Red Hat OpenShift Call for Code for Racial Justice | Red Hat Livestreaming, 28 Sep 2021

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: Call for Code for Racial Justice (miniseries): TakeTwo

Description

Join us to learn about TakeTwo, a Call for Code for Racial Justice Open Source project that identifies bias in digital content.

Bias in digital content such as news articles, social media posts, and even technical documentation can lead to overt or subtle racial discrimination. To help people proactively mitigate potential bias in their writing, a team of developers created an application called TakeTwo that uses artificial intelligence and an iterative learning approach to identify bias in written content. Hear the developers behind TakeTwo speak about the solution and learn more about how you become an active contributor on this open source project.

Sign up here to join the community and get started in Github: https://ibm.biz/OSTVAug31

A

A

A

B

Good morning good afternoon good evening and welcome to the very special mini-series that we are doing here with the call for code for racial justice team, I'm chris short host for the most of red hat live streaming, and I am joined by some of my friends over at ibm to talk about the project. Take two timmy: please introduce the topic of discussion today. Please.

C

Thank you so much thanks again for having us here again, chris, always a pleasure having or being here on your show.

C

um So today I have the take two team here with me as well, and we're gonna get a little bit into what take two is pretty soon, but before we get started, I just want to give a you know a quick intro to anyone who's not familiar with call for code for racial justice um if you've not been tuning into this openshift tv series um just want to give you just like an overview of what call for code for racial justice is, and what call for code is so call for code for racial justice is a component of call for code um and uh we'll get into that shortly.

C

Just again, just our speakers who are going to be speaking today, joanna naoki anna and nag nagma, who will be talking pretty soon uh just after I get this intro on call for code, um and we have a few additional members team members here as well. So many people have been participating in this project um from the beginning, and there are, some of them are still currently working on it and this list is continuing to grow. So if you see us in a few months, the list will be even longer than this.

C

We have lots of people joining the community um and participating as you as you'll see. We also have a lot of different roles that are available for people to uh contribute to as part of take two as well. So getting back to call for code um call for code is essentially uh one of the world's largest um tech for good initiative of its kind. So essentially we have a global challenge. Every year, where developers and non-developers can come together um and design projects that are solving specific social issues.

C

um So we've done a lot of different topics: we've done natural disasters, climate change, covered 19, and in 2020 we had a spot challenge for racial justice and out of this spot challenge, racial justice emerge call for code for racial justice, um and now we have seven offerings um that are related to this. So call for code for racial justice uh emerge in 2020.

C

As you might know, there are a lot of things happening in the u.s, specifically around race relations and uh there's a bit of a like a social and racial reckoning in regards to the um killings of specific unarmed black citizens um in the united states, so ibmers within um who are interested in um doing something about this issue. They came together and they decided to uh do a spot challenge around a call for code. So essentially they want to do something where they could feel like.

C

They were giving back and contributing to um eliminating systemic racism. So uh out of that emerged these three pillars: police and judicial reform and accountability, diverse representation and policy and legislation reform. um So during that time we were able to create solutions and create projects that aim to address each of these issues and actually funny enough. I had the pleasure of working with diverse representation at the time I was supporting that pillar. So I worked with another product manager.

C

Ian, who was leading take take two at the time and a few other projects so um and we'll talk a little bit more about take two shortly, but these were the three main pillars that we had as part of call for code for racial justice and just a timeline like it's only been a little over a year at this point, um since the launch of call for code for racial justice, so it was launched last year and juneteenth in june of 2020 with a spot challenge.

C

We had a lot of different participants who were involved um 500 volunteers globally, we had distinguished engineers master inventors.

C

As I mentioned, we created seven open source solutions, and now we have a backlog of about 20 solution, solution, building blocks that can be used for creating other open source projects, um and then we launched call for code for racial justice and externalized five of these solutions with the linux foundation um and we're able to externalize an additional two of them of in this year, and we have a really wide variety of things within our tech stack.

C

So if you're interested in python you're a front-end developer, back-end developer, there are different ways for you to get involved with our solutions here today and today, as I mentioned before, we're talking about take two so I'll. Let the team take it away to tell you a little bit more about their work.

D

E

Demi um so I'll take us off um so, like demi mentioned when we were presented uh with the call for code for racial justice challenge. um The problem statement that our team wanted to focus on was that bias can be learned and perpetuated in different ways. You know through societal beliefs, misrepresentation, ignorance and ultimately this can lead to inequitable outcomes and a lot of times.

E

What happens is that this is something that's unconscious, that people don't even realize that they are doing or perpetuating, and so we thought that this would be an interesting problem to tackle and look into deeper and see what we could do to help.

E

um So, if you go to the next slide, let me so in particularly we wanted to take a look at language right, so um I love this statement by our very own, uh dale davis, jones and humphrey, from ibm that words really they shape our world view. How we regard others and how we make others feel, and so language is just such a powerful part of society, culture, everything that we do, and so we saw that there was um this opportunity that we as a team, could potentially address this.

E

The issues that could be perpetuated via language racially biased language and uh mitigate it, and so um after the murder of george floyd, you know there was a lot of scrutiny that was being said in the media, how it was being fed um why there was a racial bias, persistent in it and there's lots of educational opportunities.

E

And um you know people were just really sharing this wealth of knowledge and background as to where the stereotypes were kind of being perpetuated.

E

And um you know pointing out social media posts, news articles and, if you think about it, we're really all exposed to media news, fake news on an hourly like basis, and so it's inherent in everything that we read and um we want to be able to mitigate the perpetuation of these hurtful ideas and biases and our team.

E

We actually came together because we were often humbled um by those who were sharing the background and the use of certain words and phrases, and we saw need to even educate ourselves and bring awareness to uh question our own biases um in any content that we've created or or consumed.

E

And so um you know all the headlines that you're seeing kind of uh highlight that other companies and organizations that people were all kind of thinking about this, and so um some of you may be aware of these. uh These movements that have been happening um for the removal of uh racially biased or insensitive language from different products, services, etc.

E

And so um you know we, we think it's just so powerful to understand the background and the words of the phrases that we choose to say as well as continue to say, and so we really operated under the premise that awareness is key and that we believe that bringing awareness uh to when language is racially biased and why would help mitigate content content that is being created from people just not realizing or knowing that? What they're saying is consists of racial bias and therefore it can minimize bias.

E

That's perpetuating via the content that they're creating and then ultimately reduce the offensive and harmful content. That's out there um yeah.

C

Actually for you here um so did you guys notice any or have you seen any work that talks about? Maybe the other different strata of types of biases, like in terms of um you know, for instance, like what someone says on social media versus a news publication just in terms of the effect that it might have, or even the context or the construct of that or is there anything that you guys know about um how that might differ from different um in different context? Essentially,.

A

E

So uh from what at least, what I think I noticed was that I felt like in social media, there was a lot of um like people emulating things that they thought were funny like just out of like humor.

C

E

I think they didn't they were, they thought it was light-hearted and it wasn't really offensive, but people were taking offense versus. I think news articles were more so like I I I don't even know if they recognize that the way that they were describing um different populations was was just different and and harmful for, um like people who are in the black population versus like how they were describing crime, for example, and so it felt like in social media that it was more coming from like a humorous standpoint versus like in the media.

E

I don't know like for some reason and um just the way that they were describing um was uh was not not humorous, but just different. I don't know if anyone else has anything to say on that.

F

Maybe I'll comment briefly on that, so I think it would be an interesting uh thing. You know to consider in machine learning part of the solution if uh it can take advantage of that sort of. You know broad contextual information to boost the accuracy of um bias detection, but maybe I'll say more about that: okay,.

C

Cool looking forward to your section, all right so moving on.

E

Okay, so um we initially began with the idea uh to build a tool that would detect bias and content, but we quickly realized that at the time at least, there was no publicly available data set um for us to leverage, and so you know this kind of made sense.

E

Language is incredibly subjective, and while it is more straightforward to identify some overt forms of racial bias, it's a lot harder to capture the more subtle forms, so kind of speaks to what we just talked about uh within, like the different contexts like social media versus like news articles, gaming chat rooms, and so um you know this is really where our chrome extensor, our browser extension came into play.

E

um We wanted to see if there could be a way to maybe crowdsource this data via trusted contributors, and so um we wanted to our goal, was to really uh provide a trusted community of members with a tool to flag content that they were seeing consisted of racial bias and provide a mechanism for them to um explain why something was racial bias racially consisted of racial bias, um and so this would feed into our machine learning model which naoki will go into uh further detail on um and ultimately would serve as the data to um to help our machine learning model, classify when phrases or expressions within certain contexts may be racially biased and ultimately, this would then feed into our api, which johanna will go into a little bit more detail that we wanted users to be able to leverage to uh check their content to see if it consists of racial bias and therefore be able to feel empowered to know and then make a decision as to whether they wanted to continue using that language, or maybe there was a different way to get across the point that they were trying to make and to be clear, we understand that this is an extremely complicated task, that's dependent on context.

E

You know we don't really know what the ground truth is here, um but we we do think that we have a great starting point, at least to explore these different concepts and ideas, um and so that's kind of that's kind of an overview of our solution. Nice.

B

E

And so question.

B

Yeah, you said you had a chrome extension with trusted people using it. How did you make that happen right like? Yes, are you gonna get into that? If I'm asking too early just time.

E

No, um I can, I can definitely- um or you know others please jump in, but um so something actually something that we came across was um we wanted to balance um giving power to the community to flag content, while also recognizing that there are bad actors and that that we need to make sure that there are credible people marking this and that people have the correct or you know the background to be able to understand when something is racially biased.

E

And so that's that's the reason why we're calling out trusted contributors, um and so we are still kind of in the process of identifying what that looks like or who that is, and um our small team of uh original challenge. uh um You know when we were just started off. It was a challenge that we're not sure if we're even the people, to make that determination. We kind of want to talk to experts to help facilitate that and understand what that looks like, but um that's just a little bit of background on that.

B

Awesome. Thank you.

E

Sure um so I will pass it over now to johanna to talk a little bit about our solution. Architecture.

G

Thanks nagla, so hopefully everyone can hear me um so my name's joe and I worked as a developer on this really fun project.

G

So the solution at the moment- and it's it's definitely not the final, the final solution, but it's basically built up of three components and our replays are split into three to match the components that go along with this architecture.

G

So, as naga mentioned, one of the issues we found at the beginning was not having a data set that that contain that more subtle form of racial bias. So it's very easy to find toxic language data sets or racial slur data sets. They do exist and we can bring those into our solution and we have to use those um to sort of for training and things like that for training. Our models.

G

What we found more difficult to find was things like um like when nagoya was talking about news articles that might be more heavily um making it look like young black males are more heavily engaged in crime than the rest of the population, that's more difficult to to identify and there's not that many data sets that that kind of contain that information.

G

um So we so we decided. Okay, let's create something so that someone could privately if they're browsing, because people were already doing it on instagram on twitter, they were already um picking out a piece of content from someone's website or from the news and saying hey. You know this is not okay, um and this is why so we thought if we could give them the mechanism where they can actually do that via a private tool. So it's it's not something that anyone else can see. It's just privately highlighted they go into their browser.

G

They can be on any any web page. They want. They highlight the word or a whole sentence or paragraph whatever it is, um and that data is then via the api um posted into a database, a back-end database. Now, in our github kind of get up and running example, that database is a couchdb instance.

G

The overall plan for the project long term is for the community to own the data set via some sort of organization who are kind of impartial.

G

So we obviously um someone who's a kind of neutral ground for something like um say you know to to kind of, because it it would, it would basically be like this is the definition of racial bias. Some people might interpret it that way, so it needs to be kind of owned by the community.

G

um The api then calls out also to them the machine learning model. So if someone so basically at the moment, we have just a very simple text- editor, which we'd also like to turn into a chrome extension. um At the moment, what happens is on the client side. They can actually type content.

G

So if they're thinking of writing an article or posting something to social media, they can type what they want in the text editor, it makes call out to our machine learning model, which then will highlight any parts of that text or sentence that it feels might contain, or it doesn't feel that it finds might contain racial bias. It's not that clever yet um and then it will actually explain why it thinks that um it's very clever machine learning model why it's um suggesting that you might want to read word.

G

um We've also talked about offering alternative words um to things that people are using, and this could be anything from social media post all the way to um to code that someone's creating online.

G

So I'll just share my screen and then I'll show very quickly uh how it works.

G

Hopefully it hasn't logged me out.

G

So, just to recap: we've got three github repos. This is the repo for the api and in here, you'll find everything um the information about what what we collect um from the chrome extension. So um the categories that we define so I'll go into that when I show it um and then a getting started guide.

A

G

How to get up and running um today? It is not running on openshift, but we have run it on openshift um when we were building uh the solution as part of the code. This is just running on my on my normal machine, but you can run it in a in a cloud native environment.

G

um We've got the uh so this is the web api we've also got the data science repo and then we've got a repo for the chrome extension.

G

So the way it works is we've got this api running when, when you get the solution up and running and there's a number of different endpoints and they're used for various things and that's all explained in the docs, but basically you can get the marks that are in the database.

G

You can post to the database, which is what happens when if you post a mark, that's what happens when someone selects some text in in the chrome extension.

G

What I'll do is I'll just show the chrome extension so um in order to be impartial at this moment, I've just got an article about lettuce, um because I I don't want to um say what I think is racially biased or not, but basically the user can toggle on the cursor once they're logged in, and this will allow them to choose any words or phrases. So if I thought addition was a racially biased term, oh I've done that wrong.

G

Let's get rid of that. So it's a bit slow.

G

Let me just reload: it's gone a bit slow.

B

I'm sure you streaming, video has nothing to do with.

G

It yeah okay, let's just pick another one uh slightly. I can hide.

E

Okay and then you'll.

G

The user, once they've highlighted they'll, be presented with six categories, which is our initial idea was to have a category of racial bias. We've also discussed just having a a kind of boolean, so yes or no, it does, or it doesn't we're still kind of I'd, say figuring this one out, but essentially the user can select.

G

They can also again, it might just be really slow, but they can also change a category if they want. So. If I had selected stereotyping, I can change it. I can also delete a mark which will, which will basically- and it calls all those endpoints that I showed earlier in the api.

G

um What then happens is so that's, basically the chrome extension and you can see that it needs some ui design work, because it's extremely basic and actually what was quite cool about this is we used an already existing just general highlighter tool that existed for browsers, and we just took the um it was just like a highlighter tool that highlights stuff in yellow, yellow pen and we used that open source project and then kind of made it our own in order to suit this project, which was a really nice way of using something that's already out there in the open source, world and kind of modifying it to our own um kind of needs.

G

C

Before you leave that page I'll, just curious.

G

C

It's a little bit more about the chrome extension, so I guess it's kind of disappeared, but it has something about copying. What happens when you do that um copy all right, yeah, I guess, remove all.

G

What is copy yeah, so I think copio will take all the all the text that you've highlighted and you can probably paste it somewhere else like in a text editor. If you wanted to something like that, and removal will just remove all of your um all of your marks that you've made okay.

B

G

There's lots of work to do on this chrome extension, so I think that there's a lot if you look in the chrome extension repo you'll find a whole host of issues that um some of us have raised: okay and that need to be worked on nice. So.

C

I'm curious here is it like for centuries and human, the highlight doesn't tell you what the racial bias is or just told you that's highlighted.

G

So at the moment it just tells you what's highlighted, um obviously, because this is. This is the side where people are capturing that data. So we don't know what the racial bias is. Yet um we want to add like a a space where they can not only highlight, but then fill in some information about why they've highlighted it and that information could then be used by the machine learning model as well.

G

um That's all stuff that we kind of haven't got around to okay, but it's a good idea um so on the other side of the kind of api is, is our text editor and for now? I don't have it hooked up to a machine learning model?

G

I just have it hooked up straight to the database that I'm using, but the general idea is, I type in a sentence in the text editor, and it makes a call out to the machine learning model and it would say you know, you've got this word and we found out that we we think it's a racial slur and here's here's the definition of of what that is, and here's any additional information which might be things like you could. You could use this word instead um and, and that is pretty much it.

G

I've obviously got my database running in the background, but this is sort of not that interesting because, like I said, we really want this to be a community-owned data set and not something that one person just spins up on their own. There could be scenarios where companies want to use their own data for a specific reason, so I know ibm. They have a set list of words that that they would like that people should not use.

G

um We have done a project within ibm where, where, before they post content onto their ibm developer website, they used our tool to scan the content of the of the new of the kind of article and flag any words that were not appropriate to use and then sent a message back to the developer. Who wrote it so you need to you need to maybe reread the content and maybe change some of these things. So we have we have. We have deployed it um within ibm on like a trial project um and yeah. That's about it.

G

I don't know. If there's anything else, um I think there's one more slide that I probably need to talk to after before naoki, um which is about data authentication, so that obviously nokia will go into data authentication from a machine learning perspective, but there's some things that we've identified, that we would potentially do to mitigate um malicious actors or yeah or unreliable data.

G

One thing I forgot to say is that the the data is: um the user is obviously authenticated before they use the um chrome extension that is set up using one of ibm, the ibm app id service.

G

um Initially, we had the idea that we would ideally want anyone to be able to contribute, but I think the way it's moving now. I think it's probably was the wrong direct, the wrong choice.

G

um However, there could still be bad actors, even if you've gone through a very rigorous process so and also we want to kind of build up some consensus around the data that we get so we've we haven't implemented this yet, but these are kind of some mock-ups of the kind of thing we want to do, and this is very typical of something like what google translate does just asking the user, while they're highlighting text um serving them a pop-up say you know, click on the words that you feel like are racial slurs, so that maybe this data is already in the database, but we just want to verify it and try and build up that credibility same for um maybe serving them a sentence and saying you know which, which category of these do you feel is most appropriate just so we can kind of build that verification into into the process, um and we still need help with that.

G

Like many things um in in in ui in the ui space, um so I will now hand over to noki who will talk more about machine learning.

F

Awesome: okay, thanks joe! um Let me if you can go to the next slide, please yeah! So this is sort of a one slide depiction of our road map if you will or a kind of vision, cautiously optimistic vision of where we want to go in the machine learning component. So there's a essentially a list of different versions of machine learning. Component of with, with increasing sophistication, I guess, and starting from the left version. Zero will be sort of a normal machine learning version. Where you do uh you know, database or dictionary, lookup version.

F

One uh is, is a classical machine learning version where you can train a classifier that can detect what sort of expressions or maybe even long, you know expressions and sentences and paragraphs are biased or not from labeled data version. 2 is, is the most next. You know sophistication of that where you leverage the cutting edge, deep learning technology to do this in sequential context and and version 1 has been implemented in the repo and there's ongoing work on version two, um but the rest of the versions actually are not implemented.

F

Yet it says it's in the future. Intense uh version. Three we call the aix with explainable. Ai model is a variation of the previous two, where it not only can flag a whole, you know paragraph or whatever, as bias can also point to specific words, expressions within it that are responsible for that. You know bias or not judgment version.

F

Four is one that tries to address the markers credibility question that joe was you know talking about another mechanism to address where we try to learn the credibility of the markers in a crowd-sourcing context and and use that information to to build better and better uh machine learning models. Version five is an extension of that, where you use that credibility, information to actually actively choose and seek label data from the markers that are more credible and about the questions that they, you know the model. The models really want to know about.

F

So that's just sort of the summary and uh in the next couple of charts I will go into a little more detail about two of these versions version, one sorry, yeah version, one and version four um next slide.

C

Yeah, so just a question here before we move on, um can you explain just for the general audience what is meant by broad context versus sequential context,.

F

Yeah so well, maybe I shouldn't say broad, but it in context so in in joe's demo, she happened to be picking, you know, words to label, but in general you could just um label a whole context like a paragraph. So this is trying to um basically the version one is going to classify whole. You know paragraphs or whatever into bias versus non-bias and sequential means. It really is paying um detailed attention to the sequential context of um you know in a sentence structure and that that's something that you need uh deep learning technology for.

B

How I mean pardon my understanding or lack thereof, of machine learning in general, like how hard is it to train these models? Right, like this, has to be a very challenging model to train. I feel like.

F

Yeah yeah a good question. I think maybe I will say more about what we have done concretely and I think it will uh it'll come out. It's.

B

F

It's not done uh by means um uh done or complete. We have a lot more to do so, so this one is the first version that we have implemented and it's using relatively classic machine learning. Classification um technologies like support, vector machine and nybase classifier, uh combined with what we call backwards representation of text.

F

um So the picture on the right is sort of illustrating how this works um sentence is coming uh as data, and they are labeled in this case, they're, simplistically, biased or unbiased, and then these sentences or longer uh uh transformed into something a representation called a bag of words, which is essentially a vector of frequencies.

F

Of many many words, so these sentences are mapped into a very high dimensional numerical vector space like you know, not necessarily zeros and ones, but often zeros and ones, and then you can run run a classic machine learning, classifiers like support, vector machine and naive bayes, plus five to come up with a classification boundary, and this you do own a bunch of training data label terminal.

F

Once you have that classifier or the model, you can apply it on a new uh sentence or paragraph. In this example, it's a master replica which is also transformed to backwards, and then the first classifier is applied to that and comes up with the adjustment of biased or unboxed.

F

So that's how it works and it's being implemented using um scikit-learn library and that you.

F

So that's our first version um and if, if there's a question I'd like to move to the next one, if you can go to the next slide,.

B

uh Just one there's one question in check: could this solution be used for other languages or is that something that's interesting.

F

That's a good map, that's a good! uh That's a good point! uh One of the advantages of you know: machine learning, based natural language type methods is that you train it with data. So if you're training with english, they will learn about language with training with german will learn about german chinese, japanese.

F

E

And one thing to know, I think our solution right now is primarily is english and but also us, I think we're very we've been kind of u.s centric, and so that's definitely something that we've been trying to. You know consider moving forward, considering different cultures, different languages and all of that so.

B

F

Yeah yeah there's a lot more that you should and you can do about them. It doesn't all come you know automatically, but right, but the it's basically extendable. That's that's the option, so this version four that we call ensemble consensus learning. This is our vision uh as a or an attempt to address the issue of there being unreliable markers or even adversarial markers in in uh crowdsourcing context.

F

You know for the reasons that joe stated, I I we believe that brow sourcing is very important and and that's the way to go, uh but it's also challenging if you're collecting data from uh you know, thousands of um uh anonymous or or people from all over, uh you may not necessarily have all reliable people. But how can you you know? How can you try to detect that and and learn from or leverage that information?

F

So here's this here's an idea. We leverage a general idea called ensemble, that's a notion where you have a number many potentially many machine learning models or methods, and and basically you make judgment using the ensemble. That means, as shown in the equation on the top, basically doing averaging.

F

So you do like a weighted voting by all of the machine learners in your ensemble and make us an average judgment. So that's one thing about ensembl.

F

The other thing about ensemble learning is that you can also learn the credibility of those models by updating their weights, credibility weights, and you do that by this second um equation, where basically, it's um cutting up cutting the weight of a model. That makes a mistake so every time it makes a mistake compared to the ground with label. It cuts, for example, cuts the weight by half um so it actually, you can show that it quickly learns and zooms uh into to the uh you know good set of models.

F

So this is the general idea of our ensemble learning, but we we applied this in our context and uh there are a couple of things that we need to do. The first is that um for those models um we don't really want to have just the markers if you have 1000 markers and drain a model for each marker.

F

You don't have enough data to to for from a single marker to create a good model, so we instead create many many models, but you train each one with data from a subset, arbitrary subset of the markets, so that's sort of how it's depicted in picture, and then you start you feed random, subset of them all of the markers into different models. You could have hundreds of models for exploring the different subsets, so that's uh one way that one that differs from the genuine form of money.

F

The other difference is that we don't really know the ground truth for the same, for the very reason that some of the markers are actually unreliable. So how do we know when to to learn that um one of the models made a mistake? So that's where we use the consensus, meaning the rated average prediction of the current ensemble you treat as as, if it was the ground truth, and then you, you penalize those models that deviate from that. So that's why we call it ensemble consensus learning now.

F

So, with some assumptions about the reasonableness of the ensemble, you can show that you can hope to show that this also will quickly uh converge and and we'll quickly, zoom into those models that are built with credible subsets of credible markers.

F

So this is a this is an idea and that would like to take forward but uh sort of abstract. So if you can go to the next slide, I would like, in the next couple slides I'd like to show some concrete work that we have done. uh This is much of this is was done by pritika, who couldn't be uh presenting with us today, but um uh what we did was, rather than sort of you know doing all this in abstract.

F

We wanted to do it in uh concretely, so we found this data set called the jigsaw data set from the kaggle competition. It's a decent uh data set about ratio of biased expressions.

F

The data is, it has 1.8 million observations, but it's not just about race or um it's not just about um you know, black or asian, as as it's shown in the uh the the histogram on the left. There are many different kinds of um biases that are present in in the data, but by and large you know, building machine learning models to learn these biases. That's that's a similar kind of uh uh modeling exercise- and you know, irrespective of the exact type of um um you, know, uh bias we're talking about.

F

So this is a very useful data set. um Another interesting thing uh is, and if you look at so there is a column in the data which stands for severely talked sequoia and for now we're using that, as a you know, target to predict so we're trying to judge whether something's toxic or not um interesting. Interestingly, the sentences in data. If you look at the presence of obvious, um obviously offensive terms like the f-word or the n-word.

A

F

You can see this monotonic uh increase in toxicity as a function of the the frequency of of those terms. But if you look at some more.

F

For example, like a big dog, it's not that simple! So uh there's a lot more that you need to do uh to learn uh uh rather than just sort of counting frequencies of you know these bits. uh If you can go to the next slide.

F

This slide is summarizing the the results we have obtained thus far with this data with the the first version that I talked about, the support vector machine version, um there's a bit of an extension that we did called tf idf, that's kind of variation of the bag of words. um Let's I'm not going to go into detail on on that one, but um so far we've obtained uh the accuracy figure it's around 80, it's not satisfactory.

F

So this sort of in part answers answers your question, press that it is a challenging problem and we do need to leverage you know more sophisticated and cutting edge um methods to to improve. But uh it's a it's. It's not a bad start. We, you know now.

B

Yeah, this is awesome.

F

Yeah yeah and there's some um this uh sort of spells out the steps of this involved in this modeling like cleaning up the text or removing stopwaters um dealing with imbalanced data by down sampling, the non-toxic data and up sampling or the toxic ones and so forth, and and we used acgds a method for training, simple vector machine models and uh we're using a specific, you know uh version with parameters with that.

F

um If you can go to the next slide, um there's some take away from from this exercise. We've done both uh support, vector machine and knife base and, as you can see now, basically about 75 accuracy. So even with this kind of classical machine learning technologies, some are better than others, and- and you know, you can hope to keep improving um and uh deep learning. For example, I think will improve significantly one of the indications that the model as of now is still not satisfactory is is.

F

This is the fact that it's a little bit technical, but the optimal loss function of the model is something called the modified uber loss, which means that there are outliers in the data which, which means also that, with the the information present in the data in in in the sentences, it's not enough to be making reasonable classification for for some of those ones.

F

And- and so we expect that you know- maybe it's a sequential context- or maybe it's some other kind of contextual information that that's lacking in the data uh that you would need to to be able to make better judgment of those. So those are some takeaway lessons.

F

Another comment is that this data comes with, as I mentioned, these different kinds of discrimination like against asia and against black weapons, we're not necessarily trying to predict those those categories, but they could be uh useful as features because the type of offensive expressions might depend on those categories and it could be helpful in refining our models.

F

Another thing is that this data is: it comes with the marker id column, so you do know which are which sentences or which examples come from a particular uh marker, even though we don't know their identities, so we could actually to try to do this version. Four, the unstable consensus learning where you're trying to learn the credibility of the markets. It's not something we haven't done, but this data set actually is rich enough for for us to do something like that.

F

um But eventually we want to move to uh crowdsourcing where you can choose to collect more information, maybe demographic and other kinds of contextual information, about the markers um uh and so forth, and also the selective you know, uh sampling uh or the active learning. That's.

C

F

Possible with the existing data set, so we are, you know, I'm supposed to move to the cross sourcing. um Sorry, I read all kind of longbow.

B

Oh you're, fine, this.

F

Is it for the machine learning component? Are there any other questions.

C

Yeah I had some questions. um Can you talk a little bit of just about the computations required for this, especially if you're moving to deep learning? So how can someone? How easy is it for someone to do this locally on their computer, or what do you guys thought about that.

F

Yeah, so so the classic machine learning models that we've done. I mean that it's easy to do like on my laptop um and uh training takes. Maybe you know seconds, so that's that's not a problem, but uh deep learning. If you really, I think, maybe at this level it wasn't uh that bad. I would have to check with ritika.

F

She probably did it on her own laptop, but as this data size increases- uh and you want to train, you know for a longer period of time and so forth, you may want to leverage uh gpus and go. You know going to um more capable setup, but I think you could also do that with with uh cloud uh clusters. um Okay, it doesn't have to be ibm cloud, but I mean cloud would provide such capabilities and you can seamlessly train these models with no no particular difficulty problem.

C

Nice and what's the volume of data that you think you might need in terms of crowdsourcing, so like just a call to action to the community, how much more data do you guys think you need for your next phase.

F

I don't know it's difficult to say how much I mean this data center has 1.8 million, which is just a decent amount, but it doesn't uh it's it's not restricted to the type of racial uh bias that we're interested in so that that's a lot less, um but with prosourcing we have the ability to select the markers and select the data that we want, so that would reduce the amount of data that we need. um But um generally I mean I know, hundreds of thousands would be okay, okay, good.

D

To know, actually, I actually have a question.

E

um So I know we've talked about context and um you know where we're kind of flagging, and um I know in our solution- we've talked about like including where the data is coming from or like what platform we're on. um Did you see that that was a field in the kaggle data set.

F

That's a good question, I'm I doubt it I I haven't checked thoroughly, but I I'm doubting it.

C

F

Would that's an example of the kind type of additional information we would like to collect.

A

F

G

Sorry, maybe I can jump in there as well, like so um obviously, we've built in to capture where data comes from, but I think to demi's point about how much data do we need. I think I don't know what nokia's thoughts are, but even when we've collected enough to train an initial machine learning model, I think racial bias is something that the what is racially biased changes over time.

G

So it's something that we'd want to keep updating and keep iteratively improving over time, also, so that we can understand how the definition of racial bias might have changed over time.

A

G

Depending on political events, different types of situations that unfold, um so I think it's something that we obviously we need an initial set of data, but we want to keep that process going and improving and things like that.

A

F

Yeah, that's important that we, our models, are dynamic and able to adapt to changes in society. Also.

B

F

Okay, maybe we should move on to the next segment uh anna. I guess.

D

Yes, so um here are some of the immediate applications of ticks to that we're working towards uh text-based content like social media platforms, blogs, um but there's also future applications that we would like to take. uh That would like to have take to be a part of such as in the gaming industry and, of course, in other languages, and we're also actively working with different partners to realize and co-create a vision to grow. Take two even further um next slide dummy.

D

uh So here I would just like to highlight where we are headed. um We have picked up our community meetings, and these are some of the things we're focusing on right now: finding um a home for our data, updating the ui ux of api and browser extension, as well as the training ml models uh next slide, and so here is where we need your help.

D

As you saw, we have set up the foundation for take two, but even with the current functionality, along with a great team of contributors that continues to grow, we're still looking for people that can help. This is a complex project, a complex problem, we're seeking to solve and excuse me, a complex problem. We're seeking to solve, and so any contributions we can get would be really appreciated.

D

We want to make it as impactful as it can be, so we are looking for various various varied skill sets and there's many ways to contribute: we're looking for data hosting data scientist designers, a full range. So if you are interested, if you know of anyone else who may be interested, please check out the links that were provided in the chat as well as listed on the next slide demi, and this the links tell you how to get started where you can find all of our repositories um and all the information you need to contribute.

D

Thanks, uh demi go ahead to discuss the next steps.

E

D

E

I could sorry could I just uh pause it um and yeah.

E

I just wanted to say that um you know anna lifted off from a bunch of the different um kinds of help that we need and just wanted to call out that, while you know we talked a lot about like the technicalities of our solution, but I think we also wanted to call out that, um because it's a socio-technical solution, we really also need people who are experts in like linguistics, and um you know political science and media, and all of that so I think really, no matter who you are and what background you have.

E

What skills you have. You have a very valuable contribution to make to our project um technical, social, whatever it is, I think we're really kind of looking for all kinds of people in a broad, diverse community to help out. So um I think that's kind of beautiful and also um awesome about take two and so just wanted to call that out.

C

B

That's great very awesome. Yes,.

C

um So yeah, just in relation to that, we do have some opportunities for you to get involved. As I'll put in this next few slides, you can join our slack community. You can also join us for hacktoberfest, so we're joining hacktoberfest, which is a month-long celebration of open source contributions um for in october, from october 1st, through the 31st and we'll be participating with call for code for racial justice.

C

All of our projects we participating, including take two during this celebration, so we'll be having a kickoff event on october 6th at 5pm eastern you can watch it. You can register for the event at ibm, dot, biz, slash, hacktobercc, that is case sensitive or you can watch the replay if you're not able to catch it.

C

Live there, it'll be on the um at that same link so, and for here you'll be able to find out more about how you can contribute to our projects exactly how to get involved in hacktoberfest and just learn a lot more, whether you're, technical or non-technical, about ways you can contribute to all of our projects going forward, and this is a last call to action here.

C

um We'd like for you to join our community, there's definitely a lot of work to be done with uh take two so we'd like for you to join us, and if you would like to be a part of a community, you can do so by joining us um through our registration page, which is ibm.biz os tv0928, which is today's date when we're filming this um and with that you'll be able to join us on the call for code slack channel and once you get there, they're like going to be a million slack channels.

C

So the things that it's important to know is that you should want to join the racial justice, um take two channel as well as the racial justice general channel as well, um and with that thank you for joining us today. It's been really awesome. Having yeah to being here and talking to you guys about take two.

C

B

This is awesome, thank you so much for the work you're doing. Thank you so much for bringing this project to the channel and thank you so much for sharing your work with us. It's it's very much needed work and I'm passionate about it. So, thank you very much for doing it um any any other points you want to make just thoughts before we get off the air here other than.

A

B

All the stuff I dropped in chat and links and such.

E

I think just thank you again for having us and letting us share our story and um we're looking forward. um Hopefully, people can join us on in our project and the journey and wherever we're, headed and so um yeah. Thank you.

F

E

Yeah, it's a grow.

F

Growing community- and uh it's really nice to see this network- you know going so we.

E

F

Everyone to to join, join.

E

Forces yeah, it's cool, to see when we all started like as a challenge to where it's gone and and where we're sort of headed. So you know back last june to where we're at right now is, uh is really really awesome.

B

It's quite an achievement and just a little over a year I mean it's very impressive work. Thank you awesome. So thank you all for joining. Thank you audience for watching and stay safe out there. Folks and we'll see you soon.

F

A