Lesson 29: Large Language Models and Semantic Algorithms of Google
Key Takeaways
Core Concepts
Main Teaching
hello welcome to the 25th lecture of the semantic sr course in this lecture we will just examine some of the language models that actually google published and we will try to understand their angle and in the future you will need to do these type of researches by yourself too because we are a kinds of in the in the age of the large language models in the future lectures also in the future we will see actually how to use llms for printing websites or what kinds of different methodologies we can use for fine-tuning them and improve the results further and further so basically the first language ...
Koray's Terminology
| Term | Meaning in Context |
|---|---|
| Topical Authority | State of ranking higher than authority websites through semantically organized content networks with lower cost of retrieval |
| Topical Map | Content network design based on semantics with core and outer sections, processing a central entity with main and minor attributes |
| Semantic Content Network | Collection of connected, semantically optimized web documents organized for comprehensive topical coverage |
| Central Entity | The entity appearing in every subsection of the semantic content network, both in main and supplementary content |
| Source Context | The purpose of the brand identity and how the brand monetizes its content |
| Knowledge Domain | A specific area of knowledge (e.g., health, finance) that determines entity relationships |
| Discourse Integration | Connecting sentences and sections so context flows naturally from one to the next |
| Representative Question | The canonical or primary form of a query that search engines rank as the main question |
Practical Application
- Create a content brief using Koray's contextual vector methodology
- Map out the central entity and source context for your project
- Study the concepts presented in this lesson until they become intuitive
- Review the related case studies mentioned by Koray for real-world application
- Practice identifying the key terminology in your own SEO projects
- Apply the frameworks discussed to a test website or content network
Connection to Framework
Full Transcript
hello welcome to the 25th lecture of the semantic sr course in this lecture we will just examine some of the language models that actually google published and we will try to understand their angle and in the future you will need to do these type of researches by yourself too because we are a kinds of in the in the age of the large language models in the future lectures also in the future we will see actually how to use llms for printing websites or what kinds of different methodologies we can use for fine-tuning them and improve the results further and further so basically the first language model here that we will be showing it is for it is actually pegasus and it is for actually summarizing an article when it comes to the summarization of an article actually the the thing that google usually uses is the key points and the key points are coming from the facts not from the opinions or the relative probabilities the same thing also happens inside the google documents in the google documents there is always a kind of a problem with the fact that the fact that the fact that the fact that the fact that the fact that the fact that the fact that the kind of summary creator or summarization engine and that somebody created also uses the same methodology and the mindset too as you look at here the pegasus actually gets four sentence first one says pegasus is mythical the other one tells pure white other than other one tell it is the model name other one tells the cool name as you see the model actually doesn't take the sentence of pure white or cool name because these are not facts these are just relative opinions or they're statements when it comes to the definitional articles or definitional questions as inside in the question types always go with the as much as possible always go with the actual absolute facts or try to create a consensus at least in your document do not have conflicting declarations even a single adjective or a single adverb that are used in a destructive way can actually eliminate that answer according to the search engine and in the future we will be ranking our answers not just for the blue title links or for feature snippets basically we will be ranked inside a kinds of chat window and once it happens you will realize that these type of single or simple word choices might affect your overall rankability another thing here that is actually augmenting the specific let's say the concepts even if it is not told there for example the sentence of we paid 20 at the buckingham palace gift shop here as you see we don't have the word actually the word pound but search engine guessed that buckingham palace is located in the united kingdom and official currency of the united kingdom is pounds so this should be the pounds the thing here is that these type of algorithms are expensive so that's why give everything explicitly as i stated in in our rule set give definitive answers by mentioning every attribute do not just tell for instance putin tell that the the president of the russian federation or 56 they don't have 56 i guess but whatever just try to give the number president or country name or others use the long form answers by giving every kinds of details there and here from bill slavsky i'll be showing you one thing because in this part you will see that this is already inside the patterns the pound is the united kingdom currency so this is actually how i do seo for the last five years and i'm going to show you how i do seo for the last five years and i'm going to using topical authority actually overall already four years the thing here is that if google publishes a research paper and if they also publish an announcement then if they also get a patent for that it means that actually they are using that technology in a really really for instance this is patent retrieval augmented language model pre-training and the fine tuning these are the inventors here and as you see it is pretty pretty modern but if we come to the here you will see that actually it's same and some of these names actually even might overlap with each other you can check them for instance i guess yeah this one here but there is one letter missing here but i assume they're the same person so in this case if they if they get patent if they publish research paper it means that they are using it especially for the knowledge domain or let's say knowledge base construction for different types of knowledge domains and once you get these patent once you read the patent you can see that the patent is the same as the research paper so if you read it once you find these research papers then you can actually extract different types of results from these then you can understand search engines or their classifications for instance we have a document here another document and another document here and basically here the search engine tries to choose the most relevant document based on nearest let's say nearest neighborhood and they are extracting this information from this document to fill the gap so they don't use this one and this is already similar to the response generation that we will be seeing soon with the google bar we already solved that and probably if it happened which will be happening there will be new lectures for that too and the next one is conversational search experience so conversational section here is important it is published in 2022 but still conversational search is important because this is what we do with the chat gbt as well we search while giving some types of conversations so this is important because as you remember in previous lectures we talk about query path right so check here we search for turkey recipe we see the recipes then we search for carving and then we get definition of the carving but now we search for turkey recipes then we search for the carving and since there is a query path from recipe to the carving with turkey carving the turkey appears here if you do the same like this the polar express for instance and if you search for a christmas story for example they already know what they should be actually suggesting us inside this specific carousel so that's why i am telling that always give the query meaning from the user behaviors as much as possible and in this case you will need to guess all these connections and associations and you will need to continue from there you already know the lambda i will keep it simple and probably when i share some of other presentations private presentations and the lectures you will get them already but basically you should understand that every question has been connected to another question as language agnostic so basically this is what we do actually inside our content briefs too we generate a question match it in an answer then move on to the next question and usually the questions inside the headings they are representative questions and supporting questions and the supporting evidences they they come after the representative main question you already know the mom and in this part i will be talking about a little bit actually the topical map creation so let's say we talk about actually let's say hiking to the mountain fuji so in this case there are a few things for instance if you create a topical map as a tracking route consultancy company let's say in this case you will need to get all the tracking routes all the mountains for tracking and all the necessities and the guides for the tracking if you talk about just mountain fuji your topical map will be the tracking route for the mountain fuji and the activities that can be done in mountain fuji and it will go beyond just a tracking if we actually talk about overall let's say the shoes or the tracking shoes in this case our topical map one time one more time it will be changing because we will have different types of products and we will need to focus on actually product reviews a little bit further basically according to your central entity and the...