Voice assistant -- Qu -- semantic role annotation and its application

In this chapter, we will talk about semantic role tagging （Semantic Role Labeling (SRL)） And its application in voice assistant , It is mainly divided into 4 part ： What is semantic role annotation 、 Why do we need semantic role annotation 、 How to achieve 、 Voice assistant applications

1、 What is semantic role annotation

         Semantic Role Labeling （Semantic Role Labeling (SRL)） It is also called chunk analysis （query chunking）, It is a shallow semantic analysis technology . Give a sentence , SRL The task of is to find out the corresponding semantic role components of predicates in sentences , Including core semantic roles （ As an agent 、 Patient, etc ） And subordinate semantic roles （ Such as location 、 Time 、 The way 、 Reasons, etc ）.  for instance ：      The yellow part is the result of semantic role annotation ,A0 For the practical ,A1 For the patient ,ARGM-ADV Is an adverbial .

2、 Why do we need semantic role annotation

         You can see from above , Through semantic role annotation , We can get the core part of a sentence , These core parts can represent the semantics of the sentence , Using this feature, we can realize semantic understanding . In previous chapters , We talk about the method of using intention slot model to identify semantics , But some types of scripts are not suitable for this method , such as ：

1. Time + Content ： This kind of script needs to judge whether it needs to be called back to the schedule according to the type of content , such as ： Tomorrow, 8 Hold a meeting （“ The meeting ” It's a verb , You can call back to the schedule reminder ）, Tomorrow, 8 Point Weather （“ The weather ” Non verb , This should not be part of the schedule ）. If this kind of script uses the intention slot model, it is easy to recall by mistake .
2. action + Content ： This kind of script needs to judge the intention according to the category of content , such as ： Open blue and white porcelain （ Should play music ）, Open the WeChat （ Should belong to open app）, If the intention slot model is used in this kind of scripts, it is difficult to accurately identify the intention , because “ Content ” Knowledge is strongly related , At this point, we can use semantic role annotation to identify this kind of script , Then the content type is determined by combining the knowledge map to further determine the intention .

         The above two types of scripts , All of them are strongly related to knowledge , It needs to be combined with the knowledge map to further determine the semantics .

        meanwhile , Semantic role tagging is much more efficient than the intention slot model for some scenes , such as “ Set up ” This scene , We know that in mobile phones “ Set up ” This app Contains many operations , Want to control the operation , such as ： Turn on the flashlight , Turn on do not disturb mode , Turn the volume to 25. There are hundreds of these operations , Imagine if you want to identify hundreds of intentions through an intention slot model or classification model , The difficulty is very high . In this case, semantic role annotation can be used , Identify “ action ”、“ Entity ”, combining “ Mobile phone knowledge map ” According to different combinations of actions and entities , Identify different semantics . Next, we will talk about the specific implementation here .

3、 How to implement semantic role annotation

        So how to get the results of semantic role annotation ？ Open source tools can be used “LTP”, Use it “ Semantic Role Labeling ” Result , Or use it “ dependency parsing ” The result of secondary processing , To fit your scene .

        Of course, we can also train the model ourselves , This model is similar to “NER” The model is very similar , We can use BILSTM-CRF, perhaps Transformer, perhaps Bert, We don't need to expand here , The key is to label the data .

4、 Application examples in voice assistant

        Here we take “ Set up ” Take the scenario as an example , The scene features are also mentioned above , Mainly for ： The sentence is short 、 Many intentions 、 There are many ways to express . such as ：” Turn on the flashlight “,” The room is too dark , Turn on the flashlight for me “,” Turn on the flash “, They all have the same intention .

         Pictured above , For user input query, Using semantic role tagging to obtain sentence components , such as ” Turn on the flashlight “, Split into ” open （ action ）“ and ” The flash （ Entity ）“, And then to the action 、 Entities are disambiguated and normalized , Finally, combined with the functional map , Get the intention of this sentence .