Voice assistant -- Architecture and design of Instruction Assistant

        In this chapter, let's take a look at voice assistant , Command type （ Task oriented ） The overall architecture and design of the assistant .

One 、 Application scenarios of instruction assistant

        ad locum , If the assistant's role is to help the user perform certain operations , such as ： Voice alarm clock , Listen to music by voice , Voice navigation, etc , The assistant that implements this kind of task is called instruction type voice assistant , That is, the user sends some instructions through voice , The assistant operates according to the instructions , So as to realize the liberation of both hands , The purpose of the shortcut , For example, the following examples ：

           

Two 、 Instruction Assistant architecture and design

        We are in the previous chapter 《 The overall architecture and design of voice assistant 》 Mentioned in , Each vertical class in the instruction assistant acts as BOT There is a form of ,BOT It's a small Sunday , Each module has , The following figure shows a BOT The overall architecture and design of ：

          Pictured above , stay BOT There are various worker, Be responsible for handling different tasks ,worker There is no direct communication between , But through DataBus To deliver data , therefore ,worker And worker They are completely independent of each other , Just need to satisfy DataBus Interface definition in , Data transmission can be realized , This design , You can make worker You can iterate independently , Sub module development .

        Let's look at each one separately worker The role of ：

1、 Basic semantic understanding

        The basic semantic understanding here , It refers to the input query Some basic processing of , such as NER、 Syntactic parsing 、 participle 、 Scene classification , Error correction, etc , This step is usually performed by external DM Common module processing in , The incoming to BOT in .

2、history_session_worker

        this worker It is mainly used to process historical information , Including getting global dialog history information , The BOT Relevant historical information, etc , Generally, if it is an open multi wheel , Global dialog history information is required , In case of closed multi wheel , This is generally required BOT Relevant historical information .

3、qu_worker（BOT Inside ）

        this worker And the outside DM Medium qu Modules are different , External qu Modules are used to handle basic semantic understanding , The identified content will be shared to each active BOT, and BOT Inside qu_worker It is specially used to deal with the worker Peculiar qu Information （ If necessary ）, for instance , If there is an alarm clock scene query rewrite , It takes effect here .

4、scene_identify_worker

        this worker Generally, it is multi classification or two classification , We talked about that before , Every BOT Need to have “ Identify incoming query Whether it belongs to BOT” The ability of , So it's time to worker The main function is to exclude those that do not belong to the scenario query,DM Recall the scene classification model in , So for some BOT Vague words will be recalled by mistake , At this point the worker The classification in the needs to be accurate , It can improve the accuracy . Of course , Besides this function , It can also be implemented as multi classification , For... Below worker Provide information .

5、intent_entity_worker

        this worker It is mainly used to identify the intention , Extraction slot , Its internal implementation can be based on BOT Type to implement , Generally speaking, there are several ： Templates 、 Model （ Intention slot model ）、 Templates + Model 、 Semantic role annotation, etc .

6、context_association_worker

        this worker The ability to implement multiple sessions , Generally including ： Closed multi wheel, i.e. slot position inquiry , Open multiple rounds means semantic inheritance 、 Omit and complete 、 Anaphora digestion , Scripted multi wheel .

7、NLU_postprocess_worker

        this worker Used for post-processing semantic content . this worker Before （ Include this worker） The content of , We are all divided into semantic layers , That is, the content identified here is the result of pure semantics , namely query Its own semantics . Distinguish it from the skill level below , Sometimes according to the product demand , Some semantics need to be executed with special skills , such as ：query：“ Search for facial cleanser ”, The query The meaning of is “ Search for items ”, If the item is a commodity , You need to perform shopping skills . If the query Semantics divided into shopping , and “ Buy facial cleanser on Taobao ” In the same way , May cause the BOT False recall of .

        therefore , take query Divide according to semantics , Each different type of semantics can be iterated separately , And some processing according to different semantics , Take the example above , If query by “ Search for xxx”, It needs to be judged in combination with the knowledge map “xxx” Whether it is a commodity , If it is a commodity, go shopping , Otherwise, go to encyclopedia ; If query by “ Buy on Taobao xxx”, It doesn't matter “xxx” What is it? , All need to go shopping .

        The worker in , Generally, entity disambiguation is implemented 、 Post processing at semantic level such as entity verification .

8、skill_worker

        This is the first step in the skill level worker, Mapping semantic results to skills , A mapping table is usually maintained here , Mapping semantics - Skill .

9、skill_rank_worker

        Realize here BOT Inside is different skill Sorting between , for instance , Map BOT Inside ,query：“xxx Where is the ”, This meaning will be mapped to two skills ：1、“ Search the map xxx”.2、“ The encyclopedia card shows xxx Introduction to ”. These two skills will be based on "xxx" Sort by type , If “xxx” For scenic spots , Then the skills 2, If “xxx” It is a common place , Then the skills 1. 

10、skill_postprocess_worker

        The implementation here is for skill Post processing of , such as query：“ What's the weather like today? ”, After identifying to semantics and mapping to “ Check the weather ” After skill , Can be here worker To access the weather server , Get weather information , In the form of a card , Distribute to client .

11、nlg_worker

        This step is generally used to generate nlg Talking skill , In the skill type BOT in , commonly nlg They are all configured based on templates , according to slot、intent as well as context Information , Different combinations generate different reply scripts . Chatting BOT Or ask and answer BOT in , this worker Instead, you can generate a model or retrieve a model to generate a reply .

        General recommendations are also available here , such as ：query：“ What's the weather like today? ”, Get recommendations ：“ What about tomorrow ”、” What to wear today “ etc.

12、execution_worker

        This is mainly used for final result encapsulation , Encapsulated as structured data , Return to external DM.

3、 ... and 、 Written in the back

        The above architectural form , It can be realized that BOT The internal links are detailed 、 decompose , each worker Independent of each other , It can be developed independently by different principals 、 upgrade . meanwhile BOT And BOT They are also independent of each other , It can be upgraded independently , individual worker perhaps BOT error , It does not necessarily affect the overall results （ Notice that this is ” not always “, That is, the robustness of the system is improved to a certain extent , But the final result may not be optimal ）.

        Because of the various worker Independent of each other , Only pass data_bus communicate , Therefore, theoretically, it can be designed to pass RPC perhaps http In the form of communication , In this way, we can really realize worker Independent upgrade of . But this will increase the network overhead , May cause system performance degradation .