Fireside Chat: Interview with Manxing Du of Talkwalker

All right. In this episode, we're talking to Manjing Du from TalkWalker about all things uh AI and data science. So Man Sing, to start off with, uh tell us about yourself. What's what's your background? You know, how'd you get into data science and machine learning?

Yeah, so thank you for inviting me. Uh so my name is Man Singh, and I did my bachelor and my master uh in telecommunications engineering, actually. And then I did my PhD here in Luxembourg in machine learning. Um I started doing data analytics projects actually uh for my master thesis. Um so I did in uh research institute um of Sweden, Rice.

Um so in that project, um I analyzed um YouTube video, uh YouTube users watching behaviors and uh discussed the potential gains of caching the popular content uh in the local proxy cache for an efficient uh content distribution. Uh even though there was no machine learning related uh in the project, but that's my very first step of entering this domain. Gotcha. That's very cool. So you would be telling telecom providers what to cache to reduce bandwidth strain?

Yes. Okay, very cool. And did they did that go into production? Uh no, no, not really. No.

Gotcha. Okay. Uh in terms of data science environments and things, you know, what's your favorite environment for working now? Jupyter, R Studio, and why? Uh so actually I use Python all the way.

Um, but sometimes for a very quick experiments or for data visualization, I use Jupyter Notebook. Okay. Why would you so what do you do your your Python development in? Is it just a straight text editor? No, I use uh PyCharm.

Okay, very cool. In terms of how do you decide when to do something in a notebook versus when to just write up straight up Python code? For instance, if I just want to quickly show, let's say take a look at the data and to see the distributions of the labels or to see some examples to check the features and so on. So that I would use the Jupyter Notebook. And to carry out like running experiments, I will switch to PyCharm.

Yeah. Okay. So uh talk a bit about what you do for TalkWalker. Um so I joined TalkWorker actually uh almost two years ago. And uh so in our data science team, we mainly work on of course finding AI-driven solutions for our products, uh ranging from image processing to uh natural language processing both for text and for audios.

And uh for me, um I have worked on improving our document type classification model um in particularly to identify news or blogs um or forum sites uh among others. And the rest of the time I have been working on an NLP-related projects, mainly processing text. And but that's work in progress, and these are not publicly released yet. And also I'm also working on some more, let's say practical issues, let's say how do we serve our model efficiently and to meet the requirements of the production environment? Can you talk a bit about sort of the evolution of natural language processing?

Like we all, I think pretty much everybody started with a bag of words and just doing very simple tokenization. And where is the field today? And how do you see the most recent big models like transformers? How do you see them being used? So this like big models like, for example, now very popular ones, uh transformer-based models.

Um the most interesting uh part for that model is um it used this contextual uh embeddings instead of bag of words, which only um embeds each words like independently uh regarding uh regardless the context. So in that case, um one word would have only one embedding, and um for contextual-based in word embeddings, um so if one word um has multiple meanings and they will have multiple embeddings accordingly, so it has a lot more potential and it understands um the semantic meanings of the word. Um, so it would help us to solve many real-world problems. How does that work with stuff like, for example, like hate speech and abuse of language? So for that, I think we have we call them like noises.

We have our noise control. So we will also, of course, train our model based on the context and then to understand the meaning and then identify them. And then of course, in our training data, I think before we do other tasks, we would do this noise control. We will try to filter out these noisy data first, and then we uh continue with other uh analysis. What if somebody wanted to specifically study like hate speech, for example?

Um, would they have to have a separate model that was trained specifically for it? Uh not necessarily, but um I would say we provide general models. Um, but if you want like a really domain-specific model, um, it is also possible to um train uh your customized model. Yes. How much how much horsepower does it take in terms of compute power for working with some of these models like BERT or GPT, uh the GPT family or the open uh the Illither AI family?

Is it something that a technically savvy person could do on a modern laptop? Do you need cloud architecture? Do you need just a room full of servers? Um you know for like epic training time? How what's what's the the uh overhead on these models?

Uh so I think uh if I'm not wrong I think some models uh if you load them it could be uh it could take up let's say 512 or like one gigabytes um memory and I think normally if you just want to run like a base model um it's a modern like laptop can can afford it and but of course for us we use um like bigger GPU servers yeah gotcha okay what are some of the more interesting um machine learning challenges you're working on right now uh so in general um the most challenging part is for instance how do I um assign labels to um unlabeled documents um for instance if you if you have um a predefined set of topics and you have tons of um documents uh how do you um assign the topic for for each document so a very naive approach would be let's say we define a few we find a few keywords related to the topic and then we could do keywords matching on on the documents and also of course if you want to go a bit uh one step further you want to find um the embedding of the document, and then you want to compute the similarities. And of course, when you choose the model um how how would you compute the let's say the document embedding? Would you compute word embeddings and aggregate them? Or would you compute based on sent uh based on sentence? So there are multiple choices.

And also, how do we, for instance, of course, we deal with global uh data and then the data documents would be in multiple languages, and how do we deal with that? Do you find like is there a substantial difference in terms of performance between using the more complex embeddings like from a transformer model versus just using bigrams? Um, you know, sort of going back to the naive approach, but using biograms. Um I never tried actually, but um I think because um, for instance, if we want to let's say find something related to Apple, I guess um the rather naive um word embedding models wouldn't understand, for instance, between the real fruit apple and um the Apple products, right? So I think that would be a challenge.

And right now I think the big more complex models, it can because of the contextual embedding, and it can understand the meaning of the word so it's um more powerful and more accurate. Okay. Um describe your exploratory data analysis process. When you get handed, say a new data set, uh, what do you do? What's you what's your recipe for unlocking value from a data set?

Um, so take right now, take this uh text data, for example. Um, we would check the source of the data set and if it matches our problem or not, because for instance, if the data is from social media or it's uh it's any like domain-specific data, uh or it's like it's from uh news website and so on. And of course, when we do um data cleaning up, and we need to maybe um translate the emojis into text and also remove some user account information, and also in this process, we need to try our best to de-bias the text as well. And um of course we need to also check the label distributions and to see um if any of the class, if any of the group is significantly um we have significant significantly more data than the other groups and so on, and also we can always run some simple uh baseline models on it and to quickly check the results and also identify let's say the misclassified documents and you to see which class we perform better and which class we perform worse. Talk a bit more about you said debiasing the text.

What does that mean? So, for instance, uh one example is um so emoji uh comes in um different gender and in uh different skin colors and so on. So we want when we want to translate the emojis into text, we will uh remove the gender and um the racial uh related text and to keep it as neutral as possible. Are there cases though where that those factors would be useful? Um yes, I guess so, but that's it's also always a trade-off.

So somebody who needed that data would have to do the the that data analysis separately outside of the the environment you're talking about. Um yeah, I guess so, yes. Okay. Why is that step in there? I'm curious as to like the decision making processes about why that's important or not important.

Um because I think we right now we don't want to make um assumptions or we don't want to confuse uh the model and it's very important to keep our data set neutral and clean. Um we don't want to introduce uh too much like bias into the data so the model may pick it up and may focus on um a wrong uh let's say feature um in the data to to make the decision. Okay. Um you mentioned labeling of of sources and documents. How do you differentiate?

Because there's there's a lot of I guess blurry lines. I'll give you an example. Um my personal website is listed in Google News, right? Now it's a personal blog. I would argue it's probably not a news source, even though it shows up in Google News.

How do you differentiate between news sources and you know some random guy's blog? Yeah, that's a very, very good question because it's uh very difficult for us as well. Um, we actually work very closely with our product team, and then we give a rather uh like detailed guidelines uh to label our data. For instance, um let's say if the um in a personal blog, if you are talking about news uh in a very objective way, and then we we may classify it as news, even though it's published on your personal blog site. Um, so yeah, it's it's um it also depends on what our like clients want.

Um so I would say it's um we need a rather clear uh in detail guideline to to label our data. How do you deal with objectivity issues? I'll give you an example. Um most of the planet agrees that uh Russia illegally invaded Ukraine, right? Uh it's generally accepted as true.

If you go to the official Russian news website, Rio Novosty, it's a completely different story. It's basically Kremlin propaganda. Um, but Rio Novosti would be classified as a news source. It is literally the state is the government official news source, just like the BBC is the government official news source of the United Kingdom. In cases like that, how do you deal with a site that is theoretically accredited but is completely disconnected from reality when you're talking about news sources and classifying something as a new source or as propaganda?

Yeah, so in this case, I guess it depends on what you want to use this. How do you want to use this data? So if you want to use it for, for instance, sentiment analysis, then I guess your data is highly biased. So I would say we will like exclude them from our training data because it's um yeah, it's it's highly biased. Okay.

Um go ahead. I don't know, it's okay. Um in terms of sentiment analysis, how what does the field look like right now? Because uh in a lot of the different uh studies I've seen and papers I've read, even with transformer models, it's still kind of a crapshoot. Um I would say um for us, I think we well, it depends.

You you need to if you use like let's say vanilla version of the model, um then like let's say BERT is not trained to do uh sentiment analysis, then of course you may not have the best performance there. And um, and also it's not really trained for uh sentence embedding, let's say, because it's better to do a word embedding, and then how do you aggregate them? Um I would say you need to find um that's why in TalkWorker we we um collect our own training data and also we customize our uh model and for like specific tasks. So in that case, we will make sure that, for instance, for sentiment analysis, we will um have better performance that than using um a model, we just use it uh just take it from the shelf. Gotcha.

And do you find that these models all how much how much human review of the training data is needed for a natural language processing models? Is it something it's not as easy, for example, as like saying, you know, taking e-commerce sales data? Uh that's much easier to model. Um so I guess we also so first we collect uh let's say from from some public data set, and um so we we know that this data, for instance, um, are used to build up some uh benchmarks. Uh so they are relatively uh reliable.

And also uh we will also make label some data by ourselves. Um so yeah, we have rather good control of our uh training data, and yeah, it takes a lot of time to to build up um our in-house data set. Yeah. Talk a bit about the the mitigation of bias in data sets. You mentioned obviously the debiasing of of some of the text itself.

Um do you is it a a valid approach in natural language processing to keep some of the demographic data and use it as a way to uh remove bias so for example let's say i have a hundred articles by a hundred authors and i have gender information um for the authors and let's say 80 of them are male 20 of them are female is it in terms of de biasing the data set there's obviously a a few different ways to do it one of the easier ways would be to take you know do something like propensity matching find the 20 articles are most similar to the the women's articles only choose 20 of the the the 80 men's articles uh but obviously you drop out a lot of information that way how do you think about the mitigation of bias particularly in the problems that you're being asked to solve um well that's a tricky question um tricky subject yes yeah so I guess I have also like watched some uh like talks about um trading bias and they said it's it's it's it's always it's always a trade off um between you don't want to remove too much of the um demographic information because you will yeah lose um a lot of information as well in that case um so I guess it depends on um your your task for instance you you can um keep all the data and then you do the training and then you test on your test set and to see um if you can observe any um mistakes, let's say. And if those kind of um demographical features uh really introduced um biased predictions, um then I would say maybe we need to deal with it. Otherwise the demographical information, if it um provides uh benefits to the prediction, then we we should keep them, yeah. Okay. Do you think though that and I don't mean uh talk walker specific, I just mean companies in general?

How carefully do you see your fellow machine learning and and data science practitioners thinking about bias and making sure that it's a step that they account for in their pipelines and even in their training data? Um I think because we are also fully aware of this problem. And um so for us, I think we always when we do um data collections and so on, uh we need to make sure the data sets are like diverse enough and we don't collect, for instance, from a specific uh domain or specific region and so on. Um yeah, so we we when we do when we build up our own uh training data sets and we are uh very careful and um try to prepare this rather clean and diverse training set. What do you how do you deal with drift when it comes to to models, particularly around dimensions like bias?

When uh let's say you calibrate a data set so that it returns the authors that are evenly split, 50-50 for gender, as a as a very simple example. But over time, just by nature of the fact that maybe you're pulling in, I don't know, uh accounting papers or uh or something, or pick a uh a domain where there's a there's a strong gender bias in one direction or the other, the model will inevitably drift if you just feed it the raw uh data. How do you how do you deal with drift in models? Um for us, so before we release our models, of course we will um test it in our uh production environment and using our production data and to see the prof to monitor the performance. And of course, later um if we have feedbacks from from our clients that um they are not satisfied with the results, and if they see some misclassified um documents and so on, and it's always possible to um label, for instance, as a domain-specific uh data set, and then um using our AI engine to retrain the model.

Do you how effective are systems like reinforcement learning um and active learning uh for these kinds of models in terms of getting feedback from customers, like if customers just thumbs up or thumbs down an article in the results? Um, how does that work as a feedback loop for retuning models? Uh so for active learning, I think right now we have, for instance, um, if we notice that um there are certain types of uh documents or a certain group of documents, they are they are misclassified. And then we would add those examples. Particularly we are going to target those examples and then add them into the training set.

And we try to learn from those difficult cases. What are the things that you say like, oh, you know, look out for this? Yeah, so I think the first step, uh, of course, right now we have tons of like big uh complex models out there, and it's very fascinating, and we all want to try them. But at the beginning, I think it is always beneficial to select a rather simple model. Um, it could be even a decision tree model, uh, to build your baseline and to understand your data.

And um, and also, of course, you shouldn't uh stop learning. Um, you should never stop learning because this is a really um fast-paced um area, and you should always keep up with the recent research. And also when you see sometimes the results are uh incredibly good, um, always double check. Always go back to see to make sure they are not too good to be true. Um what research and things uh are you keeping an eye on?

What things have got your interest that are on the horizon now that are obviously not in production but uh you know that have caught your interest? Uh for instance, right now, let's say we we need to train uh a model specifically for for each problem we want to solve. And of course, um GPT3 give us this opportunity to do this zero shot learning, and it can just we we describe our task and then the model will immediately pick it up and then uh give us give us the results. And I think in that domain, um there are still tons of things could be done, and also um how is it possible to use or even to downsize uh such giant model um into smaller manageable ones and use them in production? Um it's also a very interesting question.

What do you think of some of the more novel use cases of natural language processing to solve problems that aren't strictly language? Um there was a case not too long ago where someone took the sequence genome of of SARS CRV2, the COVID virus, um, transcribed it into essentially words, uh, you know, RNA fragments, just the letter sequences of the amino acids, and then used natural language processing to try and predict mutations with a fairly good degree of success. What how do you how much do you keep up with you know the way these models can be transferred from one domain to another? Yeah, I have seen those kind of usage. I guess you can also let's say apply an LP model in the music domain.

I think they are all of these usage are quite interesting. And then it also shows how powerful right now this natural language uh models are. Um yeah, and I think they are it's definitely these models have the potential to solve the problems in other domains. Do you think they'll be sophisticated enough at some point that we'll be able to use them, for example, to restore uh lost languages? Um yeah, I I guess because I think right now uh so these models could pick up um, for instance, some similarities between uh different models.

Um, for instance, um one multilingual model, um, if you train them on one task only in English, and then you test it on the same task but in another language, um, it's also it wouldn't give you like really top performance, but it's it's also uh the results are also quite uh are quite impressive. Uh so I think the model has the potential um to to pick up the links between the languages. So yeah, maybe why not? Okay. And what advice would you give to non-technical folks in particular when they're thinking about artificial intelligence?

Because they seem to have they fall in one or two camps. They either seem to be uh disbelieving of it entirely, or they think it's entirely magic and can do anything, including you know, create terminator robots and and other things. How do you talk to executive non-technical executives about what AI can and can't do? Uh so I think personally I I would say we should definitely definitely um embrace the enormous the enormous uh potential of AI. And um, but also at the same time we we need to be aware of the limitations.

AI cannot do everything. Um, for instance, right now the models um people are um mystically think uh the the models tells us um the correlations between features, but here the correlations um are not equal to uh causations. Um so for instance on Valentine's Day, and if you see, oh, we have rather high price for the roses, and at the same time we have also a very high sale of the roses, and they are highly correlated. And but it doesn't mean you cannot draw the conclusion that oh, so we should, in order to have a high profit, a high sale of roses, we should increase the price because high price is the cost of the high sale of the roses, which is wrong. So I think um here uh people should be aware of all these limitations, and also when you interpret the results, how to um explain how to understand the results correctly, it's also very important.

How do you deal with so with a model like GPT-3, for example, there is no interpretability or explainability of it? It really is very much a black box. Given the interest of governments and things, uh rightly so, about how machines are being used to make decisions. How do you deal with a situation like that? When uh when somebody says, Well, how did how did the model come up with this answer?

And you have this black box. What do you tell somebody? Um, yeah, so I guess this uh explainable AI is uh also a very hot research topic right now. Um so but I I guess for instance, um, if you look at a chat bot um or you let GPT3 to to write your story, um you can read the story and then easily probably tell, oh, this is not really a human-written uh text. It's it looks or it's uh it seems uh not consistent or um it rather looks weird.

So maybe you can immediately immediately uh see it's not generated, it's not uh written by by human so i would say in this case we are uh still a bit far away from the real let's say intelligent machine okay how do you personally and and and i guess uh from a a professional and corporate perspective plan on dealing with the absurd amount of um content that's going to be generated by a lot of these natural language generation models where they're gonna create you know instead of one really good blog post they'll generate a million mediocre blog posts that are you know that still meet their goals which is you know keyword density or other things for mostly for SEO but will flood all of our uh public commons I guess with with machine generated stuff that is okay but not great how you know how do you how do you see companies dealing with just this massive explosion of content um so I guess in this case um the first task is to identify um which text um are generated by machines and which are the the real let's say comments the the real um articles written by by human yeah I guess in the future maybe um the let's say the noise control um engine um should um should also um try to identify so so this is also one of the um major tasks in the future like to first to filter out um the machine generated text and then to yeah find um your uh interested uh human generated content. Particularly with comments though, like on product reviews and things, I I foresee that being really difficult because on one hand you might have a machine-generated comment that you know has might have a marker or two, like, okay, that that word choice is not how it um you you would normally say something, but it could be somebody who's not a native speaker of that language. And on the other hand, you have comments that are just put up by human idiots. Um I was reading a an Amazon product review saying uh the other day about uh a type of apple juice, and they're like, it doesn't taste like fresh apples at all. I'm like, it it it's not.

It's it's dried apple powder. Of course it's not gonna taste like you know, real apples, you idiot. You know, this this human just wrote this absurdly stupid comment on a on a product. But you could easily see that a machine learning model trying to understand comments, might actually think the machine comment was more useful and valuable, even though it's generated but not by a human than the what the idiot human wrote. And it poses this challenge, I think, of the machines might actually write better product reviews, um, but they're fake, they're not a real authentic review than what the human idiot wrote.

Um, how do you see companies try dealing with that, particularly a company like Amazon, where they're gonna have you're gonna have you know people who have a very strong interested in bombarding a product with you know as many f thousands of fake reviews as possible to to boost their ratings. Uh so I guess those um machine like the fake accounts, maybe you could also look at um their account names and find some patterns and also how often they post um you could i think from other aspects other than only looking at the the text they generated and also sometimes um this machine generated text they may put maybe put lots of let's say emojis or add links and so on so I guess you need to uh if let's say we can identify those comments uh easily if um then we should maybe filter out those comments and then maybe try to study the the the pattern and um yeah otherwise if if those comments uh if those accounts are even difficult for us to to identify then yeah how how can a um machine uh identify them right i mean that's the challenge i was having like uh did a real human write this because i can't believe well and i i looked care more carefully like you said looking for other reviews and like no this actually was a real just stupid person uh not a machine okay where can folks find out more about you and and learn more about the you and the work that you're doing um i think if you want to see my previous uh publications i think on google scholar you can find me yeah, and right now at TalkWalker uh we are not uh publishing like research papers, um but i think um you can always stay tuned with our product release and um see our uh new port acts. Oh that's over at talkwalker.com, right? Yes. All right, thanks so much for being on the show.

Thank you for having me here. It's very nice talking to you.