End-to-end task-oriented dialogue method and system

Abstract

The application discloses an end-to-end task-oriented dialogue method and system, which is constructed based on a self-attention mechanism, and comprises the following steps: S1, constructing a user simulator, using a pre-trained language model to realize corresponding natural language understanding, strategy action and natural language generation according to a dialogue target; S2, constructing a dialogue system, using a pre-trained language model to realize corresponding belief state tracking, strategy action and natural language generation; S3, constructing a task-oriented dialogue system framework, and jointly updating the user simulator and the system, so that a high dialogue success rate is achieved. The application can effectively construct an end-to-end user simulator and dialogue system, and meanwhile, a high dialogue success rate of the user simulator and the system can be achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of natural language processing, in particular to an end-to-end task-oriented dialogue method and system. BACKGROUND

[0002] Dialogue artificial intelligence has been a long-term exploration topic in the field of natural language processing (NLP) in computer science. If NLP is the brightest pearl on the crown of artificial intelligence, then the dialogue system is the brightest pearl on the crown of NLP. With the development and progress of neural networks, dialogue systems have recently received more attention in academia and industry.

[0003] Dialogue tasks can be roughly divided into two categories: open-domain dialogue systems focus on chatting with humans, such as various voice assistants: Xiaoyi, Xiaoai, Xiaodu, Siri, etc., which are very natural and fascinating, and they are usually trained end-to-end using large-scale data from social media; and task-oriented dialogue (TOD) systems need to complete the related field goals described by the user in natural language, which are commonly used in public transportation ticketing systems, flight and hotel reservation systems, etc. TOD tasks are usually divided into three subtasks: understanding user input information, deciding to generate actions, and generating natural language responses, and a pipeline modular method is usually used to train specific data sets. The pipeline modular method includes a natural language understanding module to track the belief state, a dialogue policy module to determine which actions to take based on the beliefs, and a natural language generation module to generate natural language responses to communicate with the user based on the generated actions.

[0004] In a task-oriented dialogue system, the user has a goal to be completed, and the system provides an interface to access an external database. Both users can only obtain relevant information from each other through natural language communication. A task-oriented dialogue system is shown in Figure 1 The user should express the goal (Goal, G) completely in an organized manner, which usually includes constraints (Constraint, C) and requirements (Request, R). For example, constraints C (a Chinese restaurant in the city center) and requirements R (such as querying the address, hotel phone number) are composed, and there can be multiple domains in the goal G (such as restaurant, tourist attraction, taxi, hotel, train, etc.). The system should accurately and timely respond to useful information using its external database, and given all candidate entities that meet the conditions and the corresponding information. The user and the system communicate with each other through natural language in the conversation, transfer information, and ultimately achieve the user's goal.

[0005] Much of the research work on task-oriented dialogues has focused on optimizing specific modules and evaluating only that module. These modules include understanding user intent through intent detection, tracking constraints provided by the user through dialog state tracking, determining dialog system actions through dialog policy determination, and using dedicated natural language response generation components. However, the dependencies between these independent modules make the pipeline approach vulnerable to error accumulation caused by cross-component error propagation, and fine-tuning of each component module is required, which can be done independently, but it is difficult to ensure that the entire dialog system is optimal.

[0006] Recent approaches have increasingly turned to end-to-end solutions, aiming to reduce the design of human and specific task modules. SimpleTOD, proposed by Salesforce Research, directly treats all task-oriented dialogues as a single sequence prediction problem, uses a single pre-trained language model, and trains through a single, joint, multi-task loss to achieve advanced performance in dialog state tracking, action decision, and response generation indicators in an end-to-end setting. SUMMARY

[0007] To make up for the deficiencies of the above background art, the present application proposes an end-to-end task-oriented dialogue method and system to solve the problem of error accumulation and inability to reach global optimization in modular optimization of task-oriented dialogue.

[0008] The technical problem of the present application is solved by the following technical solution:

[0009] The present application discloses an end-to-end task-oriented dialogue method based on a self-attention mechanism, comprising the following steps:

[0010] S1, a user simulator is constructed, and a pre-trained language model is used to realize corresponding natural language understanding, policy action, and natural language generation functions according to the dialogue target;

[0011] S2, a dialogue system is constructed, and a pre-trained language model is used to realize corresponding belief state tracking, policy action, and natural language generation functions;

[0012] S3, a task-oriented dialogue system framework is constructed, and the user simulator and the system are jointly updated, thereby achieving a high dialogue success rate.

[0013] In some embodiments, in step S1, the user simulator is a goal-oriented end-to-end dialogue user simulator, which is constructed into a sequence prediction problem of a user simulator using a natural language model based on a self-attention mechanism.

[0014] In some embodiments, in step S1, for the sequence prediction problem of the user simulator, the dialogue of each round of the user simulator is decomposed into five parts: User_Context, NLU, Dynamic_Goal, User_Policy, and User_NLG; wherein User_Context is the natural language dialogue between the user simulator and the system in each round, NLU is the understanding of the action made by the user simulator to the system, Dynamic_Goal is the dynamic goal state in the dialogue process, User_Policy is the action made by the user simulator in this round, and User_NLG is the natural language generated by the user simulator according to the action, which is Delex_User_Utterance, and Lex_User_Utterance is generated in combination with the goal information.

[0015] In some embodiments, in step S2, the dialogue system is an end-to-end dialogue system; and a natural language model based on a self-attention mechanism is used to construct the dialogue system into a sequence prediction problem of the dialogue system.

[0016] In some embodiments, for the sequence prediction problem of the dialogue system, the dialogue of each round of the dialogue system is decomposed into four parts: Sys_Context, DST, DB, Sys_Policy, and Sys_NLG; wherein Sys_Context is the natural language dialogue between the user simulator and the system in each round, DST is the understanding of all useful belief states of the user by the system, DB is the search result of the external database, Sys_Policy is the action made by the system in this round, and Sys_NLG is the natural language generated by the system according to the action, which is Delex_Sys_Response, and Lex_Sys_Response is generated in combination with the belief state BeliefState information.

[0017] In some embodiments, in step S3, the task-oriented dialogue system framework is directed to the goal of the user simulator, the user simulator first performs understanding, then generates corresponding actions and natural language through a strategy, the system performs belief state tracking through the user language of the user and generates corresponding actions and system responses, the two interact through natural language, the system side has an external database, and the two interact to complete the corresponding goal.

[0018] In some embodiments, step S3 includes an optimization step, which adopts the Transformer model and sets the Transformer model of the user simulator to User_Transformer and the Transformer model of the dialogue system to Sys_Transformer.

[0019] In some embodiments, step S3 includes an optimization step, wherein, for the user simulator, in round t, the user simulator generates user language U. t Construct historical text C for the user simulator u_t =[S0,U0,S1,U1......S t-1 Natural Language Understanding (NLU) is abbreviated as NLU. u_t The user's dynamic status is Goal u_t User policy is denoted as POL u_t The entire process can be summarized as follows:

[0020]

[0021] In some embodiments, step S3 includes an optimization step, wherein, for the system, in round t, the system generates a system response S. t Constructing historical text C for dialogue systems s_t =[U0,S0,U1,S1,......U t Belief state tracking is denoted as DST. s_t External database search is DB s_t The system policy is denoted as POL. s_t The entire process can be summarized as follows:

[0022]

[0023] The present invention also discloses an end-to-end task-oriented dialogue system constructed using the method described in any of the preceding claims.

[0024] The beneficial effects of this invention compared to the prior art include:

[0025] The end-to-end task-oriented dialogue method provided by this invention utilizes a pre-trained language model for the user simulator to achieve corresponding natural language understanding, policy actions, and natural language generation functions based on the dialogue objective; for the dialogue system, it utilizes the pre-trained language model to achieve corresponding belief state tracking, policy actions, and natural language generation functions; then, it constructs a task-oriented dialogue system framework and jointly updates the user simulator and the system, thereby achieving a high dialogue success rate. This invention can effectively construct an end-to-end user simulator and dialogue system, while achieving a high dialogue success rate for both the user simulator and the system. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of a task-based dialogue system based on the prior art of this invention.

[0027] Figure 2 This is a flowchart of an end-to-end task-based dialogue method according to an embodiment of the present invention.

[0028] Figure 3 This is an end-to-end model diagram of the user simulator in an embodiment of the present invention.

[0029] Figure 4 This is an end-to-end model diagram of the dialogue system according to an embodiment of the present invention.

[0030] Figure 5 This is a framework diagram of a task-oriented dialogue system according to an embodiment of the present invention. Detailed Implementation

[0031] The present invention will be further described below with reference to the accompanying drawings and preferred embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other.

[0032] It should be noted that the directional terms such as left, right, up, down, top, and bottom used in this embodiment are only relative concepts or are based on the normal use of the product, and should not be considered as restrictive.

[0033] To overcome the errors and inability to reach the global optimum in modular optimization of task-oriented dialogue, and to improve the success rate of dialogue between the user simulator and the system, this invention provides an end-to-end task-oriented dialogue method based on a self-attention mechanism (Transformer). This method can effectively construct an end-to-end user simulator and dialogue system. Simultaneously, it achieves a high success rate in dialogue between the user simulator and the system, providing a foundational framework for subsequent Transformer-based task-oriented dialogue.

[0034] The core problems to be solved in the embodiments of the present invention include: (1) how to use a pre-trained language model to realize the corresponding natural language understanding, policy action and natural language generation functions according to the dialogue goal for the user simulator; (2) how to use a pre-trained language model to realize the corresponding belief state tracking, policy action and natural language generation functions for the dialogue system; (3) how to select an optimization method to jointly update the user simulator and the system, so as to achieve a higher dialogue success rate.

[0035] like Figure 2 As shown, the end-to-end task-oriented dialogue method of this invention includes the following steps:

[0036] S1, construct a user simulator, use a pre-trained language model to realize corresponding natural language understanding, policy action, natural language generation function according to the dialogue target.

[0037] Specifically, the construction of the user simulator has been a widely concerned problem in task-oriented dialogue, but it is difficult to evaluate, and how to change the corresponding strategy generation according to the change of the target is a problem worthy of study. The user simulator of the embodiment of the application is a target-oriented end-to-end dialogue user simulator, which uses a natural language model based on a self-attention mechanism (GPT-2 model) to construct it into a sequence prediction problem of a user simulator. It includes the following functions: natural language understanding, policy action, natural language generation. The end-to-end model of the user simulator is as shown in Figure 3

[0038] Among them, for the sequence prediction problem, the dialogue of the user simulator in each round can be divided into five parts: User_Context, NLU, Dynamic_Goal, User_Policy and User_NLG. Among them, User_Context is the natural language dialogue between the user simulator and the system in each round, NLU is the understanding of the action made by the user simulator to the system, Dynamic_Goal is the dynamic goal state in the dialogue process, User_Policy is the action made by the user simulator in this round, and User_NLG is the natural language generated by the user simulator according to the action. The generated natural language is a delexicalized user language (Delex_User_Utterance), and the lexicalized user language (Lex_User_Utterance) is generated in combination with the goal (Goal) information. An example of the dialogue of the user simulator in each round is as follows:

[0039] User history text: <sos_sr> <eos_sr> <sos_uu> I need to book a train that leaves after 13:45 on Tuesday. <eos_uu> <sos_sr> What is your destination? <eos_sr>

[0040] Natural language understanding: <sos_sa> [act_request] [train_destination] <eos_sa>

[0041] ​Dynamic goal: <sos_g>restaurant_info_area restaurant_info_food restaurant_book_day restaurant_book_people restaurant_book_time train_info_departure train_info_destination train_reqt_duration train_reqt_price<eos_g>

[0042] Policy action: <sos_ua>[act_inform][train_day][train_leaveAt]<eos_ua>

[0043] Natural language generation: <sos_uu>i need to book a train that leaves after[train_leaveAt]on[train_day].<eos_u>.

[0044] S2, a dialogue system is constructed, and a pre-trained language model is used to realize corresponding belief state tracking, policy action, and natural language generation functions.

[0045] Specifically, the construction of the dialogue system has been a widely concerned problem in task-based dialogue, but how to track the corresponding belief state and generate the corresponding response according to the feedback of the user simulator is a problem worthy of study. The dialogue system of the embodiment of the present application is an end-to-end dialogue system, which uses a natural language model (GPT-2) model based on a self-attention mechanism to construct it into a sequence prediction problem. It includes the following functions: belief state tracking, policy action, and natural language generation. The end-to-end model of the dialogue system is as shown in Figure 4 .

[0046] wherein, for a sequence prediction problem, the dialogue of the dialogue system for each turn can be decomposed into four parts: Sys_Context, DST, DB, Sys_Policy, and Sys_NLG. The Sys_Context is the natural language dialogue between the user simulator and the system for each turn, the DST is the understanding of all useful belief states of the user by the system, the DB is the search result of the external database, the Sys_Policy is the action made by the system for this turn, and the Sys_NLG is the natural language generated by the system according to the action, the generated natural language is the delexicalized system response (Delex_Sys_Response), and the belief state (BeliefState) information is used to generate the lexicalized user language (Lex_Sys_Response). An example of the dialogue of the dialogue system is as follows:

[0047] System history text: <sos_uu> I need to book a train that leaves after 13:45 on Tuesday. <eos_uu> <sos_sr> What is your destination? <eos_sr> <sos_uu> I am going to Liverpool Street in London. <eos_uu>

[0048] Belief state tracking: <sos_b> [train_destination] London Liverpool Street [train_day] Tuesday [train_leaveAt] 13:45 <eos_b>

[0049] Database:

[0050] <sos_db> [restaurant_db_0] [hotel_db_0] [attraction_db_0] [train_db_2] <eos_db>

[0051] Policy action: <sos_sa> [act_inform] [train_choice] <eos_sa>

[0052] [train_departure] [train_destination] [train_leaveAt] [train_trainID] <eos_sa>

[0053] Natural language generation: <sos_sr> The first train out of [train_departure] <eos_sr>

[0054] after [value_count], going to [train_destination], is the [train_id], which leaves at [train_leaveAt]. <eos_sr>.

[0055] S3, construct a task-oriented dialogue system framework, and update the user simulator and the system jointly, so that a higher dialogue success rate is achieved.

[0056] Specifically, for the above-mentioned goal-oriented end-to-end dialogue user simulator and end-to-end dialogue system, a constructed end-to-end task-oriented dialogue system framework (Transformer-based Framework of Task-Oriented Dialogue, TFTOD) is as shown in the figure. Figure 5 Under the framework, for the goal of the user simulator, the user simulator first performs understanding, then generates corresponding actions and natural language through strategy, the system performs belief state tracking through the user language of the user and generates corresponding actions and system responses, both through natural language interaction, the system side has an external database, and both interact to complete the corresponding goal.

[0057] Among them, the following optimization method is used to update the user simulator and the system jointly:

[0058] In the embodiment of the application, the dialogue usually includes many rounds, at round t, the user simulator generates user language U t , the system generates system response S t ; the historical text C u_t =[S0,U0,S1,U1......S t-1 ] is constructed for the user simulator, and the historical text C s_t =[U0,S0,U1,S1,......U t ] is constructed for the dialogue system. Meanwhile, the Transformer model of the user simulator is set as User_Transformer, and the Transformer model of the dialogue system is set as Sys_Transformer.

[0059] For the user simulator, at round t, the natural language understanding is denoted as NLU u_t , the user dynamic state is Goal u_t , and the user strategy is denoted as POL u_t , and the whole process is summarized as:

[0060]

[0061] The sequence prediction process using Transformer implements natural language understanding, policy actions, and natural language generation functionalities. The specific implementation is as follows: Figure 2 As shown.

[0062] For the system, in round t, the belief state tracking is denoted as DST. s_t External database search is DB s_t The system policy is denoted as POL. s_t The entire process can be summarized as follows:

[0063]

[0064] The sequence prediction process using Transformer implements functions such as belief state tracking, policy actions, and natural language generation. The specific implementation is as follows: Figure 3 As shown.

[0065] For the input sequence x of the user simulator during training... u_t and the input sequence x of the dialogue system s_t The way it is constructed is x u_t =[C u_t NLU u_t Goal u_t POL u_t U t ], x s_t =[C s_t DST s_t DB s_t POL s_t ,S t The learning objective of a language model is to learn the transition probability distribution p(x) between words. This distribution can be naturally decomposed using the chain rule of probability calculation. A neural network θ is trained with parameters to minimize the negative log-likelihood function on dataset D. For the application of the pre-trained language model Transformer, an end-to-end framework was implemented, thereby solving the problems of error accumulation, error propagation, and inability to reach the global optimum in modular optimization.

[0066] For joint optimization updates, the loss function L for the user simulator is... user The loss function L of the dialogue system sys The overall loss function L constitutes user +L sysEmbodiments of the present application implement a user simulator and a dialog system model using the Huggingface Transformers library of version 4.2.2. We initialize it using DistilGPT-2 (a distilled version of GPT-2). During supervised learning, the batch size is set to 2, the gradient accumulation step is 16, the AdamW optimizer and linear scheduler are used, there are 20 warm-up steps and the maximum learning rate is set to 1e-4, and the gradient (clip) is set to 5. The total training round is 30 (it takes about 20 hours on an NVIDIA Tesla 2V100-SXM2-32GB).

[0067] The present application also discloses an end-to-end task-oriented dialog system constructed by the end-to-end task-oriented dialog method according to any one of the above.

[0068] The performance of the end-to-end task-oriented dialog method of the embodiments of the present application is analyzed as follows.

[0069] A commonly used public dataset for task-oriented dialog is MultiWoz. MultiWoz is a multi-domain, multi-intent task-oriented dialog corpus, which contains 7 domains, 13 intents, 25 slot value types, 10483 dialog sessions and 71544 dialog turns. During data collection, users are required to follow the pre-specified user goal and are allowed to change the goal during the conversation if necessary, so the collected dialog is closer to real-world dialog. The corpus also provides domain knowledge that defines all entities and attributes as external databases.

[0070] There are two ways to evaluate the success rate of task-oriented dialog: one is to evaluate the proposed model on the MultiWoz test set according to the notification rate (Info), the success rate (Succ) and the machine translation evaluation index (BLEU). The notification rate measures whether the dialog system provides correct entities matching the user's goal, while the success rate further requires the system to correctly answer all user questions, and BLEU evaluates the fluency of natural language. Recently, there is also a related standard combined performance score (Combined Score), which is calculated as 0.5*(Info+Succ)+BLEU. The second is to use 1000 dialog goals for automatic evaluation. When the dialog starts, the user simulator and the system interact with each other around a given user goal. The interaction performance between two trained strategies is estimated by two indicators: the notification rate and the matching rate Match to estimate the success rate of the dialog. Both indicators are calculated at the dialog action level. Info evaluates whether all requested information has been informed, and the matching rate checks whether the booked entity matches all the indicated constraints given by the user. And only when the notification recall rate and the matching rate are both 1, the overall task success is achieved. Among them, the dialog success rate is generally considered the most important indicator.

[0071] The experimental results of the embodiments of the present application are shown in Table 1:

[0072] Table 1 Performance of the algorithm TFTOD in the official MultiWoz test set

[0073]

[0074] Compared with other Transformer-based task-oriented dialogue methods, the end-to-end task-oriented dialogue method of the embodiments of the present application is different in the representation of actions, belief states, database search results and special words. The end-to-end task-oriented dialogue method of the embodiments of the present application realizes the highest dialogue success rate and combined performance score based on the MultiWoz test set, and realizes the current Transformer-based dialogue system. At the same time, aiming at the success rate of the dialogue between the user simulator and the system, the highest success rate of the current task-oriented dialogue is also realized, as shown in Table 2:

[0075] Table 2 Performance of the algorithm TFTOD in the official MultiWoz test set user simulator system interactive dialogue

[0076]

[0077] The above is a further detailed description of the present application in combination with specific preferred embodiments, and the specific implementation of the present application cannot be limited to these descriptions. For those skilled in the art to which the present application belongs, without departing from the concept of the present application, a number of equivalent alternatives or obvious modifications can be made, and the performance or use is the same, which should be regarded as belonging to the protection scope of the present application.

Claims

1. An end-to-end task-oriented dialogue method, characterized by, It is constructed based on a self-attention mechanism, including the following steps: S1, constructing a user simulator, using a pre-trained language model to realize corresponding natural language understanding, policy action, natural language generation function according to the dialogue target; In step S1, the user simulator is a goal-oriented end-to-end dialogue user simulator, which is constructed into a user simulator sequence prediction problem by using a natural language model based on a self-attention mechanism; For the sequence prediction problem of the user simulator, the dialogue of the user simulator in each round is divided into five parts: User_Context, NLU, Dynamic_Goal, User_Policy, and User_NLG; Wherein User_Context is the natural language dialogue between the user simulator and the system in each round, NLU is the understanding of the action of the user simulator to the system, Dynamic_Goal is the dynamic goal state in the dialogue process, User_Policy is the action made by the user simulator in this round, and User_NLG is the natural language generated by the user simulator according to the action, and the generated natural language is Delex_User_Utterance, and Lex_User_Utterance is generated by combining the goal information; S2, constructing a dialogue system, using a pre-trained language model to realize corresponding belief state tracking, policy action, natural language generation function; In step S2, the dialogue system is an end-to-end dialogue system; It is constructed into a sequence prediction problem of a dialogue system by using a natural language model based on a self-attention mechanism; For the sequence prediction problem of the dialogue system, the dialogue of the dialogue system in each round is divided into four parts: Sys_Context, DST, DB, Sys_Policy, and Sys_NLG; Wherein Sys_Context is the natural language dialogue between the user simulator and the system in each round, DST is the understanding of all useful belief states of the user by the system, DB is the search result of the external database, Sys_Policy is the action made by the system in this round, and Sys_NLG is the natural language generated by the system according to the action, and the generated natural language is Delex_Sys_Response, and the belief state BeliefState information is used to generate Lex_Sys_Response; S3, construct a task-oriented dialogue system framework, and update the user simulator and the system jointly, so as to achieve a high dialogue success rate; the step S3 includes an optimization step, a Transformer model is adopted, and the Transformer model of the user simulator is set as User_Transformer, and the Transformer model of the dialogue system is set as Sys_Transformer; for the user simulator, at the turn t, the user simulator generates the user language U t , the historical text C u_t is constructed for the user simulator t-1 , natural language understanding is denoted as NLU u_t , the user dynamic state is Goal u_t , and the user policy is denoted as POL u_t , and the whole process is summarized as: NLU u_t = User_Transformer([C u_t ]) POL u_t = User_Transformer([C u_t , NLU u_t , Goal u_t ]) U t = User_Transformer([C u_t , NLU u_t , Goal u_t , POL u_t ]); For the system, at turn t, the system generates a system response S t , constructs a history text C s_t = [U0, S0, U1, S1,... U t ] for the dialog system, belief state tracking is denoted as DST s_t , external database search is denoted as DB s_t , system policy is denoted as POL s_t , and the whole process is summarized as: DST s_t = Sys_Transformer([C s_t ]) POL s_t = Sys_Transformer([C s_t , DST s_t , DB s_t ]) S t = Sys_Transformer([C s_t ,DST s_t ,DB s_t ,POL s_t ]) ; where, for the joint optimization update, the loss function L user and the loss function L sys of the dialog system user constitute the overall loss function L sys .

2. The end-to-end task-oriented dialog method of claim 1, wherein, In step S3, the task-oriented dialogue system framework is targeted at the user simulator, the user simulator first understands, then generates corresponding actions and natural language through strategy, the system tracks the belief state through the user language of the user and generates corresponding actions and system responses, both of which interact through natural language, the system side has an external database, and the corresponding target is completed through interaction.

3. An end-to-end task-oriented dialog system, characterized by, The method is constructed as claimed in any one of claims 1-2.

Images