural network model as network business influence data; as well as An output unit, outputting the network business influence data outputted from the text encoding unit to the outside; Wherein, the word vector is a W-dimensional row vector, the text encoding neural network model outputs an N-dimensional row vector, the training vector is an N-dimensional column vector, the text encoding neural network model includes a plurality of first matrices with W rows and H columns and a second matrix with H rows and N columns, the text encoding neural network model multiplies a plurality of the word vectors by a plurality of the first matrices respectively to obtain a plurality of H-dimensional row vectors, and multiplies a sum obtained by adding the plurality of H-dimensional row vectors by the second matrix to obtain the N-dimensional row vector; or The word vector is a W-dimensional column vector, the text encoding neural network model outputs an N-dimensional column vector, the training vector is an N-dimensional row vector, the text encoding neural network model includes a plurality of first matrices with H rows and W columns and a second matrix with N rows and H columns, the text encoding neural network model multiplies the plurality of first matrices by a plurality of the word vectors to obtain a plurality of H-dimensional column vectors, and multiplies the second matrix by the sum of the plurality of H-dimensional column vectors to obtain the N-dimensional column vector.
7. The text encoding system for network planning according to claim 6, wherein: The text encoding neural network model also includes a nonlinear transformation component, which uses a Softmax function to map the elements of the N-dimensional vector to values greater than 0 and less than 1.
8. The network planning-oriented text encoding system according to claim 6, wherein: When training the text encoding neural network model, according to the loss function , the stochastic gradient descent method is used for training, wherein L is the training scalar, and ΔL is the training network service variation.
9. A network planning system, comprising: An input unit, inputting network business influence data from the network planning-oriented text encoding device according to claim 1 or 2 or the network planning-oriented text encoding system according to any one of claims 6 to 8; A network planning unit, which plans resource allocation for each node on the network according to the network service influence data from the input unit; as well as The output unit outputs the planning result obtained by the network planning unit to the outside.
10. A text encoding method for network planning, comprising: Input step: input text data to be encoded; A word segmentation step, using a word segmentation word library storing words for word segmentation processing, performs word segmentation processing on the text data, and converts the result of the word segmentation processing into a word vector group including multiple W-dimensional word vectors through one-hot encoding; as well as An encoding step, running a text encoding neural network model that inputs the word vector group and outputs an N-dimensional vector, and outputs the N-dimensional vector as network business influence data, wherein W and N are both natural numbers greater than 1, and W>N; Wherein, the word vector is a W-dimensional row vector, the text encoding neural network model outputs an N-dimensional row vector, the text encoding neural network model includes a plurality of first matrices with W rows and H columns and a second matrix with H rows and N columns, the text encoding neural network model multiplies a plurality of the word vectors by a plurality of the first matrices respectively to obtain a plurality of H-dimensional row vectors, and multiplies a sum obtained by adding the plurality of H-dimensional row vectors by the second matrix to obtain the N-dimensional row vector; or The word vector is a W-dimensional column vector, and the text encoding neural network model outputs an N-dimensional column vector. The text encoding neural network model includes a plurality of first matrices with H rows and W columns and a second matrix with N rows and H columns. The text encoding neural network model multiplies the plurality of first matrices by a plurality of the word vectors to obtain a plurality of H-dimensional column vectors, and multiplies the second matrix by the sum of the plurality of H-dimensional column vectors to obtain the N-dimensional column vector.
11. A training method for a text encoding device for network planning, the training method comprising: An input step of inputting text data and a network service change amount corresponding to the text data; A word segmentation step, using a word segmentation word library storing words for word segmentation processing, performs word segmentation processing on the text data, and converts the result of the word segmentation processing into a word vector group including a plurality of W-dimensional word vectors through one-hot encoding; and The training step is to train a text encoding neural network model that inputs the word vector group and outputs an N-dimensional vector, wherein W and N are both natural numbers greater than 1, and W>N, In the training step, the N-dimensional vector output from the text encoding neural network model and the N-dimensional training vector are multiplied to obtain a training scalar, and the text encoding neural network model is trained in such a way that the value of the training scalar is equal to the value of the network service variation; Wherein, the word vector is a W-dimensional row vector, the text encoding neural network model outputs an N-dimensional row vector, the training vector is an N-dimensional column vector, the text encoding neural network model includes a plurality of first matrices with W rows and H columns and a second matrix with H rows and N columns, the text encoding neural network model multiplies a plurality of the word vectors by a plurality of the first matrices respectively to obtain a plurality of H-dimensional row vectors, and multiplies a sum obtained by adding the plurality of H-dimensional row vectors by the second matrix to obtain the N-dimensional row vector; or The word vector is a W-dimensional column vector, the text encoding neural network model outputs an N-dimensional column vector, the training vector is an N-dimensional row vector, the text encoding neural network model includes a plurality of first matrices with H rows and W columns and a second matrix with N rows and H columns, the text encoding neural network model multiplies the plurality of first matrices by a plurality of the word vectors to obtain a plurality of H-dimensional column vectors, and multiplies the second matrix by the sum of the plurality of H-dimensional column vectors to obtain the N-dimensional column vector.
Citation Information
Patent Citations
Self-media data text representation method based on recursion variational self-encoding model
CN108363685A
Text representation learning method and system based on automatic coding and electronic equipment
CN109582786A
Text word vector acquisition method and device, computer device and storage medium
CN109885826A
Text word vector model training method and device, terminal and storage medium
CN110795935A
Semisupervised autoencoder for sentiment analysis
US20180165554A1
