Intention recognition method and electronic device

Abstract

This application discloses an intent recognition method and electronic device. In response to a natural language query input by a user, the method identifies the target intent and target object contained in the natural language query. Based on prior knowledge of the attributes of the business domain category to which the target object belongs, the method extracts candidate attributes of the target object from the natural language query. Based on the natural language query and the attribute prior knowledge, the method performs semantic reasoning on the candidate attributes to generate corresponding attribute constraints. The attribute constraints include the target attributes corresponding to the candidate attributes under the business domain category, as well as the constraint logic operators on the target attributes. Based on the target intent, target object, and attribute constraints, the method generates a structured intent representation corresponding to the natural language query.

Claims

1. An intent recognition method, the method comprising: In response to a natural language query input by a user, identify the target intent and target object contained in the natural language query; Based on prior knowledge of the attributes of the business domain category to which the target object belongs, candidate attributes of the target object are extracted from the natural language query; Based on the natural language query and the prior knowledge of the attributes, semantic reasoning is performed on the candidate attributes to generate corresponding attribute constraints; the attribute constraints include the target attribute corresponding to the candidate attribute under the business domain category, and the constraint logic operator for the target attribute; Based on the target intent, the target object, and the attribute constraints, a structured intent representation corresponding to the natural language query is generated.

2. The method according to claim 1, wherein the attribute constraint further includes a constraint specification corresponding to the constraint logic operator; in, The constraint specifications are determined based on the natural language query or generated based on prior knowledge of the attributes.

3. The method according to claim 2, wherein the prior knowledge of the attributes includes: The predefined set of standard attributes under the business domain category, the set of colloquial descriptions corresponding to the standard attributes, and the specification mapping rules between the colloquial descriptions and standard specifications; The constraint specifications are generated based on the specification mapping rules.

4. The method according to claim 3, wherein the prior knowledge of the attributes further includes: The set of logical operators supported by the standard attributes, and / or the judgment rules used to identify logical operators under the business domain category.

5. The method according to claim 3 or 4, wherein, The method for generating the target attribute corresponding to the candidate attribute under the business domain category includes: Extract attribute constraint fragments from the natural language query for the candidate attributes; the attribute constraint fragments are positive or negative constraint representations of executing the target intent on the target object having the candidate attributes. Based on the set of colloquial descriptions in the prior knowledge of the attributes, semantic matching is performed on the attribute constraint fragments to determine the target attribute corresponding to the candidate attribute under the business domain category.

6. The method according to claim 5, wherein, Generating the constraint specification includes: Based on the specification mapping rules in the prior knowledge of the attributes, the specification descriptions in the attribute constraint fragments are converted into standardized constraint specifications.

7. The method according to claim 5, wherein, Generating the constraint logic operator includes: Based on the prior knowledge of the attributes, the determination rules for logical operators under the business domain category are used to perform semantic analysis on the attribute constraint fragments to generate constraint logical operators corresponding to the target attribute.

8. The method according to claim 5, wherein, The constraint logic operator generated includes any of the following: In response to the fact that the attribute constraint fragment contains a negative word describing the candidate attribute, the attribute constraint fragment is determined to be a negative constraint representation, and a negative constraint logical operator that matches the corresponding target attribute is selected from the set of logical operators of the attribute prior knowledge. In response to the fact that the attribute constraint fragment contains a negative word describing the candidate attribute, the attribute constraint definition of the candidate attribute in the attribute constraint fragment is negated, and based on the negated attribute constraint definition, a positive logical operator that matches the corresponding target attribute is selected from the set of logical operators of the attribute prior knowledge. The constraint specification that matches the positive logic operator is generated based on the inverted attribute constraint definition.

9. The method according to claim 1, wherein: The steps of obtaining candidate attributes of the target object from the natural language query and / or generating the attribute constraints are implemented by at least one first model; The step of generating the structured intent representation is implemented through a second model; The second model has more model parameters than the first model; The attribute prior knowledge can be dynamically updated based on changes in business information under the business domain category.

10. An electronic device, comprising: Input component, used to obtain natural language queries from user input; At least one memory for storing computer programs; At least one processor is configured to execute the computer program to implement the steps of the intent recognition method as described in any one of claims 1-9.

Images