Can a Self-Organizing Map Help Improve Memory-Based Learning in Natural Language Processing?
Title: Can a Self-Organizing Map Help Improve Memory-Based Learning in Natural Language Processing?
Research Question: Can a Self-Organizing Map (SOM) be used to select a subset of training items for comparison in memory-based learning (MBL), thereby reducing the number of comparisons and potentially improving the performance of MBL in natural language processing tasks?
Methodology: The researchers used a hybrid system called Labelled SOM and MBL (LSOMMBL). The SOM was used to create a map of the input space, with each unit having a weight vector. When a novel item was presented, the unit with the closest weight vector was selected as the winner. The weight vectors of the winner and a neighborhood of surrounding units were then nudged towards the novel item. This process was repeated multiple times during training. The size of the neighborhood was gradually decreased, and the learning rate was also reduced. At the end of training, the units formed a map of the input space, with areas of high density representing areas where there were many inputs.
Results: The researchers tested their system on the task of base noun-phrase (NP) chunking using the Wall Street Journal corpus. They found that the LSOMMBL system was able to identify base NPs with high accuracy. The system's performance was comparable to that of other MBL systems, but it required fewer comparisons, as the SOM was able to select a subset of training items for comparison.
Implications: The results suggest that a SOM can be used to improve memory-based learning in natural language processing. By selecting a subset of training items for comparison, the system can reduce the number of comparisons needed, which can lead to faster processing and potentially better performance. However, further research is needed to fully understand the potential benefits and limitations of this approach.
Link to Article: https://arxiv.org/abs/0107018v1 Authors: arXiv ID: 0107018v1