Summary of Multicontrievers: Analysis Of Dense Retrieval Representations, by Seraphina Goldfarb-tarrant et al.
MultiContrievers: Analysis of Dense Retrieval Representations
by Seraphina Goldfarb-Tarrant, Pedro Rodriguez, Jane Dwivedi-Yu, Patrick Lewis
First submitted to arxiv on: 24 Feb 2024
Categories
- Main: Computation and Language (cs.CL)
- Secondary: Artificial Intelligence (cs.AI); Information Retrieval (cs.IR)
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary Dense retrievers compress source documents into vector representations, yet there’s limited analysis on what information is lost or preserved. We analyzed the information captured by dense retrievers compared to language models (e.g., BERT vs. Contriever). Our study used MultiBert checkpoints as initializations to train MultiContrievers, a set of contriever models. We tested whether specific pieces of information like gender and occupation could be extracted from contriever vectors in wikipedia-like documents. We measured extractability via information-theoretic probing. We also examined the relationship between extractability and performance, gender bias, and sensitivity to random initializations and data shuffles. Our findings show that (1) contriever models have increased extractability but poor correlation with benchmark performance; (2) gender bias is present but not caused by contriever representations; and (3) results are sensitive to both random initialization and data shuffle, suggesting future research should test across a wider range. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Dense retrievers help computers understand text by creating special codes. But no one has looked at what information gets lost or kept during this process. Our study compares these codes to language models like BERT. We used 25 special starts for our code-making models and tested if we could get specific details, like someone’s gender or job, from the resulting codes. We also checked how well these codes work and if they help or hurt certain groups of people. We found that (1) these new codes are better at getting information but don’t always do as well as other methods; (2) there is some bias towards certain groups, but this isn’t because of the code-making process itself; and (3) our results depend on the starting point and data used. |
Keywords
» Artificial intelligence » Bert
