Word Vector Retrofitting via Reinforcement Learning for Incremental Domain Adaptation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing word vector training models and retrofitting methods cannot support incremental training, are inefficient in resource utilization, and fail to improve word vector quality using external information or adapt to new domains without labeled data, leading to computational burdens and reduced performance in cross-domain applications.
Innovation Solution
A word vector retrofitting method using reinforcement learning, where two models with opposite training directions are interacted to continuously retrofit their parameters based on losses generated during training with a target corpus, allowing for incremental training and adaptation to new domains without requiring large-scale data reconstruction.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If word vectors are trained using traditional models (CB or SG), then word vectors can be obtained for machine processing, but the models cannot support incremental training and require reconstruction of large-scale corpus from scratch when adjustment is needed
Solution Approach 1:
The patent segments the training process into two independent components: (1) pre-training phase that generates initial word vectors using large-scale corpus, and (2) retrofitting phase that incrementally adjusts vectors using target corpus. This segmentation allows the retrofitting model to operate independently without requiring full corpus reconstruction, thereby enabling incremental training and reducing computational burden.
Solution Approach 2:
The patent performs preliminary action by pre-training word vectors using large-scale corpus before the actual application phase. The pre-trained vectors serve as initialization for the retrofitting process, allowing the system to start with quality vectors and make incremental adjustments rather than training from scratch, thus improving training efficiency and reducing computational burden.
2Manufacturing precision
If traditional retrofitting methods are used with manually labeled or automatically extracted word-to-word relationships, then original low-quality word vectors can be upgraded to high-quality vectors, but the method is only suitable for a small quantity of words and is useless for words without labeled data
Solution Approach 1:
The patent implements self-service by designing a retrofitting model that automatically learns word-to-word relationships directly from the target corpus without requiring manual labeling or external labeled data. The model uses the corpus itself as the source of learning signals, enabling it to adapt to any domain and handle any vocabulary including words without pre-existing labeled relationships, thus improving both versatility and quality.
3Reliability
If word vectors are trained for a specific domain, then they perform well in that domain, but they cannot be effectively adapted to new domains without labeled data
Solution Approach 1:
The patent applies dynamics by creating an adaptable retrofitting system that can dynamically adjust word vectors based on the target corpus from any domain. The model structure allows it to transition between domains by retraining on new corpora, maintaining reliability in the original domain while gaining adaptability to new domains through incremental learning without requiring labeled data.
4Manufacturing precision
If large-scale corpus reconstruction is performed to adjust word vectors, then vector quality can be improved, but additional computational burden is incurred
Solution Approach 1:
The patent extracts only the essential information needed for vector adjustment from the target corpus, rather than processing the entire large-scale corpus. The retrofitting model focuses specifically on learning word-to-word relationships relevant to the target domain, extracting and utilizing only the necessary signals for quality improvement while avoiding the computational overhead of full corpus reconstruction.
Data Source
AI summary
The present disclosure discloses a word vector retrofitting method. The method includes obtaining, by a computing device, a first model and a second model that are generated when original word vectors are trained, the first model being configured to predict a context according to an inputted word, and the second model being configured to predict a target word according to a context; inputting a corpus unit from a target corpus into the first model, inputting an output of the first model into the second model, and determining losses generated by the first model and the second model when the second model outputs the corpus unit; and retrofitting the first model and the second model according to the losses.