Search

EP-4736060-A1 - REWRITING TEXT GENERATED BY A GENERATIVE MODEL

EP4736060A1EP 4736060 A1EP4736060 A1EP 4736060A1EP-4736060-A1

Abstract

A computing system performs several acts, where the acts include providing text generated by a generative model and content of a webpage to a computer-implemented text rewriting model, where the generative model generated the text based upon user input received from a client computing device, and further where the generative model generated a citation to the webpage to indicate that the text generated by the generative model is supported by the content of the webpage. The acts also include generating, by the computer-implemented text rewriting model, a rewriting of the text, where the computer-implemented text rewriting model generates the rewriting of the text based upon: 1) the text generated by the generative model; and 2) the content of the webpage. The acts further include transmitting the rewriting of the text to the client computing device for presentment as a response to the user input received from the client computing device.

Inventors

  • POTASH, Peter

Assignees

  • Microsoft Technology Licensing, LLC

Dates

Publication Date
20260506
Application Date
20240530

Claims (15)

  1. 1. A computing system comprising: a processor; and memory storing instructions that, when executed by the processor, cause the processor to perform acts comprising: providing text generated by a generative model and content of a webpage to a computer-implemented text rewriting model, where the generative model generated the text based upon user input received from a client computing device, and further where the generative model generated a citation to the webpage to indicate that the text generated by the generative model is supported by the content of the webpage; generating, by the computer-implemented text rewriting model, a rewnting of the text, where the computer-implemented text rewriting model generates the rewriting of the text based upon: the text generated by the generative model; and the content of the webpage; and transmitting the rewriting of the text to the client computing device for presentment as a response to the user input received from the client computing device.
  2. 2. The computing system of claim 1, the acts further comprising: prior to providing the text generated by the generative model and the content of the webpage to the computer-implemented text rewnting model, providing the text and the content of the webpage to a computer-implemented classifier, where the computer-implemented classifier is trained to determine whether texts generated by the generative model are supported by contents of webpages cited by the generative model for the text; and obtaining, from the computer-implemented classifier, an indication that the text generated by the generative model is not supported by the content of the webpage, where the text and the content of the webpage are provided to the text rewriting model in response to obtaining the indication from the computer-implemented classifier.
  3. 3. The computing system of claim 2, where output of the generative model comprises the text and the citation to the webpage, the acts further comprising: as the generative model generates the output, identifying that the citation to the webpage corresponds to the text in the output; and providing the text and the content of the webpage to the computer-implemented classifier in response to identify ing that the citation to the webpage corresponds to the text in the output.
  4. 4. The computing system of at least one of claims 1-3, the acts further comprising: providing an instruction to the computer-implemented text rewriting model together with the text and the content of the webpage, where the instruction instructs the text rewriting model to ensure that the rewriting of the text is supported by the content of the webpage.
  5. 5. The computing system of at least one of claims 1-4, the acts further comprising: prior to transmitting the rewriting of the text to the client computing device, transmitting the text and the citation to the webpage to the client computing device for presentment as an initial response to the user input received from the client computing device.
  6. 6. The computing system of claim 5, wherein the text is replaced by the rewriting of the text upon the client computing device receiving the rewnting of the text.
  7. 7. The computing system of claim 5, the acts further comprising: prior to transmitting the rewriting of the text to the client computing device and subsequent to transmitting the text and the citation to the webpage to the client computing device, causing at least one of the text or the citation to the webpage to be highlighted to indicate that the text is not supported by the content of the webpage; and receiving a request from the client computing device to provide the rewriting of the text to the client computing device, wherein the rewriting of the text is transmitted to the client computing device in response to receiving the request.
  8. 8. The computing system of at least one of claims 1-7, the acts further comprising: obtaining values of logits of the generative model that correspond to the text generated by the generative model, where the computer-implemented text rewriting model generates the rewriting of the text based further upon the values of the logits.
  9. 9. The computing system of at least one of claims 1-8, the acts further comprising: obtaining values of hidden layers of the generative model that correspond to the text generated by the generative model, where the computer-implemented text rewriting model generates the rewriting of the text based further upon the values of the hidden layers.
  10. 10. The computing system of at least one of claims 1-9, wherein the computer-implemented text rewriting model generates the rewriting of the text based further upon the user input.
  11. 11. The computing system of at least one of claims 1-10, the acts further comprising: providing the rewriting of the text and the content of the webpage to the computer- implemented text rewriting model; and generating, by the computer-implemented text rewriting model, a further rewnting of the text, where the computer-implemented text rew riting model generates the further rewriting of the text based upon: the rewriting of the text generated by the computer-implemented text rewriting model; and the content of the webpage; and transmiting the further rewriting of the text to the client computing device for presentment as another response to the user input received from the client computing device.
  12. 12. A method performed by a computing system, the method comprising: receiving user input from a client computing device that is in network communication with the computing system; based upon the user input, identifying a webpage that includes content that is related to the user input; generating, by a generative model, text and a citation that pertains to the content, where the generative model generates the text and the citation based upon the user input and the content of the webpage; providing the text and the content of the webpage to a classifier, where the classifier has been trained to output indications as to whether texts generated by the generative model are supported by webpage contents provided to the generative model when generating the texts; obtaining an indication from the classifier that the text is not supported by the content of the webpage; based upon the indication from the classifier that the text is not supported by the content of the webpage, providing the text and the content of the webpage to a computer-implemented text rewriting model; generating, by the computer-implemented text rewriting model, a rewriting of the text, where the computer-implemented text rewriting model generated the rewriting of the text based upon the text generated by the generative model and the content of the webpage; and transmiting the rewriting of the text to the client computing device as a response to the user input received from the client computing device.
  13. 13. The method of claim 12, further comprising providing values of logits of the generative model to the computer-implemented text rewriting model, where the values of the logits correspond to the text generated by the generative model, and further wherein the computer- implemented text rewriting model generates the rewriting of the text based upon the values of the logits.
  14. 14. The method of at least one of claims 12-13, further comprising providing values of hidden layers of the generative model to the computer-implemented text rewriting model, where the values of the hidden layers correspond to the text generated by the generative model, and further wherein the computer-implemented text rewriting model generates the rewriting of the text based upon the values of the hidden layers.
  15. 15. The method of at least one of claims 12-14, the method further comprising: prior to providing the text and the citation to the classifier, determining that the citation pertains to the text, where the text and the content are provided to the classifier based upon the determining that the citation pertains to the text.

Description

REWRITING TEXT GENERATED BY A GENERATIVE MODEL BACKGROUND [0001] Relatively recently, generative models, such as the Generative Pre-trained Transformer 4 (GPT-4) model and the BigScience Large Open-science Open-access Multilingual (BLOOM) language model, have become available to the general public and can be readily accessed by end users who have an Internet connection. A generative model generates output (such as text, images, and/or video) based upon a prompt provided to the generative model. Conventionally, the prompt includes user input set forth to the model, previous user inputs set forth to the model during a communications session with the user, and outputs generated by the generative model during the communications session. A known issue with generative models, and specifically generative models that output text, is that the generative models can output misleading and/or factually incorrect information (where a misleading and/or factually incorrect statement by a generative model is referred to as a hallucination). In an example, a generative model receives a prompt that includes the user input “how many home runs did Babe Ruth hit before turning 30?”. Based upon such prompt, the generative model may output the text “Babe Ruth hit 189 home runs before he turned 30 years old.” This statement, however, is factually incorrect, and the user is not provided with any indication as to how the generative model created such text. [0002] To at least partially address this issue, a generative model has been configured to communicate with a search engine, generate text based upon content of a webpage identified by the search engine, and generate a citation to the webpage - therefore, the generative model provides a user with an identity of a source upon which generated text is based. The user can then check the webpage to ensure that the text output by the generative model is accurate (and not misleading). [0003] Often, users view a citation as being confirmation of the veracity of the text that corresponds to the citation. Even when, however, a generative model generates text based upon content of a webpage (where the content of the webpage is included in a prompt used by the generative model to generate the text), the generative model may still output misleading and/or factually incorrect text that is not supported by the content of the webpage that is cited as supporting the text. An end user who is provided with the text and the citation may incorrectly assume that the citation to the webpage included in the output ensures that the text generated by the generative model is supported by the content of the webpage. SUMMARY [0004] The following is a brief summary of subject matter that is described in greater detail herein. This summary is not intended to be limiting as to the scope of the claims. [0005] Described herein are various technologies pertaining to rewriting text generated by a generative model such that the rewritten text is supported by content of a webpage cited to by the generative model as supporting the text. With more specificity, a generative model generates text based upon user input received by the generative model, and the generative model further generates a citation for the text. The citation for the text identifies a source of content employed by the generative model to generate the text. For example, the citation identifies a webpage that includes the content employed by the generative model to generate the text. [0006] Output of the generative model is monitored, and a citation in the output is identified. Thereafter, text that corresponds to the citation in the output is identified. For example, a sentence prior to the citation in the output of the generative model is identified as corresponding to the citation. Identification of the citation and corresponding text can occur while the generative model continues to generate output (tokens). A prediction as to whether the text is factually accurate (and not misleading) is made based upon the text and the content used by the generative model to generate the text. For example, a classifier is trained to predict whether texts generated by the generative model are factually accurate based upon the texts and contents of webpages cited by the generative model as supporting the texts. [0007] When it is predicted that the text is not supported by the content referenced in the citation generated by the generative model, a computer-implemented text rewriting model can generate a rewriting of the text. For example, the text rewnting model receives the text generated by the generative model and the content from the webpage identified in the citation that corresponds to the text. The text rewriting model generates the rewriting of the text based upon the text and the content. The text generated by the generative model can then be replaced by the rewriting of the text generated by the text rewriting model. [0008] The technologies described resul