Subword Vector Query Matching for Ad Revenue
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing query-to-advertisement matching technologies face challenges in maximizing revenue due to their reliance on whole query vectors, which result in large vocabularies and inability to effectively match queries not frequently encountered in training data, leading to suboptimal ad selection.
Innovation Solution
The approach involves identifying and using subword vectors for queries, combining them to generate query vectors, and optimizing a query/ads model using a subword vector generator, combiner, and optimization engine to enhance matching relevance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If whole query vectors are used for matching, then the model can capture complete query semantics, but the vocabulary size becomes extremely large and the system cannot effectively match queries not frequently encountered in training data
Solution Approach 1:
The patent segments queries into subwords (characters or character n-grams) and learns vectors for these subwords separately. Instead of treating each entire query as a single vocabulary unit, the system breaks down queries into smaller subword components that can be independently vectorized. This segmentation reduces the effective vocabulary size from exponential (all possible queries) to manageable (common subwords), while still capturing query semantics through composition of subword vectors.
2Stability of the object's composition
If whole query vectors are used for matching, then the system can achieve consistent vector representations, but it provides no option for obtaining vectors for queries not encountered in session training data
Solution Approach 1:
By segmenting queries into subwords, the system creates a compositional representation where unseen queries can be vectorized by combining vectors of their constituent subwords. This allows the system to handle queries not encountered during training, as long as their subwords are present in the vocabulary, thereby improving adaptability while maintaining representation consistency through the compositional approach.
Solution Approach 2:
The subword vector approach creates a universal representation system that can handle both seen and unseen queries through the same mechanism. The subword vocabulary serves multiple functions: representing common words directly and composing representations for rare or novel queries through combination of subword vectors, making the system versatile across different query types and frequencies.
3Adaptability or versatility
If subword vectors are used for queries, then the system can effectively match less frequent queries to advertisements, but additional processing steps are required to combine subword vectors into query vectors
Solution Approach 1:
The segmentation into subwords enables the system to handle less frequent queries by composing their vectors from common subword components, improving adaptability. The additional processing complexity of combining subword vectors is offset by the use of efficient composition operations (such as averaging, summing, or neural network-based composition) that operate on fixed-size subword vectors rather than requiring lookup of entire query vectors.
Data Source
AI summary
The present teaching relates to obtaining a model for identifying content matching a query. Training data are received which include queries, advertisements, and hyperlinks. A plurality of subwords are identified from each of the queries and a plurality of input vectors for the plurality of subwords of each of the queries are obtained and an input vector for each of the queries is derived based on a plurality of input vectors of a plurality of subwords of the query. A query/ads model is optimized with respect to an objective function via training an input vector for each of the plurality of subwords associated with each of the queries, an input vector for each of the advertisements and links, and a matrix.