- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 3
Read review
Paper summary
This paper investigates the hypothesis that in embedding models, the structural consistency of the embeddings of the corresponding pairs predicts the final benchmark performance. It provides a big set of evaluations on the MTEB benchmark and proposes two different measures of embedding similarity. It demonstrates that on some tasks they correlate well with the model's performance, although the strength of these correlations varies considerably across tasks.
Critical strengths
Understanding what a good embedding is characterized by geometrically is important.
I also like how this paper presents a very exhaustive list of experimental evaluations of different models and tasks from the point of view of representations.
Critical weaknesses
The generality of empirical findings is limited. They differ by the dataset and utilized method.
The scientific insight beyond the observed correlations is limited. That is, the authors show that quite often some geometric properties are correlated with the performance. They however do not explain why is that the case and when can we expect that to happen.
Questions for authors
- How were the evaluated models trained?
- How does Figure 2 look on log-scale?
- How does varying the prompts used change the experimental results?
- If you'd look at averaged MTEB performance (per category) instead of per-task performance, would some clearer patterns emerge?
- Can the authors demonstrate some way in which the insights from this paper can be used to improve text-embeddings algorithms?
- Are the proposed methods of investigating embeddings better than cosine similarity?
- Why for some datasets gini-coefficients shows stronger correlation than for others?
3 comments
Reviewer KvH1: are you convinced by the rebuttal, particularly the generality of the empirical findings?
Thank you for your responses. While they make things clearer for me, my concern about the generality of empirical findings has not been addressed, I will therefore maintain my score.
Response to review
Thank you for your review and your patience in waiting for our response while we ran new experiments. We address the weaknesses and questions below:
W1. (Generalizability) The manuscript demonstrates quite a lot of generalizability despite the assumption that tasks are encoded differently: 3 tasks show high correlation, and additionally, the summarisation task also shows medium correlation; the same applies to ICA with some caveats. We also consider finding dataset- and task-level differences in the level of correlation as one of the most interesting results in our paper. We further explain the outliers we see(like WebFAQ:eng, and the RTE-datasets), and the differences between neighborhood retention and ICA are analyzed in its own section (Section 5 “Shuffling ICA”).
W2. (Causality) We agree that our work is purely correlational and we indeed carefully avoided claiming causality until it can be proven in subsequent work; our goal was to analyze two ways to study structures in embedding spaces, and our results are very strong in this regard. Additionally, at the request of reviewer 77Uu, we ran a case study at the query level, looking at how retention relates to retrieval. If you are interested, please have a look at the message directed to all reviewers.
Q1. (Model training) We add the full hyperparameter list of trained ICA models in the appendix. The 25 embedding models we use in the study are all “off-the-shelf” transformer-based encoder models, for which we do no fine-tuning.
Q2. (New Figure) We are happy to include this figure: With a logarithmic x-axis, the models lie more or less on the line. We thank you for the suggestion, this figure will be available in the updated pdf.
Q3. (Varying prompts) We agree this is a reasonable concern; however, comparison across all models, datasets, and multiple prompt options would result in huge increase of the required compute, time, and effort, and we cannot run this analysis on subset of data, as we are looking for a signal across models and datasets. We investigate two options in the current manuscript and the scope of the current study is already resource-heavy and we are unable to expand the experiments in this direction at this moment.
Q4.(Task average scores) One of the bottlenecks in our study is the availability of MTEB scores. Running the full MTEB evaluation takes up to 2-3 GPU days per model, and doing this for the full 25 models is beyond our scope. While we think that studying more datasets per task would be valuable, we unfortunately cannot deliver this request within the time span of the discussion period.
Q5. (Improving embeddings) We think our work gives the community knowledge of properties that are present in models with high performance; this helps us understand the embedding spaces, and most importantly, which characteristics to look for when running causal studies. In this exploratory work, the strong correlations highlight candidate properties for further intervention research and this is genuine future work for us.
Q6. (Comparison to Cosine similarity) Yes, we already compare to cosine similarity. We analyse this in Section 5 shortly and in length in Appendix B.6. We see that our Ret_k method reaches higher correlation than any of the cosine similarity types we tested, from raw cosine similarity to z-standardized cosine similarity. We additionally do partial correlation and demonstrate that retention does not reduce to cosine similarity.
Q7. (Gini-variability) We believe this is due to multiple reasons: first, ICA is unsupervised, meaning it can latch onto any signal found in the data. We find that the same signal is found consistently, and see that in most cases, vocabulary explanations verify the signal is what we expect; despite this, ICA can in some cases latch onto something else, and this is simply a feature of ICA and will be visible in the Gini values. Second, in certain tasks, despite ICA being able to latch onto a correct signal, this might not be in line with what is needed to perform well on the task. This is the reason why we investigate the structure using two methods instead of one.
Thank you again for your review. We believe we have answered your questions; please let us know if we left anything unclear. If not, we hope that you will consider updating your evaluation.