VideoPoints

• ••• • •• • ••••• • •
Faculty Mini-Talks - Solorio Thamar's Computer Science Faculty Mini Talks 2022
Help Tour
© Distribution of this video is restricted by its owner
Transcript ×
Auto highlight
Font-size
00:00 my name is and I'm a professor the family computer science. And this
00:05 my mini talk to talk to you my research. So my research area
00:12 formally known as natural language processing. basically what we do is to design
00:18 development of computer programs that take us human language and perform some hopefully meaningful
00:27 task. And I think right now is becoming increasingly commonplace for us to
00:35 with technology developed in our field. you think about smart devices like your
00:42 or your google home, these devices are operated by voice um the field
00:49 natural language processing together with advances in learning machine learning and high performance computing
00:57 made it all of this together. it possible that we now deploy natural
01:05 processing systems or technology into everyday Um I want to start by discussing
01:13 research group. Um here I show of my PhD students. I have
01:18 new PhD students but I still don't a picture of her. So hopefully
01:23 I'll be able to add her to . And I have one master student
01:26 well um that are currently helping me progress on the research agenda. So
01:34 we have very little time, I'll give you a brief overview all the
01:38 of projects I've been working on. hopefully if you're interested, I want
01:44 learn more to shoot me an email we can talk about it A lot
01:49 the research that I've done um in last 10 years or so even more
01:54 have to do with the fact of to process automatically user generated data.
02:02 do we mean by that? Is generated in social media platforms?
02:07 We know users are pouring their lives into the social media sphere either
02:14 YouTube, you name it, And so there's a great value in
02:22 able to automatically process this type of . There's also a lot of potential
02:27 um good relevant uses of this technology some of the challenges that we have
02:36 we try to do that is a that because it's social media, it's
02:41 very informal general communication which means that will see a lot of banks.
02:47 there's a large vocabulary, there are radiation, there's, you know,
02:52 type of data represents basically a large of the population and therefore it represents
02:59 large number of topics. And so make it difficult for our technology to
03:03 able to work with this type of . So I work with my group
03:07 trying to develop technology that can process this type of data. Um
03:15 also I started working on multimedia type projects. What I mean by multimedia
03:21 that this is a project that not take text um as brother input but
03:27 morality, for example audio or in case video. And this is a
03:33 that is recently funded by NSF in with the research group from dr And
03:42 we care about there is can we users to make the decision on whether
03:46 want to watch the content that is provided in a specific video, for
03:55 , um by automatically detecting the type content that could be considered objectionable type
04:02 content that according to research or cultural um could be questionable. So so
04:10 have approached on that and so where research angle um or the difficulty lies
04:17 how we combine the evidence that you from the multiple sources. So we
04:23 text that is being provided from the on the videos. We have
04:29 right, that is being provided, have the images as well and all
04:32 this contributes to determine whether a specific can be classified as objectionable or
04:38 So, this is again a recent we're working on and on the other
04:43 , I also have research working on user reviews. So again, we
04:50 that before we do proceed with an purchase or to book a hotel or
05:00 go to a specific restaurant, some us read the product reviews, they
05:04 a lot of useful information. And for the longest time we have been
05:09 in the field in sentiment classification, ? So that we can understand it's
05:14 really positive or negative and and you have the star classification but going into
05:19 detail with that type of data, helpful that we can also distinguish what
05:26 the users talking about because for some , some aspects of a particular place
05:33 be more relevant um in the space restaurants for example, some people may
05:39 more about the price um some others the food and some others may be
05:44 or interested in the type of service they provide. Is it a fast
05:48 ? Said high quality service. And , but to understand which review is
05:53 about these things? Um it's a of work and so we have this
05:59 uh area of research where given an review like the one you see
06:06 we want to be able to extract aspect terms in this case the value
06:12 social service that are mentioned here as as associate an aspect category to each
06:19 these terms, for example, value to the price of the food dumplings
06:22 the food, sushi to the food service to service. We also want
06:26 attach a priority to each of these categories. Right? Because a single
06:33 is not necessarily always entirely positive or all negative. That could be most
06:40 there's some X most likely there's there's things that the person, like there's
06:44 things that person didn't like. So want to disentangle and be able to
06:48 process all this. So this is interesting project in my group.
06:56 I just want to mention a few and comments about the field. Um
07:02 I'm not being machine learning are super fields to be working on right,
07:06 room that we're seeing. Thanks to great progress of the learning approaches has
07:12 a lot of enthusiasm and a lot need for people that have expertise in
07:17 type of fields. So you're working this, you're studying this, you
07:22 have no um no lack of job . Your job prospects are very
07:31 And obviously going after the bigger model very exciting and interesting and we keep
07:38 that big tech companies are going after , bigger bigger models, but the
07:43 models are not necessarily the silver bullet all our NLP problems. So there
07:49 a lot of research angles that are important um that we can work on
07:54 that is not necessarily trying to go the bigger model. There are many
07:58 where these models are not solving the um to satisfaction, So and where
08:03 cannot address this with a big model we have, for example, under
08:08 languages, right? Where a Multilingual as big as it may be,
08:14 the model has not seen this the model will not perform well.
08:18 , so we still have work to specialized domains, the type of task
08:23 we care about in the real you are more likely not going to
08:29 able to use your big deep learning out of the shelf. So there's
08:33 lot of work into how do we tune or how do we prompt these
08:40 to work. Um and recently our has become increasingly aware of the dangers
08:46 this technology. So we care about well understanding biases, biases in our
08:51 , viruses, in our methods as as what what are the ethical implications
08:56 the research. And lastly there's a of need for explain ability because these
09:02 learning models tend to be very Understanding them tends to be equally
09:09 And so there's a lot of need explain ability. How can we understand
09:15 the model is coming up to this decision or the specific prediction. Um
09:22 in general if you care about this anyone of this or a combination of
09:27 topics I mentioned what I think are elements for success. Um if you're
09:33 under research studying language and domain task you know, there's there's no way
09:37 then creating resources, resources that will you motivate the research community, but
09:43 resources that will help you develop your research. Um there's a lot of
09:50 for creative people so that we can understand um find the interesting angle for
09:58 research problem but also creatively come up a solution that successfully and efficiently solves
10:05 problem. Um there's a need for as well because the field is moving
10:10 an increasingly fast pace and uh it people that that can stay on top
10:18 everything instead of top of how the is changing and a lot of self
10:23 . Because it just it doing research effort, doing research takes a lot
10:29 reading right, what is happening? reading papers um and sell you,
10:36 will find a need to be so if you want to succeed in this
10:40 . I always have some type of for talented undergraduate students and master
10:47 Um Not always, but unfortunately, I always try to get funds and
10:52 so if you're interested interested in joining group or learning what we do,
10:57 has showed me an email now, be happy to talk to you.
11:00 thank you for watching the video. you have more questions, you can
11:04 the groups, the group website, ? Um or contact me
-
+