Showcase · Patent research

PatentTalk and Patent Universe

PatentTalk follows an invention interview while it happens. Every time the description of the invention changes, it searches the patent record again, so prior art keeps pace with the disclosure.

For a patent practitioner, that means prior art that reflects the invention as the inventor finally describes it, not only as it was first written down.

How it works

The search follows the interview

Every turn of the interview revises a single description of the invention. Each revision is encoded and searched against the full abstracts index, so the references keep pace as the disclosure changes.

  1. Disclosure and turnsAn uploaded disclosure, then spoken turns, corrections and answers
  2. Invention stateRevised on every turn, with the cause of each revision logged
  3. EncoderPatentSBERTa, 768 dimensions
  4. Abstracts index7,916,854 patent abstracts
  5. Prior artTop 10 references, ranked again for each revision
Patent abstracts searched on every revision
7,916,854
Dimensions in each PatentSBERTa embedding
768
Search loop time at the 95th percentile
40.4 ms
Measured over 20 turns in a separate run
Patents placed on the Patent Universe map
5,045,073
59% of the 8,563,503 patents in the source data

Interview console

Replay a recorded session

This session is built from the text of US Patent 10,814,844, Printable representations for time-based media. Read it left to right: the left column is the interview, with the disclosure, what the inventor said and a correction; the right column is the top ten prior-art references after each change. Choose a revision or press Play.

Session

The uploaded disclosure seeds the invention state.

  1. Disclosure uploadedDocument

    The present invention relates to systems and methods for generating printable representations for time-based media. 2. Description of the Background Art Conventional printers are currently used to generate documents of various different formats and based upon different types of content. However, while conventional printers can produce images of text and/or pictures, conventional printers are limited in their ability to effectively generate representations of multimedia content. Conventional printers print onto a fixed medium, such as paper, and thus they are unable to effectively capture the elements of time-based media. Yet, the capability to easily review time-based media content is commonly needed today. To search for desired features within time-based media content currently, one must actually review the content itself, skimming to find the desired information. For example, a user may have to manually skim an audio recording of a radio talk show to find content on a particular topic or to find discussions by a particular speaker. Due to these limitations in conventional printers, there is currently no easy way for users to search through a lengthy media segment to identify and extract particular features of interest from the media content. Additionally, there is no way for users to create an easily readable representation of media that provides useful information about the media. Moreover, media content is commonly only available in digital form. However, for many users, a digital format is not the optimal format in which to view information. While viewing media information in digital form is adequate for some users, many users find it easier to comprehend and assimilate information when the information is printed on a paper medium. Nonetheless, there is not currently available a mechanism for generating a paper-based representation of time-based media through which the user can review or even access media content. Therefore, what is needed is a system and methods for generating a representation of time-based media that can be paper-based and can provide users with the ability to extract defined features in the multimedia content

Prior art, revision 1

Top 10 by cosine similarity to the current invention state.

Changes
10 retrieved
Top similarity
0.8290
Similarity to previous state
First state
Turn time
3.15 s

Similarity is cosine similarity between 0 and 1. Higher means closer in meaning; it is not a legal finding.

The first search of a session is slower than the turns that follow.

  1. 1

    Paper-based interface for multimedia information

    Record 10001895

    0.8290
  2. 2

    System and method for the creation and access of dynamic course content

    Record 12191459

    0.7909
  3. 3

    Record 12831902

    0.7844
  4. 4

    Record 15690173

    On the map

    0.7840
  5. 5

    Record 12655498

    0.7826
  6. 6

    Print media information systems and methods

    Record 10831051

    0.7813
  7. 7

    Record 11682876

    0.7804
  8. 8

    Document processing apparatus and method

    Record 11210733

    0.7790
  9. 9

    Electronic programming guide system and method

    Record 10348591

    0.7778
  10. 10

    System and method for data publication through web pages

    Record 9714207

    0.7776

Patent Universe

Where the results sit in the corpus

A three-dimensional layout of 24,000 patents sampled from the 5,045,073 that have a position, colored by how densely each region is populated. Across the session's four revisions, 8 of the 40 retrieved references have a position on the map, and they light up as you step through.

Loading the map

In revision 1, one retrieved reference has a position on the map, and the gold centroid marker sits on it. It shows where the results sit, not where the invention sits.

  • Densest regions of the corpus
  • Moderately dense
  • Sparse
  • Placed reference from this revision
  • Centroid marker

Drag to rotate. Use the buttons, or Ctrl and scroll, to zoom.

Patent text: United States Patent and Trademark Office patent data. Embeddings: the PatentSBERTa model.

Reading this showcase

What the views show

  • The replay is a recorded session, not a live connection to the system.
  • The session is built from the text of granted patents. The mistranscribed turn is staged to show how a correction changes the results.
  • No attorney has evaluated the retrieved references. The panels show what the system returned, not a judgment that the results are right.
  • The map places patents that have a position in the layout. The marker summarizes where retrieved references sit; it is not a location for the invention.

Patent practitioners

Tell us where it falls short.

We are looking for patent attorneys and agents willing to review PatentTalk and tell us what it would need to be useful in practice.

Contact us Next showcase: Islamic law