By Nino Antulov-Fantulin https://orcid.org/0000-0002-4337-2475

Review Details

---

Reviewer has chosen not to be Anonymous

Summary of paper in a few sentences:

The paper describes 5-6 different similarities measures: token-based and edit-based.

Structure of the paper and technical quality is ok.
But, I do not see why this review makes significant contribution, in my opinion all things are technically correct but the overall
contribution is not on a medium scientific level.

For example, I do not see any reason why the examples for dot products should be presented.

In my opinion: the only way how this paper would deserve to be published is the following:
- make a real critical review by making an experimental comparisons of the similarity measures on large-scale datasets and pros and cons
- elaborate run-time and space complexity on a more professional level
- elaborate the real-time online setting for similarity measures
- references need to be corrected and updated
- formulas need to be written in a better type setting

Reasons to accept:

This review, from my personal view acts more like a good tutorial to the similarity measures than a critical review.

Reasons to reject:

For this kind of a review, the authors should have put more emphasis on the computational issues for large scale datasets.
Only thing that touches the large scale issues are run-time and space complexities, which are put to table 1.
But here I see a lot of technical issues: What is On(1) ? or O( |m| +|n| ) quadratic ?

Nanopublication comments:

Further comments:

By Izabela Moise https://orcid.org/0000-0003-0370-6749

Review Details

---

Reviewer has chosen not to be Anonymous

Summary of paper in a few sentences:

This paper aims at providing a comprehensive view of similarity measures used in clustering of large data sets.
The authors focus mostly on text and document data and address six different types of similarity measures.
In a first step, the paper proceeds to summarizing the underlying method for each similarity measures, its advantages and disadvantages and also specific application domains. In a second step, the authors provide a comparison of the presented metrics, in terms of the efficiency of applying them on large collections of text.

Reasons to accept:

Choosing the appropriate similarity measure is indeed a challenging and important step in any clustering process, and the authors invest significant effort into addressing this.

Reasons to reject:

There are a few categories of similarity metrics that are overlooked by the authors. For example, semantic similarity (Latent Semantic Analysis) is not addressed at all. Knowledge similarity that uses information derived from semantic networks (e.g. WordNet) are also overlooked.
Also, more modern techniques such as word embeddings should also be mentioned, even more in the context of large scale text clustering.

I believe the current work needs to be extended in order to provide a truly comprehensive survey of similarity metrics for text and document clustering.

The performance evaluation in Section 6 remains completely unclear to me. There are no indications of which data sets were used, or what was the testing environment. In this light, the results presented in Table 1 seem unreliable.

I also recommend to the authors to check the English misspellings and formulation of sentences in general. There are a lot of these throughout the text and it really distracts from the actual content.

Nanopublication comments:

Further comments:

Reviewer has chosen to be Anonymous

Summary of paper in a few sentences:

In this paper the authors present some text similarity measures in order to demonstrate which ones are efficient for clustering a "huge" dataset. This review paper starts with very oddly structured English paragraphs that make it hard to read and immediate fixing. Regardless of the writing quality, the authors do a decent job framing their argument and provide a good-enough overview of background work. The evaluation of efficiency is not sufficient to support any of the paper's arguments in my opinion, as there are no large scale experiments and comparisons.

Reasons to accept:

None

Reasons to reject:

1) Not innovative work, there are some eerily similar articles from several years ago:
- A Survey of Text Similarity Approaches (https://pdfs.semanticscholar.org/5b5c/a878c534aee3882a038ef9e82f46e10213...)
- A Review on Text Similarity Technique used in IR and its Application (http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.695.3407&rep=re...)

2) Quality of the paper is not up to standard with a journal publication

3) Not enough (or at all) experiments to prove efficiency on large scale corpus

4) The clustering aspect of the evaluation is missing (experimentally)

Nanopublication comments:

Further comments: