Uncovering Meanings of Embeddings via Partial Orthogonality

Machine learning tools often rely on embedding text as vectors of real numbers. In this paper, we study how the semantic structure of language is encoded in the algebraic structure of such embeddings. Specifically, we look at a notion of ``semantic independence'' capturing the idea that, e.g., ``eggplant'' and ``tomato'' are independent given ``vegetable''. Although such examples are intuitive, it is difficult to formalize such a notion of semantic independence. The key observation here is that any sensible formalization should obey a set of so-called independence axioms, and thus any algebraic encoding of this structure should also obey these axioms. This leads us naturally to use partial orthogonality as the relevant algebraic structure. We develop theory and methods that allow us to demonstrate that partial orthogonality does indeed capture semantic independence. Complementary to this, we also introduce the concept of independence preserving embeddings where embeddings preserve the conditional independence structures of a distribution, and we prove the existence of such embeddings and approximations to them.

Paper

References (49)

Scroll for more · 37 remaining

Similar papers

Peer review

Reviewer zRa26/10 · confidence 3/52023-06-30

Summary

This paper aims to uncover the semantic meaning of embedding vectors within a given space. The basic idea is to determine a generalized Markov boundary by computing the cosine similarity of the orthogonality projected vectors within a subspace. The top K candidates are then selected. Furthermore, the authors provide a theoretical analysis of the concept of Independence preserving embedding in section five.

Strengths

- the connection between graphical model theory and explanation of embeddings seems novel (although similar ideas are present in other fields, such as the study of knowledge graph embeddings) - conceptually, everything is well defined and formally presented - The research question is clear and meaningful. - The structure of this paper is well-organized. In particular, in section two, the authors explain the necessary background information clearly

Weaknesses

- experiments and results analysis is rather sparse, the paper has more focus on the theory and definitions - experimental setup can be criticized (see comments below) - related work with respect to knowledge graph embeddings could be more thorough. Several studies have focused on using projection or rotation techniques for KG embedding to predict the missing relationship between two entities [THW, SLH, SW]. Can these KG meth- ods be adapted to uncover meaningful word embeddings? [THW] Yun Tang, Jing Huang, Guangtao Wang, Xiaodong He, and Bowen Zhou. Orthogonal relation transforms with graph context modeling for knowledge graph embedding. arXiv preprint arXiv:1911.04910, 2019. [SLH] Tengwei Song, Jie Luo, and Lei Huang. Rot-pro: Modeling transitivity by projection in knowledge graph embedding. Advances in Neural Information Processing Systems, 34:24695–24706, 2021. [SW] Baoxu Shi and Tim Weninger. Proje: Embedding projection for knowledge graph comple- tion. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 31, 2017.

Questions

- What is the critical motivation for using Image-language pre-train model CLIP as embedding? It would be better to compare it with other embedding techniques such as GloVe, word2vec, and BERT. And How does the dimensionality of the embedding vectors decide the performance of the proposed algorithm? - This paper only provides five examples in Table 2 regarding meaningful semantic evaluation. Would it be possible to compute the precise numer- ical results using semantic evaluation metrics to show the advantage after orthogonality projection? - In section five, the author introduces the concept of IPE, which might be more straightforward to understand through toy examples or visual figures

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

2 fair

Limitations

No

Authorsrebuttal2023-08-16

Thank you again for your review and feedback. Do you have any additional concerns or questions? If you are satisfied with the response, we hope you will consider increasing the score.

Reviewer DCJZ4/10 · confidence 3/52023-07-02

Summary

This paper investigates the relationship between the semantics and linear algebraic structure of token embeddings. It proposes utilizing partial orthogonality to define the "Markov boundary" of token embeddings. Given that token embeddings have limited dimensions and the Markov boundary can consist of numerous embeddings, the authors suggest relaxing the definition of partial orthogonality. They subsequently introduce an approximate algorithm to identify this boundary, which iteratively locates embeddings with high cosine similarity to the target vector after projecting onto orthogonal complement subspaces. To justify the effectiveness of vector space, the authors present the concept of independence preserving embeddings, which serves as the foundation for studying linear algebraic independence in embedding vectors.

Strengths

1. The paper formally discussed the relationship between meanings of tokens and their algebraic independence.  It generalizes the idea of Markov boundary and relaxes its definition so it can be practically applied to word embeddings. 2. To validate the use of linear algebraic independence relationships between embeddings for studying their semantics, the author introduces the concept of independence preserving embedding. This concept demonstrates that embeddings maintain the independence structure of distribution, making the paper comprehensive and self-contained. 3. The authors conduct experiments using CLIP embeddings and demonstrate that their algorithm effectively identifies intriguing patterns between word embeddings, indicating that these embeddings possess semantic meanings.

Weaknesses

1. Although the authors aim to study the independence relationship between word embeddings, they do not provide an evaluation metric to substantiate the effectiveness of the proposed method. The experimental results are presented as case studies with a limited number of words as examples, which may not be compelling for readers. A more robust experimental section would be beneficial. 2. The experiments conducted in the paper focus solely on the CLIP embedding model. While it is understandable that CLIP, being trained with visual information, may encode semantics that are more meaningful to humans, it would be interesting to explore whether the proposed method can be applied to other models that rely exclusively on text-based training.

Questions

1. The embeddings studied in this paper is static embedding. I wonder what if the embeddings are contextualized? For example, what would happen if the embeddings are processed by Transformers? 2. Throughout the section 3, the definitions of d and n are confusing. It says if d is smaller or equal to n, vectors are linearly independent. Should it be the opposite? And the same things happens to line 132. 3. Typo: In line 96, period should be replaced by comma.

Rating

4: Borderline reject: Technically solid paper where reasons to reject, e.g., limited evaluation, outweigh reasons to accept, e.g., good evaluation. Please use sparingly.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

2 fair

Presentation

3 good

Contribution

3 good

Limitations

N/A

Authorsrebuttal2023-08-16

Thank you again for your review and feedback. Do you have any additional concerns or questions? If you are satisfied with the response, we hope you will consider increasing the score.

Reviewer DCJZ2023-08-17

Thanks for the response and the new experiments. It addresses some of my questions. However, I believe a high-quality paper should be tested systematically on a large corpus with scientific metrics, instead of the case study. I would argue to keep my current score.

Authorsrebuttal2023-08-18

Thank you for your valuable feedback! We completely agree that testing the semantic relevance of learned Markov boundaries systematically with better metrics would be ideal. However, the dilemma we have is that, as far as we know, there are no good metrics in the literature. And we believe that large-scale evaluation with convincing metrics is an open problem and deserves its own work. Although there are existing metrics on semantic similarities that are based on Wordnet like Wu-Palmer similarity, they do not fit our experiments on Markov bounries. For example, the Wu-Palmer similarity score between “**eggplant**” and “**purple**” is only 0.167 but the score between “**eggplant**” and “**lemon**” is 0.667. However, because we want to construct the Markov boundary to be a minimal description set of the target word “**eggplant**”, one would expect to include “**purple**” instead of “**lemon**” despite what the WP score suggests. To see this, we asked ChatGPT to come up with a short description of the word “eggplant” and the answer is “a **purple** or dark-colored vegetable with a smooth skin, often used in cooking and known for its mild flavor.” This also fits human intuition. The previous example showcases the difficulty of coming up with good semantic metrics. Nevertheless, we humbly argue that the experiments in the paper, along with the newly added ones in the rebuttal, support the claims made in the paper. In particular, the central hypothesis of the paper is that partial orthogonality, and its byproduct Markov boundary, of embeddings, carries semantic information. To verify this claim, we provide both _quantitative_ and _qualitative_ experiments. For qualitative experiments, we appeal to human intuition by comparing the principal angle between learned Markov boundaries of the target embedding with linear subspaces spanned by relevant word embeddings of the target. For quantitative experiments, _unlike qualitative case studies_, we come up with a numerical estimate that calculates the average principal angle between learned Markov boundaries and embeddings of target descriptions from a dataset we created that consists of target words and their corresponding short descriptions. Although we do not claim the immediate practical impacts on a large scale, we believe that studying the Markov boundary of embeddings holds promise for understanding the inner workings of embeddings. This is one of our main scientific contributions and our empirical evaluations are designed to ascertain the utility of this claim. Because coming up with a good semantic metric is an open problem, we are of the opinion that it shouldn’t limit the contributions of our paper which is more focused on theories and definitions.

Reviewer 3K9o2/10 · confidence 4/52023-07-06

Summary

The central question (quoting the paper) is "How to make sense of an embedding vector in relation to other embedding vectors?" For that purpose, the authors propose to generalize the idea of the Markov boundary to embeddings, with a relaxed adaptation of this notion to cope with word embeddings peculiarity. The paper then introduces an algorithm to find (approximately) what is called the generalized Markov boundary for a given embedding. Empirical evaluations are carried out on CLIP.

Strengths

The real scientific goal should be first clarified before one can assess the strength of this proposition.

Weaknesses

Maybe I completely misunderstand this paper, but I cannot tell what is its scientific goal. For me everything is confused in the paper: the notion of markov boundary for vectors in the context of contextualized embeddings (like CLIP), why so much formal definitions for at the end a rough relaxation.

Questions

I have no question.

Rating

2: Strong Reject: For instance, a paper with major technical flaws, and/or poor evaluation, limited impact, poor reproducibility and mostly unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

1 poor

Presentation

1 poor

Contribution

2 fair

Limitations

No

Authorsrebuttal2023-08-16

Thank you again for your review and feedback. Do you have any additional concerns or questions? If you are satisfied with the response, we hope you will consider increasing the score.

Reviewer 1qJm6/10 · confidence 2/52023-07-07

Summary

This paper presents some theory and a method for reasoning about information gain in embedding space via a relaxation of conditional independence, as well as some theory on independence preserving embeddings. As information gain is inherently linked to independence, the paper focuses on defining a generalization of the Markov boundary that is meaningful in embedding space. The Markov boundary is a set of embeddings that "separate" the target embedding from all other test embeddings not in the boundary. Concretely, this means the cosine similarity between the projection of the target and test embeddings onto the orthogonal complement of the generalized Markov boundary should be 0. The proposed generalization relaxes elementwise orthogonality to distributional orthogonality, where the cosine similarities between the projections are allowed to cancel out, rather than all be 0. This criterion is motivated by practical concerns, where embeddings are low-dimensional representations where orthogonal residuals are rare. Finding a generalized Markov boundary for a single target embedding is then accomplished by sampling a number of random sets of embeddings, finding the top $K$ embeddings that contain information about the target given the random embedding sets, then constructing the boundary from within those top $K$ embeddings. Separately, the paper addresses another question of how to embed a set of independence assumptions in a lower dimensional space. A theorem is presented that shows that one can preserve independence assumptions as residual orthogonality to some degree, depending on the dimension. Experiments are presented, using CLIP embeddings, that show that generalized Markov boundaries can indeed be found, that projecting onto the orthogonal complement of the span random embeddings is meaningful, and that the discovered generalized Markov boundaries are more aligned with the span of embeddings of related words than unrelated.

Strengths

1. The first research question of reasoning about conditional independence and information gain in embedding space is appealing. Recent works in resolving ambiguity through dialogue, such as for image retrieval through 20 questions, reason over the space of individual images. However, if there are many images, this is not scalable. Intuitively, many of those images are likely to be very similar, motivating reasoning in the much lower dimensional CLIP embedding space. 2. The proposed definition of generalized Markov boundary and method for finding boundaries are reasonable. 3. The second research question about independence preserving embeddings is also worth studying for the same reason as above: potential applications would be very interesting.

Weaknesses

1. The method and experiments for the generalized Markov boundary only involve a single target embedding. 1. The experimental evaluation only evaluates token embeddings, whereas the text encoder in CLIP can encode sequences. An experiment involving reasoning over sequence embeddings would make the paper much stronger, especially if the experiment involved a realistic task. 1. No experiments are performed for independence preserving embeddings.

Questions

1. Primarily, I would like to see experiments verifying Theorem 13 on dimensionality reduction in independence preserving embeddings. 1. Is the only difference between kernel mean embeddings and independence preserving embeddings the choice of a kernel with finite dimensional feature map? 1. What is the relationship to work on information theory with kernel methods [2]? 1. A more realistic application that utilizes the method developed in the paper would greatly strengthen the paper. One possible application is an image retrieval game such as 20 questions [1], where the goal is to retrieve the correct image out of a set by asking questions. [1] White, Julia, et al. "Open-domain clarification question generation without question examples." Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing. 2021. [2] Francis Bach. Information Theory with Kernel Methods. 2022. ⟨hal-03577992v2⟩

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

2 fair

Contribution

3 good

Limitations

I did not find a discussion of limitations.

Authorsrebuttal2023-08-16

Thank you again for your review and feedback. Do you have any additional concerns or questions? If you are satisfied with the response, we hope you will consider increasing the score.

Reviewer 1qJm2023-08-17

In response to the rebuttal, and in hindsight, my original score is too harsh and will be raised from reject to weak accept with lower confidence. The paper supports its 3 contributions of generalized Markov embeddings, empirical validation, and independence preserving embeddings. My initial review was caused by a mismatch between the potential impact of the research question in this paper and the experimental validation presented. The paper presents theory towards embedding-based reasoning, and does not make a claim about large-scale empirical impact. The CLIP experiments in the paper are a small-scale study that validates the paper's claim. Further, larger-scale evaluation of the method is an opportunity for future work, and will not be held against the current paper.

Area Chair CD4j2023-08-21

Acknowledgement

Dear authors, Thank you for your response here. You have clarified the scientific goal succinctly and I assure you we will take this answer into account in the upcoming discussion and decisions. best the ac

Authorsrebuttal2023-08-21

Thanks for your reply! We really appreciate your commitment to the quality of the review process.

Reviewer zRa22023-08-22

Thank you for your reply, it will be helpful in the discussion about the paper.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC