Episode 33: BERT, Lead Form Extensions and bumper video BETA

Play this episode:

Or get it on:

• Apple iTunes
• Anchor
• Breaker
• Google Podcasts
• Overcast
• Pocket Casts
• RadioPublic
• Spotify
• Sticher

What's in this episode?

Mark Williams-Cook and Rob Lewis will be talking about:

Google BERT: The biggest Google change in over 5 years and one of the biggest leaps in search technology.

Google lead form extensions: Will Google's new lead form extension be the answer for lead generation sites?

Bumper Ads BETA: Talking about the new Bumper Ads in BETA for Premier Google Partners.

You can get the full transcription and links for this episode in our show notes at https://search.withcandour.co.uk

Podcast transcription:

Google Lead Form Extensions: https://support.google.com/google-ads/answer/9363648?hl=en Google BERT announcement: https://blog.google/products/search/search-language-understanding-bert Use of Transformers: https://ai.googleblog.com/2017/08/transformer-novel-neural-network.html Tensorflow Transformer example: https://www.tensorflow.org/tutorials/text/transformer

MC: Welcome to Episode 33 of the Search with Candour Podcast! Recorded on Friday 25th of September… It’s October… Welcome to Episode 33 of the Search with Candour podcast, recorded on Friday 25th of October 2019. My name is Mark Williams-Cook and again I'm joined by Mr. Rob Lewis and this episode, we've got loads of great stuff to talk about.

We are going to be talking about BERT, which is the new Google algorithm update. Google says it represents the biggest leap forward in the past five years and one of the biggest leaps forward in the history of search, so that sounds worth talking about. We are also going to be talking about new Google ads stuff in terms of lead form extensions and bumper ads BETA.

I think we're going to save talking about our friend BERT until the end of the podcast, I don't know how long I'll be going on about that for so it'd be a good idea if we kick off with our Google ads changes. So it was I think last week, we saw lead form extensions or form lead extensions going live on Google ads, is that right Rob?

RL: Yes, it is!

MC: Yeah so tell us about form lead extensions, what they are, what they do and why we should care.

RL: So it's a new ad extension that you can add to your Google ads adverts, that allows you to create a customiSed lead form, which will appear in the Google search results and because you can customize it, you can customize it to capture within reason any data that you deem to be appropriate.

MC: So that's directly from the search result basically? So, this is for websites that have the goal of lead form generation, so rather than going to the site and filling in the lead form, you'll be able to do this directly from the search results, that right?

RL: Yes that's right! There are some rules around which accounts can use the lead form extensions, so I'm just going to quote directly from Google; they have said, ‘To create a news early form extension on search, your account must be in a non sensitive vertical or sub vertical, some sensitive verticals and sub verticals are not eligible for lead form extensions.’

So, just to clarify, I looked through the policies and they deem health care, gambling, political contents and any kind of general misuse of personal information to form under the moniker of sensitive verticals. So if you're in the health care or gambling or political sector then you're not going to be able to use them, presumably because they worried that people would use the information for nefarious deeds or any kind of dodgy data capturing practices.

MC: Makes sense!

RL: Yep, so price-wise users are charged to standard cost per click for opening the form.

MC: Oh right, so not actually for form completion? So, that's kind of like cost per click in a way.

RL: Yeah. I'd call it more cost per engagement, so it might be that twenty people click on the form but not a single person fills it in but you're still going to get charged. But it may be worthwhile,you may end up having a higher conversion rate from people filling in the form via the search results rather than through your website itself.

MC: I think this almost goes along with the other 0 click stuff we see from Google now. So I mean, I was showing some people the other day how you can do a flight search in Google now and you can immediately interact with and search the data feed from flights, from the search result and you can actually go through, select and pay for the flight all without leaving Google. So these websites aren't getting the visits anymore and I guess for small businesses, it might actually solve some problems. So if you can drive leads directly from a site it kind of maybe takes away from the headache of doing the conversion rate optimization, making sure the site speed is good and actually building and hosting a website in a way because I assume it's gonna be quicker for users if they know what they want to fill out the form, directly from search.

RL: Yes. I think I’m in two minds about it, in my experience, speaking from a UK perspective in my experience the UK audience are a lot more wary when it comes to submitting information through lead forms and so this new extension reminds me a lot of the messaging extension, where you can send an SMS message through the Google ads advert, rather than having to go through the website and call someone. and I found that those extensions work really well in the USA, but they don't work very well in the UK and this is a really new extension, form extension so you know no one has much data on it, yeah I don't have any data at the moment, I only have a few clients where it's applicable to use.

MC: So do you think that's a cultural thing then between the UK and the US and the texting?

RL: I do, I do, there is a big difference, I mean this is a whole other conversation; there’s a massive difference between marketing to UK audience and marketing to a u.s. audience and I find that when you've marketed for so long to a UK audience when you suddenly have the opportunity to market to a US audience it's so much easier, for various reasons. But I just think that the UK audience are a lot more difficult, they're less trusting and so like I say this is a new extension and time will tell it may be that it performs very well, but I for one as a British consumer, I would not like to part with my details until I've had a chance to have a look at the offering in more detail. I mean what can you get from a small advert that doesn't actually explain the history of the company, doesn't allow you to to find out about trust, so I'm not saying it's not going to work, I don't know, that’s just one of the things I'm thinking about at the moment.

MC: I always find it interesting when you and when I've been on holiday in the US, even just seeing the TV ads, they're like shockingly different from the UK, in how direct they are and it actually pulls my attention away just over hearing the TV ads there. So it doesn't surprise me that there's a big difference with the messaging and possibly forms as well. RL: One of the things worth bearing in mind with the lead form extension is that you can actually link it with your CRM. So any leads that get submitted can be automatically pushed into your customer relationship management system and if you don't have a customer relationship management system or you don't have one that enables you to push the lead straight into it, then you can just export any leads that you generate through the lead form extension as a CSV file.

MC: Which people should be doing anyway, if they're doing lead gen you would hope that they have some kind of CRM system to track the quality of the leads as well as the quantity.

RL: Yeah definitely, I mean you could be generating hundreds of leads a month but not closing any of them through pay-per-click, so you absolutely need to be looking at the quality of the only those leads and not just optimizing for volume, but optimizing for quality as well.

MC: I think Facebook's got something similar, is that right? They have lead capture forms…

RL: It does and one of the things that I find works really well with Facebook lead capture forms, is that you can target on interests and general behavior which is a lot more difficult to do via Google, where it's a lot more rigid and you're showing forms to people on Google, well you will be sharing forms now or if the new lead form extension based on people who are searching for a product or service. So you're making the assumption on Google that they've searched for the service and they're ready to generate a lead there and then. Whereas with Facebook, you can show these forms in a more display oriented manor and you can target people based on their behaviors and interests. I find that you can actually generate leads at a pretty low cost per lead on Facebook using that. So it would be interesting to compare campaigns, Facebook lead generation campaigns with this new ad form extension on Google.

MC: It's good I think just to have the options anyway, you know if it doesn't work for you, you don't have to use it but it's there.

RL: On that note, interestingly Google have just announced a new change to Google search audiences where you can now target by interests. So I did wonder if maybe this is something to do with what they're trying to do now, which is where they're trying to give advertisers more of an opportunity to target by people's interests, rather than their actual shopping research behavior.

So just to clarify on on Google search, you can target people based on remarketing audiences, so people who have previously visited your website, you can choose to only show adverts to those people. You can also choose to only show adverts to people who are in the market for certain products or services, so they've actively been researching products. But up until now you've not been able to limit search adverts to people who show an affinity for certain interests, so for example you may want to only show ads to people who like dogs or who like cats or who are interested in investment news. In the past you could only carry out that kind of advertising on a Display Network really, they're on video advertising but now Google have announced that you can now carry out that type of targeting on the search network as well. So that's quite interesting and I would be really interested to play around the lead form extension based on people's interests rather than what they've been actively shopping for.

MC: It sounds really quite big to me because one of the issues for certain organizations running search campaigns is they may want to target a set of search terms and offer a specific product or service that's quite niche and actually there might be say a thousand searches for the thing they want to target, but only a small percentage of those are actually after the product they're after so it's a search term with multiple intent. So if you can identify that these are our customers and we just want to target these people that are likely our customer with that search, I think you're getting a lot closer to getting your ads in front of the people with the correct intent.

RL: Yeah absolutely I mean I'm just to give you an example, there's an audience interest in boating. So let's just say you’re a t-shirt manufacturer and you've got a range of t-shirts with boat nautical fins imagery on them, imagine the opportunities that opens up if someone's looking for cheap t-shirts and you happen to know that you can show adverts to people who are interested in boating then you know - that's just that's just one example.

MC: So you could jump straight to that category to show them that?

RL: Yeah. As marketers I think we're finding more and more, everything is to do with the intent of the person, rather than the search period they're typing in. Search query can be one thing but it's the intent, it’s the person behind that search. The more that you can customise sharing adverts to the right person, rather than the search query then the more likely you are to get a conversion.

MC: So, I mean I think the search query was where we started because it's the most basic clue for intent and now they're getting a bit more advanced, you can kind of sculpt away the rest of the wasted spend.

RL: Yeah, of course.

Another recent update actually - I've just been granted access to a beta tool that Google are trying called the bumper video ad creator, which allows you to create six second bumper ads from a much larger video, to save you editing the video into lots of different chunks. So if people listening don't know what a bumper video is, it's a really shorts in-stream video, so there's a video that shows when the displays just before you watch a video on YouTube for example, and they can only be a maximum of six seconds in length.

MC: I think that's about the maximum I tolerate not being skipping an ad!

RL: Well these ones are unskippable, so they're not charged on a per view basis like a lot of instreaming videos, these ones are charged on a cost per 1000 impression basis.

MC: That makes sense really because you can't, you know, six seconds - you can't really say whether they viewed it or not you know you're halfway through it three seconds, by the time they've woken up and they're desperately trying to find the skip ad link.

RL: I'm actually finding the lot of the video advertising I do now, I'm moving towards cost per impression against cost per view because the view free rate on videos is so high at the moment anyway that I find that you get a much better bang for your buck through cost per impression but that's going off on a tangent there anyway.

Anyway so bumper videos, I think they're great for keeping the brand in the forefront of the shoppers mind, particularly when they're utilized in like a sequence of videos. so showing, revealing different sequences of the videos in different segments in the run-up to a promotion, for example if there's a big promotion coming up in a couple of weeks showing a sequence of bumper videos is great to keep awareness in the forefront of the shoppers minds.

MC: So this sequence is the same as normal YouTube video sequences, right? When you can just have a set of videos and say, look you want to show this one first, then that one, is that right?

RL: Yes, but we're talking about bumper videos here, so much shorter.

MC: Yeah so it's the same but shorter videos?

RL: Yes.

MC: So I haven't looked at this at all yet, am I correct in that, you're uploading a longer video and Google through AI/magic is chopping it up for you and suggesting which parts to use as bumper ads, is that right?

RL: That's right, yeah.

MC: I'm really intrigued to see what the results of this are because that's got to be a difficult thing to do.

RL: Well it's interesting that you should say that because I have been playing around with it for the last couple of days.

MC: I wondered why you’d been laughing at your desk.

RL: I've had mixed results so, on one video I pulled it through and I don't know how it did it because it genuinely put together a sequence of about six clips, from a one-minute video showing just parts of the video that would work really well as a sequence of videos, in small six second bumper form. I tried another video and it just didn't work at all and I ended up with about ten versions of exactly the same six second clip, so this is still in beta, it's being tested by Google.

MC: So is it breaking the videos up with the intention of creating a set of sequences?

RL: Yes.

MC: Ah okay, so it's not just, here's some interesting segments, it's here's how to break the video up into six second segments.

RL: It breaks it up, it identifies various six second sequences segments and allows you to choose the ones that you like the best. But it does as you say, it analyzes them and chooses the parts of that video which it thinks, based on Google's infinite wisdom, which parts of the video it thinks people are most likely gonna get sucked into I suppose.

MC: So videographers beware yet or not quite yet?

RL: Not quite, yeah. As I say, I ended up with about ten versions of the same six second sequence in one of my videos, but on the other one it was pretty good!

MC: So maybe, fiver beware, but not professionals just yet.

RL: Yeah I mean what it's doing is, it's saving someone having to cut up an existing video and export them into different chunks of video, that then have to be uploaded individually onto YouTube for the purposes of a bumper advertising campaign. it's a lot quicker, it's really handy, okay it's not perfect at the moment but I'm sure they're working on it and it's going to be very useful moving forward.

MC: So BERT is Google's neural network based technique for natural language processing NLP, pre training and was actually a system on its own that was, it's open source and was released back in 2018 and I think it's worth, because I've just said that BERT stands for bi-directional encode of representations from transformers and then that is a neural network based technique for natural language processing pre-training, it's probably worth just covering some of those terms and what they mean.

So neural networks are basically mathematical models that are inspired by the human brain. think a common misconception is that they're like models of how the human brain works but they're not. They are basically systems that learn to perform tasks by looking at lots and lots of examples of those tasks, rather than being explicitly programmed how to do them. so rather than these are the rules of the game for instance, it will just observe lots of the games being played and infer what the rules are from that.

NLP, natural language processing, is pretty much what it says on the tin. it's just really the term that we use for computers trying to comprehend language, which is actually a really interesting and very deep, deep and wide layer, kind of, a pool of computer science because it sounds, on the surface, like quite an easy thing but understanding the nuances of language written language is actually a really difficult thing to do. I will include links in the blog post, in the show notes which is at search.withcandor.co.uk and that will give links to some of the specifics that I've mentioned if you want to read more about this.

The big breakthrough Google had with BERT was the use of transformers, so that's the T in BERT and the core idea behind using this transformer based model is what's called self attention, which is basically the ability to attend to different positions of what's being put in, so like the strings of text and from that computer representation of the sequence, so it means that rather than just analyze the kind of words on their own, it can look back at the position of those words in the sentence and advance its understanding of the sentence as a whole, the meaning of the sentence.

So if that doesn't quite make sense yet don't worry, you can listen to it again but I've got some examples that should hopefully, Google's given some examples, which should hopefully make that make sense to you.

So before they released BERT, Google was essentially testing different ways of trying to evaluate their understanding of language and in the blog post they actually list the different types of neural network approach that they used and they have a way of scoring the results which is called the bilingual evaluation understudy score or BLEU or as Rob very correctly pronounced it the ‘blur’ score. So this essentially showed them that, using this method which was a different method to ones they’d previously tried, they were getting a much higher fidelity on the understanding of language and they are going to essentially apply this better understanding to search, so I'm just going to now divert off to the blog post that I'll link you to about BERT to talk not about the technical stuff but about the application and what it's useful for.

So Google said, by applying BERT models to both ranking and featured snippets in search were able to do a much better job helping you find useful information, in fact when it comes to ranking results BERT will help search better understand one in ten searches in the US in English and will bring this to more languages and locales over time.

So I think firstly, there's a couple of really important things in there. So one, they're saying that this BERT model is going to impact normal organic rankings, your 1 to 10 rankings and it's going to impact how the featured snippets are going to appear in search. Normally, as we've previously talked about, they're fairly sort of separate entities in a way, so it's interesting that this BERT model, which I think it’s fair to say, is rather than having an impact on ranking how a web page might rank, it's got more to do with the selection of which page is going to rank. So rather than looking at this as if it's a ranking factor, it's not - BERT isn't a thing that's going to make a page go for in position six to four, it's more actually Google’s probably going to choose a different page if it thinks that's that's a more relevant result. It says it's going to affect one in ten searches which is actually a huge amount. So if that affected one in ten searches globally, that would be in the hundreds of millions of searches a day that that this model is going to impact. It's also interesting to note that they've said it's, at the moment, dealing only in English searches which is understandable, it's a good place to start because a lot of the web is written in English but my thoughts on this were because different languages can be structured very differently, they will likely have to train this model on every language because you know, even the basic languages that are very similar like, French and English are incredibly similar, but we do things like we use the order of adjectives and nouns differently, so I'll say the red table but in French, obviously it's the table red so I guess that this model has to account for that.

Going on Google says, particularly for longer more conversational queries or searches where prepositions like for and to matter a lot to the meaning, search will be able to understand the context of the words in your query. You can search in a way that feels natural to you. So they're saying while it affects likely one in ten searches in the US in English, it's much more likely to affect these longer search queries, which is actually something that the previous system, so we

If you work in an SEO you've probably heard a rank brain, which was Google's artificial intelligence attempt to specifically tackle search terms they hadn't encounter or search queries I should say, they hadn't encountered before, which was a massive still fifteen percent. So 15 percent Google said, I believe it was 2017, that 15% of searches they see on a daily basis are unique, are new and they haven't happened before. Which actually when you look at the math side of the amount of words people have and the ways they can use them, that's not unexpected because there's a huge amount of combinations in the way you can ask the same question. So rankbrain was kind of Google's AI attempt to actually, here's a way to apply models to better understand what people are asking when they ask us these and we've confirmed already from Google that rankbrain isn't going away, it hasn't been replaced but this kind of BERT model is there to support it and I guess when you get search queries, it's gonna go through some other system that's going to decide which part of the algorithm is going to give the best result for that query.

So Google says, ‘to launch these improvements we did a lot of testing to ensure that the change’s are actually more helpful’ - which is good to hear - ‘here are some of the examples that showed up during the evaluation process that demonstrate BERT’s ability to understand the intent behind your search.’ They've given the example here, they say ‘here's a search for 2019 Brazil traveler to USA need a visa - the words to and its relationship to the other words in the query are particularly important to understanding the meaning.’ So 2019 Brazil traveler to USA need a visa, as humans we understand what that means is, in 2019 someone from Brazil is traveling to the USA and they need a visa, so it's about Brazilian traveling to the US and not the other way around and Google says, ‘previously our algorithms wouldn't understand the importance of this connection and we returned results about US citizens traveling to Brazil, with bert search is able to grasp this nuance and know that the very common word to, actually matters a lot more here and we can provide a much more relevant result for this query.’ So they've given two examples of before-and-after BERT and as they said before Burt is a search about US citizens traveling, whereas the post BERT or after BERT is launched answer gives a US Embassy.gov site information about tourist and travel visas for people from Brazil - so it's a much better result!

It's interesting because if you look at guides and tutorials on, you know how to say build your own search engine and do text analysis, one of the kinds of shortcuts that they take, a very basic search engines, is they generally discard all of these words like you know - to, from, the, a, because they're just repeated so much and they add a lot of noise to search, so what Google's doing is they're saying they understand on a comprehension level that these kind of joiner words are and actually, there’s probably like a proper term in linguistics for what I've just called ‘joiner’ words, apologies if you know what that is and I don't. But it's showing that the understanding of those words and their importance in context to the sentence is key to comprehension.

They also give results where BERT doesn't work that well, which I thought was interesting. so they've said, search is not a solved problem no matter what you're looking for or what language you speak, we hope you're able to let go of some of your keyword ease and search in a way that feels natural for you. But you'll still stumped Google from time to time even with BERT, we don't always get it right. If you search for what state is south of Nebraska BERTs best guess is a community called South Nebraska, which you can kind of understand, you're saying what South in Nebraska and they said if you've got a feeling it's not in Kansas, you're right. So obviously Kansas's is south and maybe that's what a human would answer. Language understanding remains an ongoing challenge and it keeps us motivated to continue and improve search. We're always getting better and working to find the meaning in and the most helpful information for every query you send our way every day.

There's one more example here, so they've said here, another example where BERT's helped us grasp the subtle nuances of language that computers don't quite understand the way humans do and they've done a search for ‘can you get medicine for someone pharmacy’ - from that I take can you get medicine for someone pharmacy, as can you pick up medicine maybe for someone else at the pharmacy and Google was saying before BERT they didn't understand that the ‘for someone’ was an important part of the query, so before Bert results were giving results just about getting a prescription filled and after Bert they've given a result about can a patient have a friend or family member pick up a prescription. so it is actually quite impressive.

So it reminds me a bit of when we saw the Google duplex do the telephone call and then it was dealing, for instance, with heavy accents really well, because we weren't familiar with them we were having trouble understanding as well.

RL: Is the future of where this is going, moving away from Google finding the most relevant static search result that has the answer to the question you've asked and instead analyzing multiple sources and giving you the answer that you're seeking?

MC: Yeah, one hundred percent! So there's a lot of talk in SEO, around SEO is ASO, which is search engine optimization is answer engine optimization - oh sorry, AEO, not ASO. answer engine optimization. You know we see it throughout search now, if you Google for the weather you just get the weather result, if you Google movies you just get the results directing the search page where that movie is showing, what time it is, with Google shopping like with a flights example we mentioned earlier. Yeah Google's definitely going towards the, you know, it's all based around this time to result things; so how quickly can I give you the answer. So it can be sure that it's going to give you a good answer, then it will just do that. I think that's where they need our cooperation with stuff like schemer, where we can specifically label stuff on size to say this is definitely this, this is that - this is where Google sometimes gets its knickers in a twist is where it's left like you say, to decide what is the best result or what's the correct answer because if it's providing multiple search results, it's kind of hedging its bets and letting you pick. having a computer do that final step of saying this is the answer, can be quite dangerous obviously because if they get it wrong, it's normally very wrong and they can look really stupid.

We did that example with the ‘how many legs to certain animals have’ with the feature snippets and Google saying rabbits have 200 legs and ducks have like 40 legs or something, so there's a way to go I think on that. But yeah that's definitely the strategic direction they're going in and I think that likely outcome we'll see from this is, as users, we will get a more diverse set of search results, so maybe rather than the fewer sites that are more authoritative ranking for queries that maybe don't quite match up with the content, I would imagine we'll see a larger cross-section of sites that answer questions, basically better, that have the intent correctly matched up.

So what do you actually need to do then as an SEO? Because I have no doubt we're gonna be swamped over the next few weeks in the ultimate guide to optimising for BERT. the advice is pretty much exactly the same as rankbrain, which is there actually isn't anything you can specifically do to optimise for this Bert model apart from actually make better content, answer users questions and the stuff that we've been hammering on about for years now, this is you know the technology is catching up. 10 years ago Google used to tell us to write good content and do this and actually a lot of the SEO effort just went into building links because the very objective fact of the matter was that Google wasn't very good at telling what good content was, how to match up intent with with good answers, it was it was really heavily relying on links. So people just built links and they ranked number one and they were very happy about it. Google has kept fairly consistent in what they've been telling us to do and we are at the stage now where the technology has very much caught up with that guidance of, you know, think about what your user the users are asking, think about how you're answering it; by that I mean what format you're using, you know, video images, text and actually trying to make the best experience for the user, that's what all of these models are heading us towards.

So actually, while the answer is fairly simple, the answer is a lot of work - so you do need to do research, you do need to work on the user experience of the site, you do need to make sure you are providing good answers fast in a way people can understand.

RL: I figured as though this update to me, it seems like it's a really important point in the history of search, that's how it appears to me.

MC: That's what Google has promised, they've said the biggest leap forward in the history of search!

RL: There are so many questions; this is the first I've heard of it is listening to you speak about BERT. But I mean all sorts of things, you know, Google has so much data that's disparate and all over the place and that's fine, from a privacy point of view. I guess this is another whole other question to ask and a whole other conversation but if Google is intelligent to the point where you can ask a question about someone and it knows that it has information about someone, that's located in lots of different separate social media sites, news sites, could it potentially answer a question that would historically have had to have been carried out by a private investigator, for example?

MC: Yeah so we've got, so stuff again like the schema, where the objective is for Google to identify entities, whether that is a person and organization whatever the thing it is and the connections between those entities. that's one of the overriding goals I think and that's really important for them to be able to return good results to understand who is who and understand that this Mark Williams Cooke is different to this one and these two are the same and they're connected to these companies and they write on these web pages. There are certain, already legal processes, you can go through to have search results that are about you removed from Google, but you're right, it will be interesting to see what will happen as more information becomes more easily accessible, because as you say a lot of this information is accessible if you know where to look and if you've got good Google foo or if you've got money to hire a private investigator. I think that's been a that's been an issue with all sorts of things because obviously we've had updated laws because you know, ‘how to make bombs’ became very easily accessible public knowledge, you know, we got a Terrorism Act and because there's an issue, well not necessarily issue, you know I'm quite an advocate of information personally being accessible to everyone and free in that way, but I guess the size they need to navigate around that like you say that's probably a really big issue.

RL: I guess we've never been at this point in history before where we're getting to the point where we can literally ask the computer ‘give me the answer to this question’ and it will and it will give it to you, without having to do any legwork at all whatsoever and it could be used for all sorts of good things.

MC: Good things only! Let's leave it before we consider the alternative.

Thank you very much for listening we've gone a little bit over again, I hope you've enjoyed this episode. We will be back in one week's time, so that will be Monday the 4th of November. You can get all of the show notes as usual at search.withcandour.co.uk so you have all the links to all the announcements we talked about and all the really geeky machine learning, your networking, deep learning stuff that we've been talking about if you want to geek up on that. I’ll actually link as well, I found a really nice tutorial for tens of flow, which is Google's machine learning framework which actually uses transforms or transformers, sorry, in language translation they give you a tutorial on how to build your own language translation with their framework so I’ll link to that as well! Otherwise I hope you have a really great week and enjoy playing with BERT!