A lifelong learning method for dialogue state tracking

Abstract

The application provides a lifelong learning dialogue state tracking method, comprising the following steps: building a DST model, constructing a DST knowledge prompt template, fine-tuning the DST model, inputting the dialogue history, the user input sentence and the DST knowledge prompt template into the fine-tuned DST model to obtain the corresponding output, fine-tuning the DST model according to new business knowledge, and performing backward knowledge transfer. The lifelong learning dialogue state tracking method provided by the application can realize the rapid development of new businesses by combining the parameter-efficient fine-tuning method and the cross-task knowledge transfer.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of dialogue tracking, in particular to a lifelong learning dialogue state tracking method. BACKGROUND

[0002] Prompt-based, adapter-based or other parameter-efficient transfer learning methods have recently made amazing progress in dialogue state tracking (DST task) of dialogue systems, which is of great significance for the rapid expansion of business of dialogue systems in the industry. These methods do not directly fine-tune all the parameters of the pre-trained language model, but learn a small number of new parameters (1% or less of the parameter amount of the pre-trained model) on specific downstream tasks to achieve comparable or better results than fine-tuning the pre-trained model. In the deployment of dialogue systems, we can only deploy a pre-trained model, and then use the learned new parameters to complete various different tasks. However, in actual business, the business that our dialogue system undertakes will not come at the same time, so we need to constantly cope with new tasks. In such a scenario, we cannot completely require the model to quickly achieve good results on new tasks, and we need to transfer the knowledge learned by the model in the existing business scenarios to new tasks. Therefore, it is very necessary to design a lifelong learning dialogue state tracking method. SUMMARY

[0003] The purpose of the present application is to provide a lifelong learning dialogue state tracking method, which can realize the rapid development of new business by combining parameter-efficient fine-tuning methods and cross-task knowledge transfer.

[0004] To achieve the above purpose, the present application provides the following scheme:

[0005] A lifelong learning dialogue state tracking method comprises the following steps:

[0006] Step 1: build a DST model;

[0007] Step 2: construct a DST knowledge prompt template;

[0008] Step 3: fine-tune the DST model, input the dialogue history, user input sentence and DST knowledge prompt template into the fine-tuned DST model to obtain the corresponding output;

[0009] Step 4: fine-tune the DST model according to new business knowledge, and perform backward knowledge transfer.

[0010] Optionally, in step 2, the DST knowledge prompt template is constructed, specifically:

[0011] Constructing a DST knowledge prompt template, wherein the DST knowledge prompt template comprises a business name, a slot label required to be tracked by the DST, and a description statement thereof.

[0012] Optionally, in step 3, the DST model is fine-tuned, specifically as follows:

[0013] The main parameters of the DST model are kept unchanged, and the adapters parameters are adjusted, wherein the adapters parameters learned under each business are denoted as Different businesses are numbered according to the order of the plurality of businesses, to obtain wherein is the latest business, the plurality of businesses correspond to a plurality of adapters parameters, i.e. The parameter is obtained when the parameter needs to be calculated The business and the corresponding k dialogue histories are obtained, the similarity between the k dialogue histories and the expected dialogue histories of the previous n businesses is calculated through the sentence representation model, to obtain the similarity score of the business to the previous n businesses, wherein the similarity score on each business is the average of the similarity between the expected dialogue history of the business and the k dialogue histories of the business The n similarity scores are taken as weights to obtain The parameter is as follows:

[0014]

[0015] Optionally, in step 3, the dialogue history, the user input sentence, and the DST knowledge prompt template are input into the fine-tuned DST model to obtain the corresponding output, specifically as follows:

[0016] The dialogue history, the user input sentence, and the DST knowledge prompt template are input into the fine-tuned DST model, and the DST model generates the answer of each slot under the current dialogue history and the current input through a text generation method as the output.

[0017] Optionally, in step 4, the DST model is fine-tuned according to the new business knowledge to perform backward knowledge transfer, specifically as follows:

[0018] According to the data of the latest business, the historical business is improved, the business and the corresponding x dialogue histories are obtained, the similarity between the x dialogue histories and the expected dialogue histories of the previous n+1 businesses is calculated through the sentence representation model, to obtain the similarity score of the business to the previous n+1 businesses, wherein the similarity score is ​wherein the similarity score on each service is the average of the similarity of the x dialogue histories of the service to the dialogue history of the service with n+1 similarity scores as weights, obtaining The parameters are:

[0019]

[0020] The data of the service are used to train the parameters The data of task n+1 are used as regularization, and if The parameters perform well on the dev data of the service , then the original is replaced by .

[0021] According to the specific embodiments provided by the present application, the following technical effects are disclosed: the lifelong learning dialogue state tracking method provided by the present application, which comprises the following steps: building a DST model, constructing a DST knowledge prompt template, fine-tuning the DST model, inputting the dialogue history, the user input sentence and the DST knowledge prompt template into the fine-tuned DST model to obtain the corresponding output, fine-tuning the DST model according to the new business knowledge, and performing backward knowledge transfer. Through the combination of the parameter efficient fine-tuning method and the cross-task knowledge transfer, the method can make the model quickly reach good results on the new business, and can realize rapid development and online on the new task. BRIEF DESCRIPTION OF DRAWINGS

[0022] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0023] Figure 1 It is a process schematic diagram of the lifelong learning dialogue state tracking method of the embodiments of the present application.

[0024] Figure 2 It is a workflow diagram of the DST model of the embodiments of the present application on a single service (e.g., catering service).

[0025] Figure 3 It is a fine-tuning architecture diagram based on a pre-trained model + Adapter.

[0026] Figure 4 It is a knowledge transfer process diagram of the DST model of different tasks. DETAILED DESCRIPTION

[0027] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of protection of the present application.

[0028] The present application aims to provide a lifelong learning dialogue state tracking method, which can realize rapid development of new businesses by combining parameter-efficient fine-tuning method and cross-task knowledge transfer.

[0029] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the present application will be further described in detail below with reference to the drawings and specific embodiments.

[0030] As shown in the figure, the lifelong learning dialogue state tracking method provided by the embodiments of the present application comprises the following steps: Figure 1 Step 1: building a DST model;

[0031] Step 2: constructing a DST knowledge prompt template;

[0032] Step 3: fine-tuning the DST model, inputting the dialogue history, user input sentence and DST knowledge prompt template into the fine-tuned DST model to obtain the corresponding output;

[0033] Step 4: fine-tuning the DST model according to new business knowledge, and performing backward knowledge transfer.

[0034] Wherein, the meaning of DST is actually dialogue state tracking.

[0035] In step 2, the DST knowledge prompt template is constructed, specifically:

[0036] In order to provide DST knowledge prompts to the model, so that the generation model can generate the answers we want more accurately, the DST knowledge prompt template is constructed, wherein the DST knowledge prompt template includes the business name, the slot label and its description sentence required by the DST to be tracked.

[0037] As shown in the figure, in step 3, the DST model is fine-tuned, specifically:

[0038] Figure 3

[0039] ​​In each business, the DST model adopts adapters dedicated to each task to adapt the model to the task. Adapters are placed in each transformer layer of the T5 model (i.e., the DST model), similar to adding an additional residual connection to the fully connected layer or attention layer of the transformer;

[0040] The network structure of the adapter is relatively simple: the input intermediate layer representation first passes through a dimension-reducing fully connected layer, an activation function, and a dimension-increasing fully connected layer. Then, the part of the representation modified by the adapter is added to the T5 model backbone. Through such an architectural design, the adapter can ensure that it does not introduce too many new parameters.

[0041] The parameters of the adapters are adjusted while keeping the parameters of the main body of the DST model unchanged. Here, the adapters parameters learned under each business are denoted as Different businesses are numbered according to their order, obtaining For example, the first business to which the DST model is adapted is catering services, then comes the demand for a ride-hailing service, followed by hotel booking, online shopping, and various other businesses. This setting is consistent with the actual industrial scenario.

[0042] Here, is the latest business, as shown in Figure 4 Multiple businesses correspond to multiple adapters parameters, i.e., parameters.

[0043] When the parameters need to be calculated, the parameters can be directly randomly initialized and trained according to the existing data corresponding to the business However, this will result in a non-optimal effect.

[0044] Therefore, the business and the corresponding k dialogue histories are obtained, and the similarity between the k dialogue histories and the expected dialogue histories of the previous n businesses is calculated by a sentence representation model (e.g., SimCSE), obtaining the similarity scores of the business and the previous n businesses as Here, the similarity score of each business is the average of the similarity between its expected dialogue history and the k dialogue histories of the business Taking the n similarity scores as weights, the initialization of the adapters parameters of the business is the weighted average of the adapters parameters of the previous n tasks, i.e., The parameters are:

[0045]

[0046] service The adapter parameters of the service After the initialization according to the above steps, fine-tuning is performed on the data of the service The initialization method of the parameters contains the knowledge in the first n tasks.

[0047] In step 3, the dialogue history, user input sentence and DST knowledge prompt template are input into the fine-tuned DST model to obtain the corresponding output, specifically:

[0048] The dialogue history, user input sentence and DST knowledge prompt template are input into the fine-tuned DST model, and the DST model generates the answer of each slot under the current dialogue history and current input by text generation as the output.

[0049] The dialogue history, user input sentence and DST knowledge prompt template are spliced after being surrounded by <s>With< / s> symbols;

[0050] The present application constructs the DST task as a text completion task, and according to the input, the model needs to complete the following template "[1]___; [2] ___", note that the symbols "[1]" and "[2]" are consistent with the slot labels in the DST knowledge prompt, and the model needs to generate the answer of each slot under the current dialogue history and current input by text generation.

[0051] As shown in Figure 4 , in step 4, the DST model is fine-tuned according to the new service knowledge, and backward knowledge transfer is performed, specifically:

[0052] Suppose the service has accumulated a certain amount of DST data after going online, we can improve the performance of the previous service by the knowledge in the data of the new service, according to the latest service data, the historical service is improved, and the service and the corresponding x dialogue histories are obtained, the similarity between the x dialogue histories and the expected dialogue histories of the first n+1 services is calculated by the sentence representation model, and the similarity scores of the service and the first n+1 services are obtained , wherein the similarity score on each service is the average of the similarity between the expected dialogue history and the x dialogue histories of the service , and the n+1 similarity scores are used as weights to obtain The parameters are:

[0053]

[0054]

[0055] One embodiment of the present application is shown in Fig. Figure 2

[0056] The lifelong learning dialogue state tracking method provided by the present application comprises the following steps: building a DST model, constructing a DST knowledge prompt template, fine-tuning the DST model, inputting the dialogue history, the user input sentence and the DST knowledge prompt template into the fine-tuned DST model to obtain the corresponding output, fine-tuning the DST model according to new business knowledge, and performing backward knowledge migration. Through the combination of the parameter efficient fine-tuning method and the cross-task knowledge migration, the model can quickly achieve good results on new businesses, and can realize rapid development and online on new tasks.

[0057] The principles and implementation modes of the present application are described by applying specific examples in this paper, and the above description of the examples is only used to help understand the method of the present application and its core idea; meanwhile, for those skilled in the art, according to the idea of the present application, the specific implementation modes and application ranges will be changed. In conclusion, the content of the present description should not be understood as a limitation of the present application.​​​​​​​​

Claims

1. A dialogue state tracking method capable of lifelong learning, characterized in that, Includes the following steps: Step 1: Build the DST model; Step 2: Construct a DST knowledge prompt template; Step 3: Fine-tune the DST model by inputting the dialogue history, user input statements, and DST knowledge prompt templates into the fine-tuned DST model to obtain the corresponding output; Step 4: Fine-tune the DST model based on the new business knowledge and perform backward knowledge transfer; In step 3, the DST model is fine-tuned, specifically as follows: Keeping the core parameters of the DST model unchanged, the adapter parameters are adjusted. The adapter parameters learned for each business are denoted as Θ. Different businesses are numbered according to their order, resulting in task1, ..., task... n Among them, task n For the latest services, multiple services correspond to multiple adapters parameters, namely adapters1, ..., adapters n Parameters, when it is necessary to calculate adapters n+1 When obtaining parameters, retrieve the business task. n+1 And the corresponding k dialogue histories, using a sentence representation model to calculate the similarity between the k dialogue histories and the expected dialogue histories of the first n business tasks, to obtain the business task. n+1 The similarity scores with the first n business transactions are w1, ..., w n In this context, the similarity score for each business task is the expected dialogue history and the business task. n+1 The average similarity score of k dialogue history entries is used as the weight to obtain adapters. n+1 The parameters are: I n+1 =w1*Θ1+...+w n *I n / w1+...+w n。 2. The dialogue state tracking method capable of lifelong learning according to claim 1, characterized in that, In step 2, a DST knowledge hint template is constructed, specifically as follows: Construct a DST knowledge prompt template, which includes the business name, the slot label that DST needs to track, and its description.

3. The dialogue state tracking method capable of lifelong learning according to claim 2, characterized in that, In step 3, the dialogue history, user input statements, and DST knowledge hint template are input into the fine-tuned DST model to obtain the corresponding output, specifically: The dialogue history, user input statements, and DST knowledge hint templates are input into the fine-tuned DST model. The DST model generates the answer for each slot based on the current dialogue history and current input through text generation, which is then used as the output.

4. The dialogue state tracking method capable of lifelong learning according to claim 3, characterized in that, In step 4, the DST model is fine-tuned based on the new business knowledge, and backward knowledge transfer is performed, specifically as follows: Based on the latest business data, historical business tasks... j Improve and acquire business tasks j And the corresponding x dialogue histories, using a sentence representation model to calculate the similarity between the x dialogue histories and the expected dialogue histories of the previous n+1 business tasks, to obtain the business task. j The similarity scores with the first n+1 business transactions are y1, ..., y2. n+1 In this context, the similarity score for each business task is the expected dialogue history and the business task. j The average similarity score of x dialogue history entries is used as the weight to obtain adapters. j The parameters are: I' j =y1*Θ1+...+y n+1 *I n+1 / y1+...+y n+1 Through business tasks j Data for adapters j The parameters are trained using data from task n+1 as regularization. If adapters j Parameters in business tasks j If excellent performance is achieved on dev data, then Θ' is used. j Replace the original Θ j .

Images