Vocabularies and Use of Machine Learning in the New IMOS AODN Portal

Um I'm Natalyia Atkins and I've got Yushwan Hu next to me. Um we're both here from the integrated marine observing system and we're here to speak on vocabularies and the use of machine learning in the new um IMOS um Australian ocean u data network portal. Um so this will be sort of a a short talk but we'll give a brief introduction to both the AOD and IMOS so there's a little bit of context especially for our international guests. um um in the Australian space um I'll speak on the AEN vocabularies concentrating on uh parameters and platforms and then look at uh GCMD so the earth science keywords from NASA mapping to the science keywords mapping to parameter categories and then I'll hand over to Yuwan who will talk about the use of artificial intelligence machine learning to improve discoverability of all AOD hosted and harvested metadata records. So just a little bit of context for IMOS. Um I know a lot of you in the room um are are aware of IMOS, but just for again we've got um some speakers joining us from Europe today. Um so just giving a bit of an overview of the IMOS um program. It's it's a federally funded research infrastructure project that's been running for over 20 years and we deploy um a bunch of sensors and instrumentation on a range of platforms um Australiawide um both um at the you know coastal level out into the deep ocean from sea surface level down to ocean depths and and we basically deploy um and utilize a variety of um instrumentation to collect um long-term marine data sets um and the Australian Ocean Data Network is is the data um management um capacity of IMOS. So um not only do we help to um make discoverable um the IMOS funded u metadata and data but we bring together um all sort of marine related metadata and data um that is um published u throughout Australia by various um organizations and um universities and the like. Um here at the AADN um we have a range of published vocabularies in research vocabulary Australia and today I'm just going to concentrate on uh a couple of uh sets of these uh vocabularies that relate um to discoverability of metadata and hence data on our current AOD portal and I'll focus on um uh a couple of uh vocabularies related to parameters and then a couple of vocabularies u related to platforms and these are split into two Um there's a sort of a category level uh vocabulary um and then a more sort of a discovery level um more granular level um uh vocabulary that we use. So sort of just to show you um how this works in our current ADM portal. So this is a view of our current um a portal which has been in um production for I think about 10 years now. Um I know the screenshot is small but basically um uh the process of discoverability of the different data collections through the uh AODM portal is that we use a set of uh facets uh on the left hand side. Um the first one in that is parameters and the and the terms that are there um are sort of at that category level and we also have the ability to filter and search by um platforms from that vocabulary and the terms that are there represent terms from our platform category uh vocabulary. Um just blowing it up a little bit. Um when you sort of drill down into those categories at least for parameters um you then uh are able to look at that that sort of granular um data set um relevant um parameter vocabulary terms. Um you can discover where those have been utilized in metadata records and the same with um platform. So you drill down into um down from the categories, the platform categories and you can also discover things either at the sort of level such as research vessel um or in some cases for example um metadata is actually marked up with specific research vessels. So you can discover and um find metadata and data um using those which is very handy. So sort of having a bit of a a look at um the usage of the a vocabulary. So remembering that it's not just IMOS funded data, there is the um AODN sort of partners um that we bring into that. Um several of our AODM partners uh utilize the parameter andor platform vocabularies. Um noting that um because there are um they were sort of set up to be IMOS focused, not all the metadata records in in these harvested collections utilize these vocabulies. But we're looking at um Australian Institute of Marine Science Ames has over 2,000 records. Um the Institute for Marine Antarctic Studies at the University of Tasmania has over 1300. And we also harvest records from CSIRO Marine National Facility um close to 5,000 there. So um some of our AADM partners do not utilize um our vocabularies um dependent because they've got different metadata schemas or just different requirements or internal requirements and so sort of looking at the um number of metadata records from the Antarctic division got about 2,000 GS Australia about 2,000 and then smaller subsets from u national computing infrastructure NCI and e Atlas basically that relates to the usage and uptake of the AON vocaperies. Um in the AODN space um as is common in um in the marine space there is a high utilization historically and still currently of the the science GCMD keywords from NASA. Um and the usage of that in in regards to our ADN partners is the Australian Antarctic Division are quite a heavy user of the GCMD keywords and there is a number of um uh IMOS AODDN um IMAS and CSRO records that whilst they don't utilize because there isn't relevant terms for example um the AODM parameter or um platform vocabies there is that sort of discovery um keywords that they use from the GCMD. So a a process that um we did and this is in relation to us looking at bringing together all metadata records from the AODN space and making them discoverable through um the AODM portal as opposed to our current system where there's a select number of data sets. Um uh just sort of going over just just sort of quickly um we audited our ADM metadata for GCMD terms. we generally generally did a list of the granular terms and you know quite a lot distant time ago um the the way that the GCMB keywords are sort of set up in that sort of hierarchical structure we had a look at the um the granular terms and then map these to the AODN parameter categories we identified um new categories that were required and then improved and and um published some new vocabularies for the at the category levels um and that was all in relation to our new AODM portal which is currently in a beta release. Um that's the URL there for you for you to go have a look at that and sort of just showing you there that the idea is that um in the new portal um someone you can um by selecting the filter the filter button you can actually look at u parameters and platforms um in the new portal and discover um metadata um by that way. So, so that's about the AADN usage of vocabaries, but what about organizations and meta data records that don't utilize either the AODN parameter vocabularies or platform vocabies or the GCMD keywords? And for example, we have the following um organizations that don't um that don't utilize them. And then there's also records that that you know there isn't comparable terms or something like that. So you know the and this is where I'm going to hand over to Yuan because this is where she comes into the mix and and how we've gone about um improving improving that. So I'll hand over to you. >> Thank you. Um so continue with N's talk about the GCMD work and AOD vocabulary work. I'm going to show how we use AI and machine learning to support the vocabulary supported discovery. Um so uh the key message here is that AI and machine learning do not replace a vocabularies or human metadata curation. Uh we use the vocab alium vocabulary as a structure and human curated metadata as a trusted example so that we can use machine learning learn from them and um we identified two discovery gate. So from the record side uh some records have good title and abstract but they do not have uh completed vocabulary information. So in that case if it is missing you cannot find it through the parameter or platform filter on the new ADM portal. And from the user side, users may not know the exactly words or the their scientific definitions. So for users, they may how to find the records when they tap the text on the search bar. So there are two research questions. The first is how do how can we help record reach the vocabulary and the second one is how can we help users to reach the vocabulary and machine learning helps as a vocabulary support from uh two set uh on the um record set it helps connect the records to vocabulary concept and on the user side it helps to um connect the user's language to vocabulary concept. So you can find from the left side um if a record have uh title and abstract but missing um vocabulary information. We developed the machine learning classifier to predict the concept which is based on the uh training example from the human created metadata and so that we can use that uh AI predicted field in the um parameter platform filter so that the front end can search with look into this filter and find the relevant records. And on the right side uh we use uh given a user tapped what we uh we used machine learning to build a semantic search suggestion so that to suggest users with the related concept and this is the first example about how we bridge the record to AODM vocabularies. So assume I'm a user and I select the um parameters through the filter and you can find in the left side in the right scale uh there's a record uh if I go to you can find the keyword is um this thing which suggest that um It looks good from the user side but the back end is actually um the machine learning predicts the missing field information and the search engine uh field uh filter by this field and then it return in the result. So that even if a record is missing such information it still helps uh to show in the search result. And also another important in point is AI do not rewrite the metadata. So if it is empty it keeps empty. We just add the um search result in the result page so that helps user to discover this metadata but not to rewrite the metadata itself. And another use case is for uh from the user side about how we can connect a user standing language to a vocabulary concept. So this example is for a non-technical user who do not know the exact term. For example, I'm searching with um underwater devices um because I don't know what is defining the AODM platform vocabularies. Uh and you can see that in the related suggestion it shows the concept that is uh looking for the meaning which is similar to users uh inquiry and another case is sometimes um uh text have different meaning in different context. So uh using a vocabulary driven search can helps us to focus the uh search within the marine science and AODN control concept context. Another case is when users search a technical user know the concept very well but they want to search with a broader topic. we still show this related concept to help user find related topic or a specific topic. So in this uh in this case user do not necessarily know what we our vocabularies they find before they start searching because the portal can help help get them to it. And in the end I want to summarize this uh in uh to show the main three principles we used to develop the um AI supported discovery based on the vocabulary work. Uh the first is uh the alium parameter and platform vocabulary provide the well controlled concept. Well the human experts provide the authority so that um it gives us the trusted example to learn and the machine learning helps build the bridge from the um user and the records to discover via this structure. So this is all of my sharing today and we will go to the question slide