The design points of voice dialogue system and the importance of multi round dialogue

Let's start with the recent chat between the aunt on the short video platform and the robot kuaibao .

In a bank , An aunt waited too long for business , Started a dialogue with kuaibao robot from the soul . Dissatisfaction with bank staff , Aunt said to kuaibao ：“ Your staff are inside to coax the children , Why don't you come out to do business ？”; Quick answer ：“ Our brothers and sisters in the counter are also working hard to do business .” Hear the answer , Aunt tried to turn kuaibao's body towards the counter , say ：“ You look inside , Are you kidding the children ？” Kuaibao's good grievance ：“ Don't touch me , Just talk to me ”.

“ Quick treasure ” The speed and voice of speech are very clear , Logical coherence , The way you talk to ordinary people is exactly the same , It's better than Apple's SIRI Too much , Some netizens even suspected kuaibao “ behind ” It's a special person talking to people through the camera .

With the update and iteration of artificial intelligence related technologies , Now ,ASR And TTS Technology will be relatively mature , Great progress has been made in the representation and understanding of natural language , It is also gradually weakened in the competitive barriers of the industry , The core competitiveness of intelligent dialogue robots in the future lies in the differentiated services provided after understanding the user's intention . Let me talk about some knowledge about voice dialogue robots from the perspective of products , I hope to bring you some thoughts .

1、 Composition of intelligent voice dialogue system

The intelligent voice dialogue system can be roughly divided into five basic modules ： speech recognition (ASR)、 Natural speech understanding (NLU)、 Dialogue management (DM)、 Natural language generation (NLG)、 speech synthesis (TTS).

Speech recognition converts speech into words , Let the machine read what the user said , Natural language understanding is to understand what the user is saying , Analyze the user's intention , And the analysis of the core word slot in the user language . And dialog management (Dialog Management,DM) It is in man-machine conversation CPU, It controls the whole process of man-machine dialogue . The main tasks of dialogue management are as follows , Conversation state maintenance (dialog state tracing,DST)、 Generate system decisions (dialog policy)、 As an interface with the back end / Interact with the task model 、 Provide semantically expressed expectations (expections for interpretation). After analyzing the user's intention, the dialog management makes relevant actions , Natural language generation generates the processing results of user tasks in the form of text , Then speech synthesis synthesizes the result into speech . The whole process of man-machine dialogue is formed .

2、 Classification of man-machine conversation

Individuals divide the common man-machine conversation into daily provocation type and task driven type .

The most common is task - driven multi - wheel dialogue , The user has a clear purpose, such as ordering meals 、 booking 、 Call a car and other complicated requirements , And there are many restrictions , The user cannot finish the key information required by the task at one time 、 Speak clearly , Therefore, it should be carried out in multiple rounds QA Question and answer . One side , The user is in the conversation , Can constantly revise and improve their own needs ; On the other hand , When the user's statement of requirements is not specific and clear enough , Robots can ask 、 Clarification and confirmation to help users find satisfactory results , And complete daily interaction with users driven by tasks , So as to continuously improve the satisfaction of user needs .

The key to daily provocative conversations , It is to judge the current mood of the user according to the date, time and tone of the first sentence when the user wakes up the robot and after waking up , such as ： Friday night 9 Go home from work at , And when I wake up, I feel a little tired and unhappy , At this point, we need the comfort and encouragement of the robot , So as to meet the emotional needs of users . When users are happy , You can also occasionally “ A skin ”, There must be amazing sentences and popular words in the dialogue , Fun and playfulness are just needed in daily flirtation dialogue , This requires trainers to constantly update the corpus , So as to maintain the attachment to intelligent voice devices .

3、 The importance of dialogue management

Dialogue management is very important for multiple rounds of dialogue , Because each chat of word dialogue requires the user to wake up the voice dialogue robot , The user must say what needs to be done each time , Otherwise, after a few conversations, the user will feel irritable , Voice dialogue robots will become chicken ribs . Let's break down the general tasks of dialogue management ：

1、 Conversation state maintenance (DST)

2、 Generate system decisions (dialog policy)

according to DST Dialogue status in , Generate system behavior , Determining what to do next can monitor user input , Namely NLU The process of , And the system for NLU Your feedback behavior , Namely NLG.

3、 As an interface with the back end / Interact with the task model .

Act as an application program interface to interact with the server or task model , Get feedback results , Generate text results .

4、 Provide semantically expressed expectations

According to the expression entered by the user , Including language expression and semantic analysis , Make semantic expressions that meet user expectations , Meet the needs of users .

In order to clearly understand the user's intentions and needs in multiple rounds of dialogue , Form the missing information in the dialogue modeling process into a slot filling process , The slot is the information that needs to be completed to convert the preliminary user intention into clear user instructions in the multi round dialogue . A slot corresponds to a kind of information that needs to be obtained in task processing . Slots have no order , Ask the user what information you lack .

Framework based dialog management (Frame-based DM) The following points are required ：

1、 frame ： Set of slots , Defines what information needs to be provided by the user ;

2、 Conversation status ： It records which slots have been filled , Those slots are to be filled ;

3、 Behavior choice ： The next step is to ask the user what information , Which slots to fill , What to do , Which slots are weighted .

A framework based system is essentially a generative system , Different types of input trigger different generation rules , Each generation is flexible to fill in the corresponding template , These models and frameworks are designed only to meet the needs of users , Complete the acquisition of necessary information as soon as possible .

4、 The key points of voice dialogue system design

The design of voice dialogue system needs attention 5 A point ：

Design of behavior patterns 、 Design of interaction process 、 Design of knowledge structure 、 Design of personality and emotion 、 Familiar with process design , We can put this 5 The main points of design are divided into scenarios ：

In the overall architecture design , Join the collection of these sub scenarios , Through the behavior segmentation mode of user and machine dialogue , Including the collection of knowledge structure and personality emotion , Come up with a virtual personality . This mode is equivalent to a treasure or an East Mall according to the user's click 、 Thousands of people and thousands of faces generated by the combination of search and browsing behavior and big data , Through the use of voice interaction , Understand user habits to achieve a better experience .

Voice conversation experience can be divided into three directions ： Voice image 、 Dialogue interaction mode and dialogue content , They correspond to each other GUI Brand design of the times 、 Interaction design 、 Service design , Product managers need to grasp the balance between robots and people , Don't be too human , So as not to meet the transition expectations of users at some points , And the resulting disappointment .

The human brain depends on what it has learned to think 、 Logical reasoning and language understanding . Robots, on the other hand, rely on data for training , A large amount of data accumulated in the Internet era can provide a strong guarantee for training robots , Dialogue robots are based on data , Using deep learning models and algorithms , To perceive the human world 、 Identify and judge , And sort out the human knowledge through the knowledge map 、 Integrate 、 Reasoning , Become wise AI.

5、 The problems that dialogue robots face

Human complexity (complex)、 Randomness (random)、 And irrationalization (illogica) The characteristics of human-computer dialogue lead to various problems under the application scenarios , Including but not limited to the following issues ：

1、 The model describes the trade-off between capability and business complexity ;

2、 The user dialogue deviates from the path and boundary involved in the business ;

( Such as ： When the system asks the user the destination of navigation , The user asked a rhetorical question about the weather conditions in a certain place )

3、 Fault tolerance of multi round dialogue ;

( Such as ：3 The scene of a round of dialogue , The user has completed 2 round , The first 3 Wheel cause ASR or NLU error , Lead to the abandonment of previous achievements , So the user experience is very poor .)

4、 Switching and replying to multiple scenarios ;

5、 Reduce the difficulty of interactive change , Adapt to the rapid change of business ;

6、 Information inheritance across scenarios .

6、 Conclusion

In addition to passively responding to users' requests, future dialogue robots , It has become an inevitable development direction to actively predict user needs and provide immediate solutions , When the user does not give a clear demand , Remind of upcoming events or recommend useful services , People will gradually rely on them to manage their work life , Improve life efficiency and happiness .

The goal of a dialogue robot is not necessarily to solve all the problems that users face , Instead, it becomes the virtual assistant of users . By establishing emotional links with users , Understand the user , Help them in the long term , Establish various forms of communication with users , Including text 、 Voice and image and video functions .