Autoencoder Domain Determination for Sentence Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for determining the domain of a sentence in language-based recognition systems are inefficient, requiring frequent recalibration of threshold values and leading to variations in performance depending on data, which affects training efficiency and accuracy.
Innovation Solution
The use of an autoencoder to generate embedded features from input sentences and determine their domain based on their location in an embedding space, where embedded features closer to a specified location are classified as in-domain, and those farther away are classified as out-of-domain, utilizing distance and reconstruction error thresholds.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a threshold value is used to distinguish between in-domain and out-of-domain sentences, then domain determination can be performed, but a new threshold value needs to be obtained at each training time, causing inefficiency
Solution Approach 1:
The patent applies preliminary action by pre-training the autoencoder model on domain-specific data before actual domain determination tasks. The autoencoder learns the characteristic features of in-domain sentences during this preliminary training phase, enabling it to automatically adapt to different domains without requiring threshold recalibration for each new domain. This preliminary feature extraction and model adaptation eliminates the need for repeated threshold optimization during subsequent training tasks.
2Ease of operation
If a threshold-based method is used for domain determination, then classification can be performed, but variation in domain determining performance increases depending on data
Solution Approach 1:
The patent applies parameter changes by transforming the domain determination problem from threshold-based classification to distance-based classification in an embedded feature space. Instead of using a fixed threshold value that varies in effectiveness across different datasets, the method changes the parameter from a scalar threshold to a geometric distance metric in a high-dimensional space. This allows the system to dynamically adapt to different data distributions while maintaining consistent performance, as the distance metric naturally adjusts to the characteristics of each domain.
Data Source
AI summary
Methods and apparatuses for determining a domain of a sentence are disclosed. The apparatus may generate, using an autoencoder, an embedded feature from an input feature indicating an input sentence, and determine a domain of the input sentence based on a location of the embedded feature in an embedding space where embedded features are distributed.