Traffic semantic understanding model training method, traffic data multistage semantic analysis method, equipment, medium and program product

Abstract

The invention discloses a traffic semantic understanding model training method, a traffic data multi-level semantic analysis method, equipment, a medium and a program product, the training method adopts a three-stage training mechanism, and comprises the following steps: general pre-training: adopting a BERT-base Chinese pre-training model, constructing a basic language representation capability, inputting a word-level sequence into the model, and inputting the word-level sequence into the basic language representation capability; the output is a context semantic BERT vector; traffic domain knowledge injection: modeling by adopting a GNN algorithm model, and injecting a traffic domain ontology library with built-in traffic entity terms and semantic relationships thereof into BERT vectors; and scene fine tuning: performing fine tuning on the model based on a traffic instruction data set containing a real voice instruction, and the objective function being cross entropy loss. And performing multi-level semantic analysis on the to-be-analyzed text based on the traffic semantic understanding model obtained by training through the training method. According to the method, the semantic gap problem of a general model in a traffic scene is solved, the semantic barrier in the traffic field is broken, and the traffic transportation terminology recognition accuracy is improved to 90% or above.

Description

Technical Field

[0001] This invention relates to the fields of intelligent transportation systems and natural language processing technology, and in particular to a method for training a traffic semantic understanding model, a method for multi-level semantic analysis of traffic data, equipment, media, and program products. Background Technology

[0002] With the explosive growth of urban traffic data, how to quickly and accurately transform massive amounts of heterogeneous data into visual charts that can be used for decision-making has become one of the core challenges in the construction of Intelligent Transportation Systems (ITS). Traditional methods rely on manual dragging and dropping fields, selecting graphs, and adjusting parameters within Business Intelligence (BI) platforms, requiring an average of no less than eight steps, which is both time-consuming and requires users to have a high level of domain and tool knowledge.

[0003] In recent years, Natural Language Processing (NLP) technology has been introduced into the field of visualization, forming the sub-field of "Natural Language to Visualization" (NL2VIS). Representative solutions include Tableau's "Ask Data," Power BI's "Q&A," and Amazon QuickSight's voice query function. However, these general-purpose systems have exposed significant technical shortcomings when facing specialized traffic scenarios. These shortcomings include: the lack of a general language model that characterizes the traffic domain ontology, the inability to correctly identify specialized expressions such as "major accidents," "tidal flow lanes," and "capacity reduction rate," leading to errors in intent parsing or terminology mapping failures, creating a significant semantic gap, and consequently failing to provide effective visualizations for decision-making.

[0004] In summary, there is an urgent need for a traffic semantic understanding model and a multi-level semantic analysis method for traffic data in the field of transportation, in order to solve the problem of inaccurate identification of technical terms in the existing technologies. Summary of the Invention

[0005] In view of this, the present invention provides a method for training a traffic semantic understanding model, a method for multi-level semantic analysis of traffic data, an apparatus, a medium, and a program product, which at least partially solves the problems existing in the prior art.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] Firstly, the present invention provides a method for training a traffic semantic understanding model, the method employing a three-stage training mechanism, including:

[0008] General pre-training: The BERT-based Chinese pre-trained model is used to build basic language representation capabilities. The input of the model is a character-level sequence, and the output is a contextual semantic BERT vector.

[0009] Knowledge injection into the transportation domain: The GNN algorithm model is used to model the traffic domain ontology library containing built-in traffic entity terms and their semantic relationships, and then the BERT vector is injected into the traffic vector.

[0010] Scene fine-tuning: Based on a traffic instruction dataset containing real voice commands, the model is fine-tuned, with the objective function being cross-entropy loss.

[0011] As a further improvement to this invention, the algorithm formula for modeling using the GNN algorithm and injecting the traffic domain ontology library containing built-in traffic entity terms and their semantic relationships into the BERT vector is as follows:

[0012]

[0013] in:

[0014] Terminology output by BERT vector;

[0015] Domain ontology library

[0016] The set of neighboring nodes of the entity involved in the traffic accident;

[0017] [0,1], attention weights It measures the strength of semantic association between entities;

[0018] W: Neighborhood weighted aggregation

[0019] ReLU is a non-linear activation function.

[0020] The new representation after injection into the ontology is used for downstream classification purposes.

[0021] Furthermore, the real voice commands in the traffic command dataset include intent commands, which include query, comparison, drill-down, and / or adjustment.

[0022] Secondly, this invention provides a multi-level semantic analysis method for traffic data, the method comprising: The system receives user voice input, uses a speech recognition model based on an integrated traffic terminology dictionary to recognize the speech, and outputs the text as the text to be parsed.

[0023] The traffic semantic understanding model trained using the above-mentioned traffic semantic understanding model training method performs multi-level semantic parsing on the text to be parsed, identifies the user's query intent, extracts spatiotemporal parameters and traffic domain terms, and performs standardized mapping of terms based on a traffic domain ontology library to obtain semantic analysis results.

[0024] Furthermore, the speech recognition model adopts a speech recognition model based on an end-to-end streaming recognition architecture using Conformer;

[0025] The speech recognition process also includes a speech signal preprocessing step, which involves using a noise suppression algorithm based on RNNoise for speech recognition. This algorithm uses a gated loop unit to separate the speech signal from the background noise in terms of frequency bands, and constructs the following filtering model:

[0026]

[0027] in:

[0028] The input voice signal is at the 1st Frequency domain representation of frames; : The hidden state of the previous frame; : No. The weights of each GRU unit; : Bias term; : The output denoised speech signal.

[0029] Furthermore, the speech signal preprocessing step also includes: processing 50–200 speech signals through a band-stop filter bank. and 2–8 Noise in the frequency band is suppressed.

[0030] Furthermore, the multi-level semantic analysis also includes: processing the temporal and spatial information in the user's voice input using a spatiotemporal parameter standardization method that combines rules and machine learning, including:

[0031] Time standardization is performed by dynamically calculating the start and end dates based on the system's current timestamp.

[0032] And the geofencing extension method is used to map fuzzy place names to standard latitude and longitude areas for spatial standardization.

[0033] Thirdly, the present invention also provides a computer device, the device comprising: a processor and a memory;

[0034] The memory is used to store one or more program instructions;

[0035] The processor is configured to run one or more program instructions to perform the steps of the traffic semantic understanding model training method or the traffic data multi-level semantic analysis method described above.

[0036] Fourthly, the present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the traffic semantic understanding model training method or the traffic data multi-level semantic analysis method described above.

[0037] Fifthly, the present invention also provides a computer program product, the computer program product comprising computer program instructions, which, when executed by a processor, implement the steps of the traffic semantic understanding model training method or the traffic data multi-level semantic analysis method described above.

[0038] This invention utilizes a three-stage training mechanism to construct a traffic semantic understanding model, achieving high-precision recognition and standardized mapping of professional terms in the transportation field. Through multi-level semantic analysis, it enables the model to possess deep semantic understanding capabilities, solving the semantic gap problem of general models in traffic scenarios, breaking down semantic barriers in the transportation field, and improving the recognition accuracy of transportation professional terms to over 90%. Attached Figure Description

[0040] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings in the following description are merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.

[0041] Figure 1 This is a flowchart of a multi-level semantic analysis method for traffic data according to an embodiment of the present invention;

[0042] Figure 2 This is a flowchart of a traffic semantic understanding model training method according to an embodiment of the present invention;

[0043] Figure 3 This is a flowchart of the TSU-Model training process according to an embodiment of the present invention. Detailed Implementation

[0045] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0046] This invention aims to address the significant shortcomings of existing traffic data visualization technologies in terms of professional semantic understanding by proposing an implementation scheme that integrates traffic domain ontology knowledge with advanced natural language processing (NLP) technology.

[0047] Figure 1 This is a flowchart of a multi-level semantic analysis method for traffic data according to an embodiment of the present invention. Figure 1 As shown in this embodiment, a multi-level semantic analysis method for traffic data includes:

[0048] S100 receives user voice input, recognizes the speech based on a speech recognition model with an integrated traffic terminology dictionary, and outputs the text as the text to be parsed. It can also output the recognition confidence score when outputting the text.

[0049] Specifically, the speech recognition model adopts a speech recognition model based on an end-to-end streaming recognition architecture using Conformer.

[0050] Speech recognition also includes speech signal preprocessing steps, including speech recognition using a noise suppression algorithm based on RNNoise and processing 50–200 noise signals through a band-stop filter bank. and 2–8 Noise in the frequency band is suppressed.

[0051] Specifically, to improve the accuracy of speech recognition in complex traffic environments (such as vehicle-mounted and roadside command vehicles), this embodiment adopts a noise suppression algorithm based on RNNoise (Real-time Noise-suppression Network). This algorithm uses a gated recurrent unit (GRU) to separate the speech signal from the background noise in different frequency bands, constructing the following filtering model:

[0052]

[0053] in:

[0054] The input voice signal is at the 1st Frequency domain representation of frames; : The hidden state of the previous frame; : No. The weights of each GRU unit; : Bias term; : The output denoised speech signal.

[0055] In addition, regarding typical noise in traffic scenarios (engine noise concentrated between 50–200), Wind noise is concentrated between 2 and 8. We designed a band-stop filter bank to suppress the above frequency bands, which improved the signal-to-noise ratio (SNR) by an average of about 6.2 dB (measured data, based on the ISO 26262 vehicle voice standard test set).

[0056] The output text obtained after the above noise suppression preprocessing and speech recognition is used as the text to be parsed.

[0057] Step S200: Based on the Traffic Semantic Understanding Model (TSU-Model), perform multi-level semantic parsing on the text to be parsed, identify the user's query intent, extract spatiotemporal parameters and traffic domain terms, and perform standardized mapping of terms based on the traffic domain ontology to obtain semantic analysis results.

[0058] First, a traffic semantic understanding model needs to be trained.

[0059] Specifically, to overcome the low accuracy problem of general NLP models in traffic terminology recognition, this embodiment constructs a traffic semantic understanding model, employing a three-stage training mechanism, in conjunction with... Figure 2 , 3 As shown:

[0060] Phase 1: General Pre-training

[0061] A BERT-base (Bidirectional Encoder Representations from Transformers - base version) Chinese pre-trained model is used to construct basic language representation capabilities. The model input is a character-level sequence. The output is a context semantic vector:

[0062]

[0063] Phase Two: Knowledge Infusion in the Transportation Sector

[0064] Building an ontology library for the transportation sector ,

[0065] in:

[0066] : A collection of entities, containing 2,875 terms including types of traffic accidents, road facilities, and traffic regulations;

[0067] : Relation set, which defines the semantic relationships between entities (such as "belongs to", "cause", "located in" etc.).

[0068] In this embodiment, a GNN (Graph Neural Network) algorithm model is used to inject a traffic domain ontology (terms and their semantic relationships) into the BERT vectors, so that entities such as "major accidents" and "tidal lanes" can obtain consistent and context-related representations in the embedding space, thereby improving the accuracy of subsequent intent classification and term mapping.

[0069] The ontology structure is modeled using GNN, and the entity embedding is fused with the BERT output algorithm as follows:

[0070]

[0071] in:

[0072] The set of neighboring nodes of the entity involved in the traffic accident; Attention weights measure the strength of semantic associations between entities; W: Learnable parameter matrix. : Non-linear activation function (ReLU).

[0073] Regarding the above function:

[0074] Rectified Linear Unit (RPU) In this embodiment, ReLU is used for inter-layer activation of GNN, fine-tuning of TSU-Model, and output of graph decision network, balancing speed and gradient stability.

[0075] Phase Three: Task Fine-tuning

[0076] Based on a self-built traffic instruction dataset (containing 50,000 real voice instructions covering intents such as query, comparison, drill-down, and adjustment), the model was fine-tuned, with the objective function being cross-entropy loss.

[0077]

[0078] in: : True labeling (e.g., "regional comparison" intent); : Model prediction probability; : Sample size.

[0079] The final model achieved an accuracy of 95.1% on the traffic term recognition task, which is significantly better than the general BERT model (68.5%).

[0080] The traffic semantic understanding model trained using the three-stage training mechanism described above can be used to perform multi-level semantic parsing on the text to be parsed generated in step S100.

[0081] Furthermore, the temporal and spatial expressions in user input are often ambiguous, such as "the first two quarters of this year" or "the Zhongguancun area". This embodiment designs a spatiotemporal parameter standardization method based on a combination of rules and machine learning for the temporal and spatial information in user voice input.

[0082] (1) Time standardization

[0083]

[0084] in , The start and end dates of the first and second quarters of the current year are dynamically calculated based on the current system timestamp.

[0085] (2) Spatial standardization

[0086] A geofencing extension method is used to map fuzzy place names to standard latitude and longitude regions. For example:

[0087]

[0088] in:

[0089] ( ): Coordinates of the center point of Zhongguancun (116.316E, 39.989N);

[0090] Extended radius, default setting is 0.01 degrees (approximately 1 kilometer).

[0091] The aforementioned geofence extension method (GEM) is a type of boundary extrapolation or adaptive expansion algorithm introduced to overcome problems such as positioning jitter, false alarms, and insufficient coverage caused by rigid "single-point-single-radius" boundaries. In this invention, the specific implementation is as follows: given keywords in the user's spoken language, they are first parsed into latitude and longitude centers (x0, y0), and then a circular extension area is generated using the Haversine spherical distance formula.

[0092] This invention constructs a traffic semantic understanding model to achieve high-precision identification and standardized mapping of professional terms such as "major accident", "tidal lane" and "capacity reduction rate", thus solving the semantic gap problem of general models in traffic scenarios.

[0093] Furthermore, this embodiment also provides a computer device, the device including: a processor and a memory; the memory is used to store one or more program instructions; the processor is used to run one or more program instructions to perform the steps of the traffic semantic understanding model training method or the traffic data multi-level semantic analysis method described above.

[0094] In addition, this embodiment also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the above-described traffic semantic understanding model training method or traffic data multi-level semantic analysis method.

[0095] Furthermore, this embodiment also provides a computer program product, which includes computer program instructions that, when executed by a processor, implement the steps of the above-described traffic semantic understanding model training method or traffic data multi-level semantic analysis method.

[0096] In this embodiment of the invention, the processor can be an integrated circuit chip with signal processing capabilities. The processor can be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in this embodiment of the invention. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in this embodiment of the invention can be directly implemented by a hardware decoding processor, or implemented by a combination of hardware and software modules in the decoding processor. The software modules can be located in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. The processor reads information from the storage medium and, in conjunction with its hardware, completes the steps of the above methods. The storage medium can be memory, for example, volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. Non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. Volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as Static Random Access Memory (SRAM), Dynamic Random Access Memory (DRAM), Synchronous DRAM (SDRAM), Double Data Rate SDRAM (DDRSDRAM), Enhanced Synchronous DRAM (ESDRAM), Synchlink DRAM (SLDRAM), and Direct Rambus RAM (DRRAM).The storage media described in the embodiments of this invention are intended to include, but are not limited to, these and any other suitable types of memory. Those skilled in the art will recognize that the functions described in one or more of the above examples can be implemented using a combination of hardware and software. When applied software, the corresponding functions can be stored in a computer-readable medium or transmitted as one or more instructions or code on a computer-readable medium. Computer-readable media include computer storage media and communication media, wherein communication media include any medium that facilitates the transmission of computer programs from one place to another. Storage media can be any available medium accessible to general-purpose or special-purpose computers. Although the invention has been described in detail above with general description and specific embodiments, modifications or improvements can be made to it, which will be apparent to those skilled in the art. Therefore, all such modifications or improvements made without departing from the spirit of the invention are within the scope of protection claimed by this invention.

[0097] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Any simple modifications, equivalent changes, or alterations made by those skilled in the art using the disclosed technical content shall fall within the protection scope of the present invention.

Claims

1. A method for training a traffic semantic understanding model, characterized in that, The method employs a three-stage training mechanism, including: General pre-training: The BERT-based Chinese pre-trained model is used to build basic language representation capabilities. The input of the model is a character-level sequence, and the output is a contextual semantic BERT vector. Knowledge injection into the transportation domain: The GNN algorithm model is used for modeling, and the transportation domain ontology library containing built-in transportation entity terms and their semantic relationships is injected into the BERT vector; Scene fine-tuning: Based on a traffic instruction dataset containing real voice commands, the model is fine-tuned, with the objective function being cross-entropy loss.

2. The traffic semantic understanding model training method according to claim 1, characterized in that, The algorithm formula for modeling using the GNN algorithm and injecting the traffic domain ontology library containing built-in traffic entity terms and their semantic relationships into the BERT vector is as follows: in: Terminology output by BERT vector; Ontology library for the transportation sector; The set of neighboring nodes of the entity involved in the traffic accident; [0,1], attention weights It measures the strength of semantic association between entities; W: Learnable and trainable weight parameter matrix; Neighborhood-weighted aggregation; ReLU is a non-linear activation function. The new representation after injection into the ontology is used for downstream classification purposes.

3. The traffic semantic understanding model training method according to claim 1, characterized in that, The real voice commands in the traffic command dataset include intent commands, which include query, comparison, drill-down, and / or adjustment.

4. A multi-level semantic analysis method for traffic data, characterized in that, The method includes: It receives user voice input, recognizes the speech based on a speech recognition model with an integrated traffic terminology dictionary, and outputs the text as the text to be parsed. The traffic semantic understanding model trained based on the traffic semantic understanding model training method of any one of claims 1-3 performs multi-level semantic parsing on the text to be parsed, identifies the user's query intent, extracts spatiotemporal parameters and traffic domain terms, and performs standardized mapping of terms based on a traffic domain ontology library to obtain semantic analysis results.

5. The multi-level semantic analysis method for traffic data according to claim 4, characterized in that, The speech recognition model adopts a Conformer-based end-to-end streaming recognition architecture. The speech recognition process also includes a speech signal preprocessing step, which involves using a noise suppression algorithm based on RNNoise for speech recognition. This algorithm uses a gated loop unit to separate the speech signal from the background noise in terms of frequency bands, and constructs the following filtering model: in: The input voice signal is at the 1st Frequency domain representation of frames; : The hidden state of the previous frame; : No. The weights of each GRU unit; : Bias term; : The output denoised speech signal.

6. The multi-level semantic analysis method for traffic data according to claim 4 or 5, characterized in that, The speech signal preprocessing step further includes: passing the speech signal through a band-stop filter bank for 50–200 and 2–8 Noise in the frequency band is suppressed.

7. The multi-level semantic analysis method for traffic data according to claim 4 or 5, characterized in that, The multi-level semantic analysis also includes: processing the temporal and spatial information in the user's voice input using a spatiotemporal parameter standardization method that combines rules and machine learning, including: Time standardization is performed by dynamically calculating the start and end dates based on the system's current timestamp. And the geofencing extension method is used to map fuzzy place names to standard latitude and longitude areas for spatial standardization.

8. A computer device, characterized in that, The device includes: a processor and a memory; The memory is used to store one or more program instructions; The processor is configured to run one or more program instructions to perform the steps of a traffic semantic understanding model training method as described in any one of claims 1 to 3 or a traffic data multi-level semantic analysis method as described in any one of claims 4 to 7.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of a traffic semantic understanding model training method as described in any one of claims 1 to 3 or a traffic data multi-level semantic analysis method as described in any one of claims 4 to 7.

10. A computer program product, characterized in that, The computer program product includes computer program instructions that, when executed by a processor, implement the steps of a traffic semantic understanding model training method as described in any one of claims 1 to 3 or a traffic data multi-level semantic analysis method as described in any one of claims 4 to 7.

Images