Player is loading...
SolveHR: Natural language inference for professional search
Wednesday, Aug. 24, 2022 · 1:26 p.m. · 12m 25s
During ICC we will build an AI-based model which can flexibly match job descriptions (and selection criteria) to people's CVs. The project will deliver a platform which allows HR professionals to state their selection criteria preferences, and using natural language inference, to find a best-matching set of candidates (and transparently, provide a supporting explanation).
Embed
Transcriptions
Note: this content has been automatically generated.
00:00:03
i yeah
00:00:06
i. e.
00:00:10
i would offer so just like start a introduce my colleagues
00:00:14
so one garcia a full stack developer uh alex got to uh
00:00:21
yeah i engineer experienced researcher so we are so over each are
00:00:26
so i'm a research group leader here at e. d. up focusing on how to be the complex reasoning mortals so
00:00:33
the idea i hear a used to work on the problem that's
00:00:39
present in organisations that's a strategic either for organisations how to recruit good people
00:00:45
and what's surprising is that it's a a very concrete challenging for most of
00:00:49
these organisations so fifty percent of organisations had acknowledged that the struggle with that right
00:00:56
so and it's also very expensive process so it's estimated that
00:01:00
costs can range right and some types of higher can get up
00:01:04
to three to four times the salary of the higher person in
00:01:09
this means the annual salary right so it's a very costly process
00:01:15
in the mean cost behind it it's actually uh so what we call soft costs right
00:01:21
so sixty percent are estimated to be this
00:01:24
indirect costs essentially managers team leaders uh in teen
00:01:29
uh people assessing applications screening candidates scheduling interviews
00:01:34
a meeting candidates and making final decisions right
00:01:38
so in in any case is this doesn't work so what we uh try
00:01:42
to do bring this challenge is should apply what we consider to be a
00:01:48
quite strategic advances in uh the last three years uh imagining
00:01:53
processing to attack this problem in to make this a out
00:01:58
search more efficient right so we package that we
00:02:02
integrated this key technologies uh specially the most recent ones
00:02:06
uh we package at that into a single uh to that week also v.
00:02:10
h. r. and a of this too is should match the complex criteria so
00:02:15
what managers in team members know that they want for critique right you could be a a complex description
00:02:22
of a job search a very specific one and uh to match that with the
00:02:28
c. d.'s that pool of c. d.'s that you would have right the notion of seafood
00:02:33
so how to match that cement kidding right and all the challenges which are involved in that that's what we fall it's
00:02:42
so in terms of competitors in the existing alternative is incorrect is uh most of the the existing
00:02:48
internal tools are based on keyword search right external
00:02:52
tools uh are mixed right so job specialist job portals
00:02:56
uh at some of them are based on keyword orders are based on some
00:03:01
sort of semantic search but what into the sea here is a sweet ending
00:03:05
sample of features that combine exactly to match a complex right you're right this
00:03:10
is not just a semantic search uh in any language in a cross language
00:03:15
in a transparent way as well right so it's very important that
00:03:19
these tools are transparent by design right at the are emerging regulatory requirements
00:03:26
that essentially is screwed in icing all emerging h. r. applications which are
00:03:31
a i base to make them explainable like to know how the matching happened
00:03:38
so in terms of the narrative for the value delivered for this to if we take a fraction of the cost that
00:03:44
be estimated beginning so thirty caper higher right uh at in
00:03:49
if we estimate that we can improve their by twenty percent
00:03:52
uh that types of savings that we can get for higher up can be estimated
00:03:58
to be a six thousand per higher right away so i just b. c. v. s.
00:04:02
and actually this topic was proposed as a concrete problem by the c. v. s.
00:04:07
uh makes an average number of fires off eighty people per year right so the savings could uh
00:04:14
amount to half a million per year if these assumptions are correct right in these are just assumptions
00:04:21
okay what we accomplished in this time was really starting from scratch
00:04:26
uh in trained shoe uh which is the other what we
00:04:29
knew about this very recent state of the art in excellent processing
00:04:34
uh and we managed to bat specially one alex do the heavy lifting really doing the engineering
00:04:40
uh on this uh and essentially what we have and you see as
00:04:43
a demonstration is a search by meaning complex search by meeting like moody criteria
00:04:49
uh at the same time cross and multi language and transparent right uh what we
00:04:55
would like to do in terms of next steps we also established a road map
00:05:00
that was to push the sophistication for this quite year to
00:05:03
the next level right really allowing for example manages to say
00:05:08
uh for me it's more important to have this type of experience in numbers of yours
00:05:14
with this type of uncertainty or more or less i i'm flexible on that but this one
00:05:19
uh i'm very a rigid about that quite you're right so it's balancing what
00:05:24
managers know about their experience on what works for the teen right capturing that
00:05:29
yeah the corporate quite right here in matching with the c. d.'s that are coming
00:05:34
uh so what we estimate a is that in six months time we're able to the point that for example
00:05:41
if we should i think environment of the c. v. s. so to
00:05:44
make that work at the c. v. s. environment and with additional six months
00:05:49
uh we can make the two more universal rights for all other uses as well
00:05:54
so we mission here uh as a kind of nice to use these which was our driver
00:05:59
uh the the c. v. s. use case but the market for this it's quite extend so if you call the number
00:06:04
of companies that you have in season and medium large companies is around
00:06:07
eleven thousand uh in the u. k. just use other countries forty two thousand
00:06:12
and it's actually this type of solution feeds shoe a quite large spectrum of
00:06:18
large medium large sized companies they are competitors
00:06:22
in this space most of them are focusing
00:06:25
on transform it based semantic search right if you think about putting that flexibility uh together
00:06:33
so that's it uh from the beach side now we're going to jump to the them all
00:06:46
i
00:06:55
so uh uh uh uh on at all
00:07:02
i e. e. or one e. e. that
00:07:10
on that e. e. or not but
00:07:19
ah oh uh oh i wonder what the the ah ah
00:07:28
oh right or or or you know not not
00:07:34
let's deal for knowledge of all uh thirty five
00:07:41
that was the first worry that was five yeah one of the more
00:07:46
yeah one of if not more than ah or half life
00:07:52
yeah it's you know where we're going to be the the
00:07:59
but the v. on the word corn or go with the t. v. on
00:08:03
the former your well well one one you know ah ah ah file either
00:08:10
basically this little bit it off or or were
00:08:13
there with me he which far versus t. v.
00:08:17
mash recorder but you know the problem yeah v. and
00:08:22
before that i or the relevant or or or or you
00:08:29
oh he's from from uh_huh uh_huh are related with the bridge
00:08:37
now uh we can may fifth more are on the move or not
00:08:43
uh now we get in the the one more than the one on one with you
00:08:52
uh the um what we the the financial and the
00:08:58
the batteries that are the and number them or no
00:09:03
or a little with if you know me i oh my my my ah g. e.
00:09:14
however every other men i knew that we are not been up all day
00:09:24
what the market will be a bit on a five oh
00:09:32
oh i see the theory of our market in our
00:09:38
yeah the the you that are very well me i love you
00:09:42
more than that but with her now where we are on that
00:09:49
uh and ah farther or oh
00:09:54
people our office but yeah i or one
00:09:58
model i believe that the one person would be
00:10:03
and then or
00:10:07
oh oh oh or the or the already quite well for
00:10:14
for our or who has the file i will now uh_huh
00:10:21
but no relevant for limited and ah
00:10:25
our levee he that cometh to your criteria
00:10:30
yeah yeah and a lot of a lot of flexibility is probably a lot
00:10:35
or or ah e. v. for a little while at that moment the the t. v.
00:10:46
though the way you are what the
00:10:50
you don't well we who who are we to you but you get the
00:10:57
job i i a four oh yes and we you and ah on that
00:11:06
would only be that were the five you one
00:11:11
way you're gonna use the one oh yeah yeah ah
00:11:16
oh uh for uh uh uh oh yes the the door
00:11:22
you are you are a fan of the t. v. one bit
00:11:27
yeah yeah it's a for a while or have recommended off now
00:11:34
well i knew what he and he you know
00:11:39
ah v. e. well yeah we brought them oh
00:11:45
so this is good for one of his scalable
00:11:51
ah okay well all we only use english spanish is that ah oh
00:11:58
i work well why are you landed ball or any better than the ah m.
00:12:05
so you you haven't done on the t. v. for that but ah ah i don't have to be done
00:12:14
so it is always up to
Conference Program
LightAI4Comfort
Aug. 24, 2022 · 12:13 p.m.
119 views
LaternaMagica
Aug. 24, 2022 · 1:08 p.m.
118 views
