Topic Model for Named Entity Recognition in Search Queries
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing query processing methods for search applications, such as query segmentation, syntactic parsing, and query classification, are limited in handling short and poorly formed search queries, and traditional Named Entity Recognition (NER) techniques are not effective due to the lack of identifiable features in search queries.
Innovation Solution
A method for recognizing named entities in search queries using a topic model and predefined taxonomy, with an offline training component that extracts seed named entities and their contexts from search queries, and an online prediction component that predicts classifications based on Weakly Supervised Latent Dirichlet Allocation (WS-LDA) learning.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional Named Entity Recognition techniques are applied to search queries, then identifiable features can be used for classification, but search queries are too short and poorly formed to contain these features
Solution Approach 1:
The patent applies preliminary action by pre-training a topic model offline using a large corpus of search queries and their associated named entities. This training phase prepares the model in advance so that during online query processing, the model can immediately classify named entities without requiring the queries to contain traditional identifiable features. The offline training creates a knowledge base that compensates for the lack of features in short, poorly formed queries.
Solution Approach 2:
The patent introduces a topic model as an intermediary between the search query and the named entity classification. Instead of directly applying traditional NER rules to queries, the topic model serves as a mediator that has been trained to recognize patterns and relationships between queries and named entities. This intermediary layer enables classification even when traditional identifiable features are absent.
2Device complexity
If query classification is performed on the whole query, then classification can be done simply, but no further analysis on the internal structure of the query is performed
Solution Approach 1:
The patent applies segmentation by dividing the query classification process into two distinct stages: whole-query classification and named entity-level classification. The first stage performs simple classification on the entire query, while the second stage segments the query to identify and classify individual named entities using the pre-trained topic model. This segmentation allows both simple overall classification and detailed internal structure analysis to coexist.
Solution Approach 2:
The patent adds another dimension to query classification by moving from single-level whole-query classification to a multi-level approach that includes both query-level and entity-level classification. This dimensional expansion allows the system to maintain simple overall classification while simultaneously performing detailed analysis of internal query structure through named entity recognition.
3Reliability
If deterministic template-based query log mining is used, then named entities in specific classes can be acquired, but it only works when a named entity belongs to only a single class
Solution Approach 1:
The patent applies dynamics by transitioning from static deterministic template matching to a dynamic probabilistic topic model. The topic model can adaptively handle named entities that belong to multiple classes by calculating probability distributions across different classes. During online processing, the model dynamically determines the most appropriate class based on the specific query context, allowing reliable classification even when named entities could belong to multiple classes.
Solution Approach 2:
The patent changes the parameter of classification from deterministic (single correct answer) to probabilistic (multiple possible classes with associated probabilities). The topic model outputs probability distributions for different named entity classes, allowing the system to handle multi-class entities by selecting the class with the highest probability or by considering multiple high-probability classes, thereby improving both reliability and versatility.
Data Source
AI summary
Named Entity Recognition in Query (NERQ) involves detection of a named entity in a given query and classification of the named entity into one or more predefined classes. The predefined classes may be based on a predefined taxonomy. A probabilistic approach may be taken to detecting and classifying named entities in queries, the approach using either query log data or click through data and Weakly Supervised Latent Dirichlet Allocation (WS-LDA) to construct and train a topic model.