Oral argument — Software Rights Archive, LLC v. Facebook, Inc.

15-16-52. And again, where the issues overlap, I think that we have those as well in mind as we might. So if there's anything new, please concentrate on that. Okay, sure. So the one thing I do want to talk to, and it's related both to the 571 and the 494, is back to the box cluster splitting cluster process that are identified here. There are two rulings by the board. One of the rulings by the board is very specifically that the display one or more nodes using actual cluster links element was not met. And it's very important. This gives us a very clean way to dispose of their appeal. The board found that there was no, on page 16 and 17 of the 494 paper, the opinion of the board found that there was no disclosure of a criterion for forming a subset for display. And we believe that there is substantial evidence supporting this point of view. Now, to understand the board's theory, because there's a lot in the briefing about similarity links and subvectors, a theory they allege were not considered by the board. But the board's ruling directly implicates these as well as the theories that they did clearly set out forth in their brief. And it's very simple. Their whole theory is that the Foxmark system returns as a search result a cluster produced from cluster splitting. What you get as a search result is not just some documents. You get a cluster as a search result, and a whole cluster at that. And the idea is that if the cluster defines what the search result is, and the similarity links and subvectors used to create that cluster, also they're being used to display one or more nodes or locating one or more nodes for display. That's their whole logic line. Their entire theory depends on a whole cluster being displayed. However, Foxmark explicitly rejects that that is the case. What it shows, if I can get you to turn to JA5416, which is page 53 by the top pagination of Foxmark. And what we have is, and this is pretty complex stuff, but I think that it's just crystal clear. On page 53, what we see here is a description of the search procedure of Foxmark, and this is what they specifically rely upon. And I'm going to read. The search procedure is different than the tree-creating procedure. I don't want to read ahead of you, so you have arrived at the fact. Page 53 of Foxmark. It's JA5416. I feel like my brain is going to split. What page of Foxmark? Page 53. On page 53, it describes a search procedure, and it makes it very clear that whole clusters aren't being done. The first thing it does is this. The search procedure is shown in figure 23, and that is different than the cluster tree-creation procedure and the clustering procedure shown in figures 15 through 21. They rely on 21's cluster-splitting procedure and the tree-creation in 15 through 21. So what the search procedure does is create a new tree for purposes of providing search results, and this tree is going to be based on document-to-query similarity, not the pairwise document-to-document similarity scores that they allege constitute the actual cluster links. And here's the critical sentence. Searching a cluster tree is very much like finding a correct place to add a new document. See figure 17 in the intended description. So it's saying the search procedure is like the tree-creation procedure. However, it describes a very important difference. However, searching has a different objective. Instead of finding a single twig where insertion should follow, one would like to retrieve and rank documents so all relevant documents, regardless of what cluster they appear in, are retrieved as soon as possible. So what it explicitly states here is all relevant documents are going to be presented to the user irrespective of the cluster membership. So the whole theory, the rest of the idea that a cluster is a search result, that's not true. What this system does is it provides search results irrespective of what their cluster is, and that's explicitly contradicted. This is substantial evidence that supports the verdict or the decision in our case. And secondly, I want to point out and direct myself to the very sentence they rely upon for this idea that clusters, whole clusters, are displayed search results. If you flip the page to page 54 and go to the bottom paragraph, what we'll see is this sentence here. It's the bottom paragraph, second to last sentence. It says, first or further, most of the documents in the retrieved cluster are presented to the user fairly quickly. So what they do is they crop out the word most of in their brief, and they just say documents in a retrieved cluster are presented to the user, close quote, for the support that a cluster from cluster splitting is what the search result is. But what this sentence clearly shows, it's not the fact that a cluster is being displayed based upon cluster splitting. What it clearly shows, it's going to be displayed, if you keep on reading in the part that they did not cite, it says very quickly, though those with negligible query-to-document similarity will be bypassed until much later. Query-to-document similarity is very key here. Query-to-document similarity is not used in the tree creation process, not used in the cluster splitting process. What they rely on is a pairwise document-to-document similarity. What query-to-document similarity is a comparison of a query to a document or using search terms. It's very different than the things that created the cluster tree. It's not the same thing. And more importantly, query-to-document similarity cannot use BC or CC. And we're going to see this. If we turn to page 38 of FoxSmart, what you see here on page 38 of FoxSmart is that a query under the FoxSmart system is a user-supplied query that begins in a natural language form. What they're talking about is a natural language query on page 38. Now, I'm going to move to page 41, and it further describes the process of submitting a query on page 41 of FoxSmart. It says user submits initial query, and then in processing the query, step 2 on page 41 says it considers terms, authors, and dates. These are all semantical factors associated with a natural language query. What's missing is no BC and CC in computing document-to-query similarity. And there's a good reason why. One, a document cannot cite to a query, and a query can't cite to a document. So this similarity score inherently can't use BC and CC. And so not only do these things they point to rely upon show that a whole cluster is not being displayed, they reveal that the whole house of cards comes down that search actually relies on a method that does not consider BC or CC. It is basically a semantical-based search. And so I think that these facts are very important in both the obvious context and the anticipation context that's actually before hearing the court. These things aren't being used to search. And you see, he says it's obvious because once you understand that BC and CC can be used in similarity, that we can use that to cluster. However, it doesn't address the question. This claim is about searching ultimately with this method, not just clustering. In other words, providing search results on that. And what's ultimately shown here is no one here had the idea of using these things to actually provide search results. And there's a good reason why. Because as other papers show, BC and CC aren't particularly good. But in any case, this is substantial evidence that supports the decision with respect to display elements by the court. And one last thing. I want to address one thing she said. She read a quote where she said it shows a benefit in indirect relationships versus direct relationships. The quote actually says direct and indirect co-citations. And I don't want anyone to be confused by it. A direct co-citation is still an indirect relationship. Direct co-citations are explained on page 239 through 40 of Fox's thesis. And the one thing she pointed to that says it showed that indirect relationships would provide a benefit over direct relationships is false. Is that what that actually was referring to is co-citations direct and co-citations. And a co-citation is direct is when A and B directly co-cite each other versus having a common third node that co-cites the both of them. And that's described on page 239. And so I want to make sure that that one quote is not misunderstood. And the part about being reasonably useful in her quote was, again, versus terms, not directions. So, again, this is not substantial evidence directed to the question. And I don't want anyone to be confused by the citation that she made. Thank you. Thank you. Ms. Kief, do you need a minute to respond to what sounds to me like a new argument? Your Honor, I thought that that entire portion was over. It is over, but since it's been raised, perhaps we want to be sure that we have it straight. Absolutely, Your Honor. My response to that would simply be the document speaks for itself. The document talks about the fact that using indirect relationships yields a benefit. And if you look directly to the citation that I gave, it absolutely shows that. That's all I would have to say. Okay. Thank you very much, Your Honor. I appreciate the time. Thank you. Mr. Seibert. Thank you, Your Honor. If I may start, with the Court's permission, with the display element raised by counsel and then try to clear that up quickly, I hope, and then move to the other issues. Okay. The reason that related documents are clustered together in an information retrieval system is in order to retrieve those documents in response to a query. And the only reason, and essentially what it means to retrieve documents, or at least a list of documents in response to a query, is to display that information to the person making the query. We're all familiar with how searching works in any type of electronic library. You enter a query. You get search results that are displayed to you. That's simply how it works. It's quite elemental, and the Court doesn't need to take my word for it because Dr. Fox says in FoxSmart, clustered search allows retrieval of groups in response to query submission. And that is, and I apologize, I also have a mixed bag of record citations, but this is in the 1649 case, and the cite to that is JA10270. So with that background in mind, let me turn to what counsel said about the display step. And the first thing to say about the display step is the Board did find that the display step in Claim 1 was not satisfied. It found that solely because the display step recites, it says, using the actual cluster links. And the Board said, well, we previously found that there weren't actual cluster links that are derived from the candidates, and therefore you don't meet this step either. The Board did not independently talk about the display element as not, itself as not being present. And it is clearly taught that you're going to, when you have the cluster and you use the cluster, you're going to display using the cluster links. And I think the quotation that counsel himself used demonstrates that. It says, further, most of the documents in a retrieved cluster are presented to the user fairly quickly, though those with negligible document query similarity will be bypassed until much later. What that's saying is if I type in a query, the system will return most of the documents in the cluster to me. What he's saying is the system will do something slightly more advanced, which is if some of the documents within the cluster seem, based on some additional processing, to be unrelated to my query, what he says negligible query document similarity, those won't be shown to me until at least later in the process. So it's trying to pick the most helpful documents within the cluster to show the query, the submitter of the query first, early in the process. But the claim limitation is displaying documents using the actual cluster links. So a query is entered, the system looks at the cluster that most closely matches the query. In order to do that, it's using the cluster links that have been created, and it then does some additional processing and makes some judgments about these documents within the cluster are most related, and I'm going to show those first. When it talks about cluster links, am I supposed to be thinking of documents? What am I supposed to be thinking of? Links between. It's really the links between documents. Yeah. So in a graph, you have nodes, circles, or blocks, or whatever you want to call them, and edges, which are lines between them that are relationships between them. The cluster links, the nodes, in this case, are the actual documents. Right, so they're clusters of not of nodes, but clusters of links? The links, the way that I would put it is that the links define the cluster. So you might have a cluster, I think it's helpful to think of real world examples, you might have a cluster of documents of, you know, French poetry, documents relating to French poetry. And the system, the system doesn't think French poetry, but the system performs a programmatic analysis and concludes that these documents are closely enough related to each other that I'm going to call them a cluster. And so, and how does it decide whether or not they're closely related? It looks at a number of things. Is there overlapping language? Do they cite each other? Do they both repeatedly cite the same thing? That, which is indirect citation, that's an indication that they're related, etc. So, the process that's described in FOXSMART, and in fact, both in the tree formation process and in the cluster splitting process, which is just kind of an iterative process that continually happens to maintain the health of the cluster tree, is identical to what's disclosed as the preferred embodiment of the patent. It calculates a bunch of relations, it's got a bunch of documents, it calculates relationships that it wants to, in order to figure out how to cluster them together, it compares those, the strength of those similarity relationships, and it picks the strongest ones, and the strongest ones, it says, I'm going to cluster these together and make one cluster. I'm going to cluster these together and make another cluster. And I did, with Your Honor's permission, want to go back to the tree formation process, which was the predicate of Your Honor's earlier question about 102 versus 103, and just point out, because the board did consider that. And I want to at least explain how the board erred and where it went astray in thinking about that process. So the tree formation process, if you distinguish it from cluster splitting, works by, there's an existing tree, a new document is put into the library, and the system again says, okay, what cluster am I going to put this in? And so in that case, what the system does is again, for that particular document, which would be a selected node for analysis, it calculates a similarity value between that document and every single other document in the system, and then it performs further processing in order to determine which cluster to add that document to. So what are called sub-vectors, that essentially rows in a table, columns in a table, that contain information. The claim is for analyzing data in a database and selecting a node for analysis. I mean, I guess to me, that reads more like everything is already pre-existing in the database, and now we're going to choose one of those documents and try to figure out the relationships of all the other documents in the database to that particular selected one. So I don't quite read it as we're going to try to shove a brand new document into a pre-existing database with a bunch of other documents already in there. Well, but the document is in the database when it's analyzed. I mean, I don't know that FOX necessarily would make that distinction at all or recognize it as significant. However you start, you could start with a tree. The more normal use case, if you're maintaining an electronic library, is you already have something, and new things are coming into the collection constantly. But you could start with a bunch of documents and say, I want to form clusters out of them. It's the same process. The process is you calculate the similarity value between the document and the other documents. And incidentally, I do want to say that the document is in the database by the time you're doing that. It gets added, and then it's there, and then you've got to figure out where in the tree you want to insert it. And that's where the tree formation process begins, right? Well, I guess the way that I was describing it, in that particular example, the tree exists. There is a tree. There is a Lexis. So you take Lexis. So clustering is used for things like more like this. I like this case. Show me more cases that are like this case. So it will return documents that it's determined are similar to the case that you've selected somehow that are essentially in a cluster. So Lexis exists. The Lexis database exists. But obviously new cases are constantly being added to the database, and so those get ingested by Lexis. I'm simply using Lexis as an example, but those are ingested by Lexis, and then the system does some processing and says, what cluster or clusters am I going to put this in? And so it then calculates a similarity value between that new case that it's analyzing and all the other cases. I don't know specifically whether Lexis does all the other cases, but again, this is just an example. And then it compares the values, and it says, where in what cluster am I going to put this? That is precisely what the patent describes as the preferred embodiment of this generating candidate links and then deriving actual links. And where the board went astray, and I have to say in defense of the board, these proceedings were quite unwieldy below. There were something like over 8,000 pages submitted only of briefing expert declarations and deposition transcripts, 8,000 to 9,000 pages, and then thousands of pages of exhibits on top of that. And the board certainly got most of it right, but where they went astray on this particular issue was they said, well, we're looking at the fully formed tree, the already existing fully formed tree, as we're going to say that's the candidate. Those are the candidate links. And then they kind of said, well, there's nothing that happens after that that illuminates things from the tree that would get to the actual links. And so therefore, you haven't demonstrated or Fox didn't teach the deriving step of the claims. Where that analysis went astray is that they should have gone, and this is the argument that we made, they should have gone one step back in the workflow for the candidates. The candidate links are the system looks at, takes, ingests a new document, looks at all the other documents, and computes similarity values based on subvectors, BC, CC, potentially other ones, to try to figure out how related is this, how similar is this document to these other documents. And it's doing it, let's be clear, for the purpose of clustering. It's saying, I want to figure out what group to put this document in. I'm first going to figure out how similar is it to all these other documents. Those are the candidates. And then it says, okay, I'm going to put this in the French poetry cluster. And so it says, that's where I'm going to put it. Those become the actual links. So to say that the fully formed tree are the candidates, yes, that probably does lead to a dead end. But that's not what Fox taught and what we argued. What Fox taught is you look at, it's intuitive, I believe. I mean, you have a new document, you calculate the similarity, you compare the similarity that you've calculated among the various documents, and you pick the stronger ones in order to cluster. And I guess the last thing I'd say about this, again, is I do believe in thinking about the cross appeal, it's important to take a step back and look at the sole element, this element that was found to be lacking. And again, one would read the patents, the Egger patents, and believe that there are plenty of things that Mr. Egger thought were inventive and novel and innovative about those patents. He turns out not to have been correct, but he presumably thought they were, and specifically the idea of using indirect citation relationships to calculate similarity and to improve search. This particular idea, that's the deriving step that the board found not to be taught, is not something that one would ever read the patent specifications and think was supposed to be an inventive or novel idea or step. In fact, the specifications say almost the exact opposite. They say almost nothing about the deriving step other than they say, they call it a simple matter of selecting or choosing the top rated links and eliminating the rest. Do you recall, as in the other case, whether on the cross appeal, whether the petition was limited to 102? I believe, Your Honor, that we did petition on 102 grounds, and I believe we did it on the basis that we thought it was clearly there. We know it was granted on 102, but there was no reference to 103. I don't believe – we may have, and this is something I need to go back to the petition to check. We may have said something along the lines of we think this is here explicitly. To the extent it's not, it would be obvious, but I believe that the ground of the petition was 102. I would say – I mean, I think for all the reasons I've said, I believe that this step, find some candidates, pick the strongest candidates, and use those as the actual ones, is explicitly taught in the Fox papers, both in the cluster splitting and reformation process, and also in – which the board, I think everyone agrees, did not engage with, and also in the tree formation process. Which the board did consider, but in our view, got wrong. Okay, thank you. We have a few seconds left for rebuttal. Your Honor, the first thing I want to leave with is discussing Mr. Silbert's comments about what the board found with respect to the display element. He's characterized this as being the same ruling that's found with deriving the actual cluster links that's talked about preceding. These are – that's completely false. These are two separate and independent bases of why Claim 1 was not anticipated. What he specifically found is that the display element was not met, and this is why. But he's talking about claims of the breadth for which you're arguing, not claims limited to a lot of the limitations you've been explaining to us this morning. I am sorry, I don't understand your question, Your Honor. I'm trying to put this in the context of the claims themselves. You've been explaining to us a lot of limitations, a lot of background, a lot of differences that are not in the claims themselves. Oh, Your Honor, what I'm specifically talking about with respect to the display element is this element. Displaying one or more nodes using actual cluster links. It's the last element of Claim 1 of the 494 patent that we were discussing. And specifically, he characterized the board's decision as not providing an independent basis beyond the discussion of deriving a subset in the deriving element. With the display element, there's a separate basis that he has not addressed anywhere in his papers, nor is it addressed up here. And he's incorrect when he says there's not an independent basis. And I'd like to direct the court to page 16 of the final written decision for the 494 patent. Okay. So at the bottom of page 16, I'm going to begin reading and explaining what the board actually found here. I'm going to begin with the last paragraph on page 16. We are not persuaded that FoxSmart's description of ranking documents discloses deriving a subset because a set of ranked documents provides an indication of an order of presentation but is not a subset. It goes on to say, additionally, FoxSmart indicates documents from multiple clusters are ranked. So what it's saying here right now is that the FoxSmart system only ranks the documents. It never has a cutoff to display a subset of less than the whole database. The entire database, the entire tree, everything is ultimately provided as a search result in ranked order based upon document to query similarity. There's no attempt to use the cluster as defining a specific subset in the database to display. In this statement here, FoxSmart indicates documents from multiple clusters are ranked. There's a direct reputation that the search results are single clusters. What's being presented is all documents in the database in ranked order. And it goes on to say, furthermore, petitioner does not point to disclosure in FoxSmart of criteria for forming a subset. In other words, if you look at the ranking procedure on 5416 that we just looked through, that search procedure, it doesn't attempt to display a cluster. It ranks all documents in the entire tree irrespective of cluster and spits out a ranked order of all the results. Never does it find a subset. And because there's not a subset of nodes ultimately being located for display, but rather all nodes are being displayed, there's not an actual cluster link that's subjected to criterion used to locate nodes. None of the things they point to actually influence whether you're going to be displayed. You're going to be displayed, period, regardless by your presence in Fox's database. And it has nothing to do with what cluster you are in or what your similarity link is or any of these things. This is what the board has found right here. And they have not addressed this. They have not identified any criterion that would define a subset of nodes for display. And since nodes aren't being located for display, they're all being presented, you don't have an actual cluster link or display using the actual cluster link's element. So I think that's very important. I think I also want to point out, he made this assertion that clustering's only use is for providing search results. Well, the primary was searching and clustering are distinct aspects of information retrieval. Clustering is used for automated classification, in ways to be able to take a database and cluster or organize it in a way without having humans to look at it. Okay, that's one aspect of IR. That doesn't necessarily mean it's going to be used for providing search results. If you look on page 68 of the 571 reply brief, we've depicted an output of the FoxSmart system. And what it shows is all documents are being displayed. And that there are only four clusters, period, among the 52 documents of this alleged database being used. Let's think about that for a second. Really, there's only four responses that you can have no matter what your search query is? That doesn't make any sense. The idea that these clusters are going to be presented that way, there's only four possible responses. In other words, they fundamentally misinterpreted this idea that you're going to compare the clusters, have the highest cluster to query similarity is going to be displayed. It doesn't say that. In fact, what it says is document to query similarity. Not cluster to query similarity, document to query similarity in determining what's displayed. But in any event, since all documents are going to be displayed, regardless of what cluster they are, regardless of their BC and CC links, or any of these things, this is not a search system that is using indirect relationships to provide search results, or the clusters, or the similarity links, or the subvectors to provide search results. All things are displayed. And again, I also remind you for page 68. We need to wrap it up with one final statement. Yes, Your Honor. So I believe that this constitutes substantial evidence to affirm the portions of the board's decision that supported us. And again, I believe the defendants have not provided any substantial evidence that shows that indirect relationships can provide a benefit over direct relationships. Consequently, those very strong and powerful teachings would discourage one from practicing the combination as a whole and the decisions you rendered in favor of us. Thank you, Your Honor. Okay. Thank you, Mr. Howard. Mr. Shilbert, a couple of minutes. Thank you very much, Your Honor. I'll try to take less of that. Concentrate on rebuttal, please. Thank you. Just on the displaying step, the portion of the record that counsel read says, we are not persuaded that Foxsmart's description of ranking documents discloses deriving a subset because the set of ranked documents provides an indication and it goes on. So the deriving a subset is part of the claim construction that the board applied to deriving actual cluster links from candidate links. So in other words, they said the actual links need to be a subset of the candidate links, and they found that that wasn't taught incorrectly, but that they found that that wasn't taught in the Fox papers. The point they're making about the display is there is no subset of actual links that's used to create a display. It's, forgive the pun or phrasing, but it's derivative of their conclusion on the deriving step. This was the point we made. Once the court concludes, and we believe it should conclude, that the deriving step, deriving actuals from candidates was taught, then the display step is also taught. But again, I don't think the court needs to look further than Dr. Fox's statement that clustered searching allows retrieval of groups of documents in response to a query. It's quite basic. When related documents are clustered together in an information retrieval system, a query returns or can return a group, the cluster of documents. In sum, we don't believe that the deriving step or the display step, which are both within the context of these patents and even outside that context, quite routine elemental steps that are there in the claim but are simply being recited, should hold up the finding of invalidity on these particular claims if all the other limitations are met, which we believe that they are. Thank you. Thanks to our counsel. The case was well presented, complicated issues. The three cases are taken under submission. Thank you, Your Honor. Thank you, Your Honor.