Review Comment:
(1) Originality: this paper presents an original, incremental work that (a) represents subjects and objects as vectors in high-dimensional complex spaces, (b) represents semantic and temporal relations as diagonal block rotation matrices, (c) transforms the real part of subject vectors with the semantic and temporal rotation matrices and compute the attended sum of both transformed vectors, and finally (d) use the attended vector to compute similarity with the object vector. This framework is an extension of rotation-based knowledge graph embedding methods. It enables modelling the interplay between the semantic of relations and the temporal conditions, i.e., some relations are time sensitive, and some are more permanent.
(2) The experiment results are comprehensive and show minor to moderate increments compared with previous work, especially with TLT-KGE, which also treats semantic relations and temporal relations as separate rotations in the embedding space. It indicates that the improvement comes from considering the interplay between semantic and time by using the attention mechanism, which is also discussed in ablation studies. In general, the experiments serve well to support the claims made in the paper.
(3) Quality of Writing. The paper in general flows well, but with some typos and less clearly explained parts. For example, on page 4, line 29, the "by the two matrix" should be "by the two-dimensional matrix", and the notations of transformed subject vectors $s_{\tau}$ and $s_r$ are bold in the text and not bold in the equations, which is inconsistent. On page 5, from line 37 to line 46, this part does not explain clear enough why the additional relation vector is introduced, why it is combined by element-wise sum, and how exactly it is learned (look-up table of embeddings? Linear layers?), whereas this part is proven to be very critical for the final performance in the ablation studies.
Additionally, there is no such “Long-term stable URL for resources” in the submission. Please kindly point to the code and dataset needed to reproduce the results if available.
As a conclusion, my opinion is that this paper needs minor revision.
|