27 Nov 2021
ISWC 2021 took place from 24 to 28 October 2021. Here I am going to mention the main trends of the conference and share experiences that I've had during the sessions that I've been on them.
The International Semantic Web Conference (ISWC) is the leading Semantic Web Community. It gathers the most advanced researchers in the world in the fields of the semantic web and Linked Data and Knowledge Graphs every year. This year, the twentieth conference was held online (supposed to be in New York) for the second year in a row. It was a very valuable experience for me to attend this event, especially because my paper had been accepted in one of the most visited workshops of this event. And it was a great opportunity for me to connect with this community.
This year's conference also consisted of many tracks like Research and industry tracks, workshops, tutorials, posters, and demos, the doctoral consortium. In each track, there were multiple sessions with concert topics like ontologies, Information Extraction, Question Answering, Logics and Reasoning and so on which we try to skim some of them.
The first two days were for workshops and tutorials. On the first day, the Wikidata workshop was held. My paper had been accepted in this workshop. In general, Wikidata is much related to my PhD research, and I tried to be there wherever there was a trace of Wikidata, and the Wikidata workshop was one of the main places. 13 of the 15 papers submitted to the workshop were accepted and they were presented in three sessions. The presentation was five minutes of pre-recorded videos and ten minutes of Q&A in separate break-out rooms. In total, we had between 50 and 60 audiences in the workshop, considering that it was Sunday. The atmosphere of the workshop was very friendly. I remember from the first session, the Lukas presentation about Wikidated 1.0, which was an RDF graph of Wikidata revisions, I mean the history of all Wikidata edits, and such a dataset will be very useful, even necessary for investigating Wikidata quality, which is the main part of my thesis.
From the third session, the most interesting presentation in my opinion was Filip's presentation about KGTK, which allows you to keep and query the entire Wikidata on your laptop. You know, Wikidata currently has more than 100 GB compressed, and maintaining a personal copy of that in a traditional triplestore will need a server-like requirement. KGTK is a very useful tool that uses several indexing steps and TSV files and a specific query language named Kypher (very similar to cypher) and at the end, the tool create a portable version of Wikidata that can be easily queried and used on laptops. (of course, powerful laptops!).
My presentation was at the second session. The paper is a quite initial effort for investigating the quality of Referencing in Wikidata. You might know that in Wikidata every single fact could have one or multiple references. The quality of these references has been rarely studied so we tried to statistically compare Wikidata references as a first step. We extracted 6 subsets from Wikidata based on 6 WikiProjects and we then performed a bunch of queries to extract statistics like the ratio of referenced facts, usage frequency of reference properties, and shared references. My presentation went very well and I received useful feedback from audiences in my Q&A and I've been encouraged to continue the research in a wider framework.
On the first day, a separate 6-hour tutorial on KGTK was held, which I've been in some parts of it. The KGTK tool, as I said, is very effective for analytical research on big knowledge graphs as well as for generating and populating custom knowledge graphs. I also heard a lot about this tool and its team at the recent ESWC conference and I am in touch with their team leader, Filip. The KGTK team has presented and published several useful researches using this tool so far, and it is a useful and evolving tool.
I started the second day with the Deep Learning for KGs workshop. In the sessions where I was in the workshop, I found the following two presentations very interesting. One was to use deep learning to examine the semantic validity of knowledge graphs. The authors examined the hierarchical structure of a knowledge graph using KG-Bert and Word-embeddings. We know that different graphs use different vocabularies but whatever vocabularies are, they all have a hierarchy of classes and properties. This study checks the degree of adherence to this hierarchy in the knowledge graph using convolutional neural network.
The second presentation was on the use of knowledge graphs as a training resource for neural networks and object recognition, and especially style classification in digitized paintings.
Overall, considerable part of the workshop was related to the application of deep learning in classification of real world entities or class hierarchy of KG itself. There were other papers like Understanding Class Representations, or another was Challenges of Applying Knowledge Graph and their Embeddings to a Real-world Use-case, all related to classification problem.
On the second day, I was also in the doctoral consortium. PhD symposiums have always been an interesting part of conferences to me where it is a little further away from the formal atmosphere of the main conference. The ISWC PhD Consortium was also very good. Vito Walter first talked about his experiences in his PhD. Walter said his original paper was twice rejected by ISWC and in the third submission it was selected as the best student paper. The following year, his thesis get the SWSA Dissertation award. Walter's story was a wonderful paths from despair to victory. There were many good questions in the mentoring section. You can see this post for example.
The main conference started on the third day. After welcoming and introducing, Yoelle was the first keynote speaker from Amazon. She had a very interesting presentation on the challenges of humor detection in smart tools like Alexa. Probably if you have seen Interstellar movie, in a scene, Coper adjusts the sense of humor of TARS with one sentence! But In the real world, humor detection is not easy, and the device's reaction may not be what the user expects. The problem is quite common in tools with human language interfaces. Anyway, it was a good presentation about teaching the device using knowledge graphs. There were discussions about using comedic texts or even stand-up comedy performances for training Alexa!
After that, I joined the learning from Wikidata session. This session had two nominees for the best research paper and best student paper, and in the end, the best research paper award went to one of the papers from this session. I think out of ten or eleven candidates for the best article in different tracks, maybe 7 of them were related to Wikidata in some way, either using as input dataset or doing an operation over Wikidata. Anyway, this session was quite related to my thesis and I asked speakers several questions especially about detecting human edits from bot edits in Wikidata.
Another session I attended on the third day was data transformation. There was another candidate paper related to Wikidata or in a better word related to Wikibase. Wikibase are a collection of software packages including the Blazegraph Triple Store, a web-based user interface, a Java toolkit, and so on combined with a dedicated RDF data model that powers Wikidata. This paper describes the experiences of using Wikibase for importing some parts of the European Union knowledge graph with an emphasis on Wikibase specific data model and references and qualifiers. This knowledge graph would be probably one of the use cases that I intend to evaluate in the context of references!
The fourth day was a more social day In the conference. The keynote speaker first talked about what the Semantic Web should look like in the future. On how this branch of science should evolve, should it remain as a provider of other fields or should it become an independent branch science.
The panel discussions continued in the same. Professor Berners-Lee, who is the inventor of the Web and the Semantic Web, was on the panel. It was a very interesting discussion, especially his advice to younger researchers to be curious and open new doors and increase communications to help Semantic web in its goals. It was very motivating.
I further went into the Data Analytics session, and there was an interesting tool introduced for big data analysis using graph-based analyzer tools and uses the Apache Spark tool to distribute queries. It was an interesting system, and more interestingly, we in the Subsetting project, are going to use Apache Spark in a similar way for submitting Wikidata.
I saw two very good demos in the poster and demo section. The first was the work of Aidan Hogan and his student tries to create a graphical representation of the dispersion of authors in computer science fields, combining Wikidata geographical information and author information in DBLP, which was interesting.
another demo that won the best demo award was a FAIR checking system for assessing ontologies to be FAIR. It is much faster than Mature FAIR checking and other FAIR evaluators.
The fourth day of the town hall was also held. My opinion about ISWC (and to be honest, I did not dare to say at the meeting for exactly the same reason that I am saying!) was that I thought the atmosphere of the conference was too formal and the opportunity for communication for younger researchers was very less. much Lesser than ESWC.
And we come to the fifth and final day of the conference where there was a lot of good discussion about the Missions Track of the conference which was reproducibility. There were many discussions on What should be the purpose of being reproducible? What helps it, for example, is sharing code necessary? Does anyone repeat experiments at all!?
I was also at the validation meeting on the last day, especially the winner of the best student research, which was a study on the extent of using semantic markup in datasets that are published already. A problem that I have also currently. For example, very few datasets have included license information in a machine-readable form in their pages, and one needs to search the web pages for keywords to find the license of the dataset. The third presentation was also interesting about developing a shape language for the property graphs with functionality like ShEx in RDF.
ISWC is really the game of the masters! It was a very interesting and valuable opportunity for me. From my side, I can say I learned that Knowledge Graphs are being more and more important in the field of Semantic Web and Linked Data. Wikidata is going to be the main Knowledge Graph for research and even industrial purposes. Machine Learning and Deep Learning are getting a very important role in the Semantic Web and vice versa. However, there are still lots of gaps and future work like the quality of data, evaluation methods, reproducibility, applications, etc.