A Semantic Role Repository Linking FrameNet and WordNet

This paper takes the output of a previously published system (one that produces a ranking of frequent WordNet categories for FrameNet role fillers) and converts it to RDF.

The modeling of the data is very confusing. In Fig. 1, LearntSemType is depicted as being of type fn:FrameElement, but the FrameNet model seems to describe a FrameElement as a class of entities that could fill the role.
Properties like hasTotalNumExamples, hasNumExamples were introduced as properties of LearntSemType, but they refer to a description level that describes the mapping itself with all of its idiosyncracies (FrameNet version, mapping algorithm, etc.). This description level is probably best kept separate from the description of the classes of entities that fill roles. I think it would be cleaner to have WordNet-derived classes as the classes of role-filling entities that are made subclasses of frame element classes or even just of SemanticType classes. Properties like hasTotalNumExamples, hasNumExamples etc. could instead be introduced as properties of the disambiguation mapping process, e.g. of reified RoleDisambiguation entities that connect the frame element class to the WordNet class and may differ for different mapping algorithms and FrameNet versions.

Overall, this is an interesting and unique resource to have available. I am not fully convinced that a simple conversion of a previously published system's output to a different format needs to have a separate journal article when this could just be put into a readme file. The conversion seems like something one can easily do in just a few lines of code. However, I do think there is some value in accepting this paper at this point in time because it shows the diversity of resources that can be converted to triples.

The majority of the paper was just copypasted from the paper "A Semantic Role Repository Linking FrameNet and WordNet" and the only new things added are related to the availability of resource and a very brief description of the quality of the data source and its possible applications. It is very hard to assess the quality of the dataset based on this brief description. The clarity and completeness of the descriptions is also very low. In particular there is no description or very scarce description on several issues that should be covered by papers submitted as Linked Dataset Descriptions papers like purpose and method of creation and maintenance, reported usage, metrics and statistics on external and internal connectivity, use of established vocabularies (e.g., RDF, OWL, SKOS, FOAF), language expressivity, growth, examples and critical discussion of typical knowledge modeling patterns used. The paper [5] - "Investigating the Semantics of Frame Elements" contains many of these descriptions, but then what is the purpose of the submitted paper. One possible solution to this problem would be to summarize all these issues described in [5] and present these brief descriptions in the paper. The quality of [5], which is a conference paper, is much higher than the quality of submitted paper, which should be a journal paper.

The paper "A Semantic Role Repository Linking FrameNet and WordNet.with the
given tracking number 402-1514” describes linking between FrameNet semantic
roles and WordNet synsets.
In addition to this linking, three statistics are also provided using the
following properties:
1. hasTotalNumExamples: shows the number of examples of an individual!
2. HasNumExamples: (again) shows the number of examples of an individual?
3. hasWeight: associates an individual with a weight
There is not any evaluation section.
This paper lacks a strong introduction part which the problem has been stated
and well motivated there.
Only very shortly it has been stated that “in order to tackle the problem
of role generalization and provide a semantic characterization of typical
role fillers”
What is this generalization problem? Role filler also has not been introduced
previously.
It is better that the authors elaborate on this problem and even add some
examples to make a better understanding.
Then, to resolve this problem the authors mention that “they automatically
created a mapping between FrameNet roles and WordNet synsets ”
How this mapping is done? In other word, which synset of WordNet has been
chosen? how the disambiguation process has been done?

About the evaluation: The authors need to present how much this linking
enhances the generalization problem? How well is the quality of linking?

Basically the statistic information can be retrieved by SPARQL queries. What
are the benefits of storing these information in the repository? And how much
these statistics can help to the problem behind the paper.

I would suggest that the authors employ a better terminology, they use the
terms 'individual', 'frame element', 'synset'....
a usage of terminology from Semantic Web, FrameNet and WordNet which is quite
confusing. Maybe they can define each term individually and then relate them to each
other.

Moreover, I do not find any relation between this work and this special issue
of the SW journal.
To sum up, I think this paper does not have any originality and novelty and
even the presentation of the work is poor.

Some minor tips:
It is better that the authors add the link of FrameNet web page to the first
page of the paper (https://framenet.icsi.berkeley.edu/fndrupal/home).
The paper can be better structured and formatted.

The authors describe the conversion of previous work (in which a mapping between FrameNet roles and WordNet synsets was created) into RDF. The conversion to RDF seems trivial, since the mapping was already undertaken and published in previous work.

(1) Information on the data set

The data set contains a mapping from FrameNet roles to WordNet synsets, as well as a minor amount of statistical information about the number of examples for a specific frame / its linked hyponym. The WordNet 3.0 ontology is used to link the resources to the English WordNet. Because there is no LOD version of FrameNet online, there are no links to FrameNet resources, only strings identifying frames. There also is no FrameNet ontology available, so the authors developed their own. The data set is not interlinked with any other data sets.

(2) Usefulness of the data set

Although the authors try to explain the usefulness of the paper, I don't think the data is very useful, as long as there is no published LOD version of the FrameNet corpus and the data links to this corpus as well. A large part of the data has questionable usefulness. The number of examples / total examples for a given LearntSemType is only relevant in a small number of
contexts, if at all, as it could be obtained from the published data. Links to the FrameNet examples could be useful, but can't be provided due to the aforementioned lack of a LOD FrameNet. So without FrameNet to link to, this data set's statistical information is rather pointless, besides its use for the authors java tool. But this statistical information is most of the metadata the data set contains. If one removes it, the data set enriches WordNet synsets with resources like this:

without descriptions of the classes itself, thus hiding its meaning to everybody without access to FrameNet. This comes down to very little data. Especially the missing links to other data sets enhance the problem of lacking usefulness.

The eventual usefulness of their java-based tool has no bearing on the usefulness of the data set as Linked Open Data. Its possible applications and uses are thus no arguments in favor of the data set. With this in mind, if we take a look at section 4 of the paper, the possible use of the data set isn't justified. There is nearly no profit over the original mapping done by the authors, neither for linguists, nor for the Semantic Web community.

(3) Clarity of the descriptions

The description is not very clear. Regarding the modelling, Figure 1 does not declare the namespaces, concealing that fn: is defined as the authors own namespace, not FrameNet's. This is only made clear implicitly in the text, causing general confusion. The authors were forced to act this way, as: "For the structure of the OWL version of FrameNet 1.5, we initially have chosen to rely on [3]. However, as the populated ontology is not yet available, we used FrameNet 1.5 XML representation as a reference for defining frame and semantic role URIs within the ontology."

Thus, it is unclear if the data will be useful when (if) an LOD version of FrameNet or a FrameNet ontology will be published. This raises questions about data set maintenance that are not addressed in the paper.

There also is no evaluation section, although the problems of the automatic mapping are mentioned in the discussion. The authors address steps taken to mitigate the problems, although it is not clear, why this should be done with their java tool instead of on the data itself.