Hierarchical questions improve interpretability of text embeddings
Original title: Interpretable Text Embeddings Based on Hierarchical Questions
Hierarchical questions reduce the number humans must examine while maintaining task performance in QA-based embeddings.
QA-based text embeddings represent texts as vectors of 1/0 answers to yes/no questions. These provide detailed information and high task performance but require humans to examine a large number of questions. The relationship between question number and granularity level has not been studied sufficiently. We propose a method for constructing hierarchical questions where deeper levels become more specific and their number increases. During interpretation deeper questions are replaced with aligned shallower ones. The hierarchy is constructed through recursive clustering, contrastive prompting and question alignment. Experimental results on general-purpose tasks show improved interpretability while maintaining task performance. When limited questions are available shallower levels achieve higher task performance than deeper levels.
- Questions become more specific with deeper hierarchy levels
- Deeper questions replaced by aligned shallower ones during interpretation
- Hierarchy constructed via recursive clustering, contrastive prompting and alignment
- Shallower levels perform better with limited question counts
From the original article
1 min read at arxiv.orgQA-based text embeddings represent texts as vectors consisting of 1/0 answers to yes/no questions. Although these embeddings provide detailed information for interpretation and achieve high task performance, they often require humans to examine a large number of questions. Furthermore, the relationship between the number of questions and the appropriate level of question granularity has not been sufficiently studied.
By Kohei Oda · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.