Rewriting text generated by generative model

Abstract

A computing system performs a number of actions, where the actions include providing text generated by a generative model and content of a web page to a computer-implemented text rewriting model, where the generative model generates the text based on user input received from a client computing device, and wherein the generative model generates a reference to the web page to indicate that the text generated by the generative model is supported by the content of the web page. The actions further include generating a rewrite of the text by the computer-implemented text rewrite model, wherein the computer-implemented text rewrite model generates the rewrite of the text based on: 1) the text generated by the generative model; and 2) content of the webpage. The actions also include transmitting a rewrite of the text to the client computing device for presentation as a response to a user input received from the client computing device.

Description

Background Art

[0001] Recently, generative models such as the Generative Pre-Trained Transformer (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 easily accessed by end users with an internet connection. Generative models generate outputs (such as text, images, and / or videos) based on prompts provided to the generative model. Conventionally, prompts include user input presented to the model, previous user input presented to the model during a communication session with the user, and output generated by the generative model during the communication session. A known problem with generative models, particularly generative models that output text, is that the generative model can output misleading and / or factually incorrect information (wherein misleading and / or factually incorrect statements of the generative model are referred to as hallucinations). In an example, the generative model receives a prompt including the user input "How many home runs did Babe Ruth hit before he was 30?" Based on this prompt, the generative model can output the text "Babe Ruth hit 189 home runs before he was 30." However, this statement is factually incorrect, and no indication is provided to the user as to how the generative model created this text.

[0002] To at least partially address this issue, the generative model is configured to communicate with the search engine, generate text based on the content of a web page identified by the search engine, and generate a reference to the web page—thus, the generative model provides the user with an identification of the source on which the generated text is based. The user can then review the web page to ensure that the text output by the generative model is accurate (and not misleading).

[0003] Typically, users view citations as confirming the authenticity of the text corresponding to the citation. However, even when a generative model generates text based on the content of a web page (where the content of the web page is included in the prompts the generative model uses 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 web page cited as supporting the text. An end user provided with text and a citation may incorrectly assume that the reference to the web page included in the output ensures that the text generated by the generative model is supported by the content of the web page. Summary of the Invention

[0004] The following is a brief summary of the subject matter that is described in greater detail herein. This summary is not intended to limit the scope of the claims.

[0005] This document describes various techniques for rewriting text generated by a generative model so that the rewritten text is supported by the content of a web page referenced by the generative model as supporting the text. More specifically, the generative model generates text based on user input received by the generative model, and the generative model also generates a reference to the text. The reference to the text identifies the source of the content that the generative model used to generate the text. For example, the reference identifies the web page that includes the content that the generative model used to generate the text.

[0006] The output of the generative model is monitored and references in the output are identified. Thereafter, the text corresponding to the reference in the output is identified. For example, the sentence preceding the reference in the output of the generative model is identified as corresponding to the reference. As the generative model continues to generate output (tokens), identification of the reference and the corresponding text may occur. A prediction is made about whether the text is in fact accurate (and not misleading) based on the text and the content of the web page that the generative model uses to generate the text. For example, a classifier is trained to predict whether the text generated by the generative model is in fact accurate based on the text and the content of the web page that the generative model uses as supporting text.

[0007] When the content referenced in the reference generated by the generative model is predicted to be unsupported by the text, the computer-implemented text rewriting model can generate a rewrite of the text. For example, the text rewriting model receives the text generated by the generative model and content from a webpage identified in the reference that corresponds to the text. The text rewriting model generates a rewrite of the text based on the text and content. The rewrite of the text generated by the text rewriting model can then replace the text generated by the generative model.

[0008] The described techniques result in an improvement over conventional generative models because the output provided to a user when employing the techniques described herein includes fewer misleading and / or factually incorrect statements than the output provided to a user by conventional generative models. Specifically, text that is not supported by the content of a web page identified as supporting text by the generative model is rewritten to ensure that the content referenced by the generative model supports the rewriting of the text.

[0009] The above summary presents a simplified summary of the invention in order to provide a basic understanding of some aspects of the systems and / or methods discussed herein. This summary is not an extensive overview of the systems and / or methods discussed herein. It is not intended to identify key / critical elements or to delineate the scope of such systems and / or methods. Its sole purpose is to present some concepts in a simplified form as a prelude to the detailed embodiments that will be presented later. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 is a functional block diagram of a computing system configured to rewrite text generated by a generative model.

[0011] Figure 2 is a communication diagram depicting communications between devices and computer-implemented modules related to rewriting text generated by a generative model.

[0012] Figures 3 to 7 A graphical user interface (GUI) corresponding to the generative model is depicted.

[0013] Figure 8 is a flow chart illustrating a method for rewriting text generated by a generative model.

[0014] Figure 9 is a flow chart illustrating a method for rewriting text generated by a generative model.

[0015] Figure 10 It is a functional block diagram of a computing system. DETAILED DESCRIPTION

[0016] Various techniques related to rewriting text generated by a generative model and thereby correcting hallucinations output by the generative model will now be described with reference to the accompanying drawings, wherein like reference numerals are used throughout to refer to like elements in the drawings. In the description that follows, for purposes of explanation, a number of specific details are set forth in order to provide a thorough understanding of one or more aspects. However, it is apparent that such aspect(s) may be practiced without these specific details. In other instances, well-known structures and devices are shown in block diagram form to facilitate describing one or more aspects. Further, it will be understood that functions described as being performed by certain system components may be performed by multiple components. Similarly, for example, a component may be configured to perform functions described as being performed by multiple components.

[0017] Furthermore, the term "or" is intended to mean an inclusive "or" rather than an exclusive "or." That is, unless specified otherwise, or clear from the context, the phrase "X employs A or B" is intended to mean any of the natural inclusive permutations. That is, the phrase "X employs A or B" is satisfied by any of the following instances: X employs A; X employs B; or X employs both A and B. Furthermore, the articles "a" and "an" used in this application and the appended claims should generally be construed to mean "one or more" unless specified otherwise or clear from the context to direct to a singular form.

[0018] Further, as used herein, the terms "component," "system," and "module" are intended to encompass a computer-readable data storage device configured with computer-executable instructions that, when executed by a processor, cause certain functions to be performed. Computer-executable instructions may include routines, functions, and the like. It should also be understood that a component or system may be located on a single device or distributed across multiple devices. Further, as used herein, the term "exemplary" is intended to mean serving as an illustration or example of something and is not intended to indicate a preference.

[0019] Described herein are various features related to rewriting text generated by a generative model and correcting hallucinations output by the generative model. The generative model communicates with a search engine so that the search engine can identify a web page that includes content relevant to the user input, and the content can be included in a prompt provided to the generative model. The generative model generates text based on the content of the web page. Further, the generative model can generate a reference to the web page and associate the reference with the text, thereby indicating to the user that the text generated by the generative model is supported by the content of the web page. However, it has been observed that the generative model can generate text and associate a reference to the web page with the text, but the web page does not support the text generated by the generative model. For example, the text generated by the generative model can include misleading and / or factually incorrect statements that are not supported by the content of the web page.

[0020] Can detect the situation that generation model produces hallucination by using classifier, wherein classifier is provided with the text that generation model generates and the content that generation model quotes as supporting text.Based on the text that generation model generates and above-mentioned content, classifier can predict whether this content supports this text.When the output of classifier indicates that the text that generation model generates is not supported by the content that generation model quotes, the text rewriting model that computer realizes can generate the rewriting of text.For example, the text rewriting model generates the rewriting of text based on the text that generation model generates and the content that generation model quotes.Correspondingly, when generation model outputs hallucination and quotes webpage to support this output, hallucination can be detected and text can be rewritten so that the text that rewrites is supported by webpage.

[0021] Now refer to Figure 1 , a functional block diagram of computing system 100 is configured to: 1) detect hallucinations in text generated by a generative model; and 2) generate a rewrite presenting such text. Computing system 100 is in communication with a client computing device 102 operated by a user. Further, computing system 100 is capable of retrieving web pages 104 to 106 from a computer-readable storage location via a network connection.

[0022] The computing system 100 includes a processor 108 and a memory 110, wherein the memory 110 includes instructions executed by the processor 108. The computing system 100 also includes a data store 112. The data store 112 includes a web index 114, instant answers 116, and a knowledge graph 118. The web index 114 can be used to identify web pages that include content relevant to a query. Example answers include answers to fact-based queries such as "What is the population of the United States?", "What is the distance between the Earth and the Moon?", etc. The knowledge graph 118 is searchable to identify entities and information about entities.

[0023] The memory 110 includes a generative model 120 that is configured to receive user input and generate output based on the user input. In an example, the generative model 120 is a transformer-based model, such as a GPT-4 model, a BLOOM model, etc. The memory 110 also includes a search engine 122 that is configured to search the web index 114, instant answers 116, and / or the knowledge graph 118 based on user input received from the client computing device 102 and / or a query generated by the generative model 120. Accordingly, the search engine 122 identifies (multiple) web pages, (multiple) instant answers, entity information, etc. that are related to the user input received by the computing system 100 from the client computing device 102.

[0024] The memory 110 also includes a content extractor module 124 that extracts content from a source (such as a web page) identified by the search engine 122. For example, the content extractor module 124 can extract content from a web page identified by the search engine 122 as relevant to the user input. The extracted content can be included in a prompt provided to the generative model 120 as an input to the generative model 120. In an example, the search engine identifies the first web page 104 as including content relevant to the user input, the content extractor module 124 extracts web page content 126 from the first web page 104, and the content extractor module 124 provides the web page content 126 to the generative model 120 (as part of the prompt, where the prompt may also include the user input received from the client computing device 102).

[0025] The generative model 120 generates output based on the prompt. The output generated by the generative model 120 may include text, images, videos, etc. Further, the output generated by the generative model 120 may include a reference to the first web page 104 from which the content extractor module 124 extracted web page content 126, wherein the reference is associated with the text generated by the generative model 120. The generative model 120 generates the reference based on the prompt provided to the generative model 120. Accordingly, a user reviewing the output of the generative model 120 can determine that the text generated by the generative model 120 is supported by the content of the first web page 104 due to the reference to the first web page 104 associated with the text.

[0026] The memory 110 also includes a text identifier module 128 that searches for references in the output of the generative model 120 and identifies text associated with the reference. For example, when a reference occurs at the end of a sentence, the text identifier module 128 may identify the sentence preceding the reference as text associated with the reference. In another example, the reference may be in the form of "according to XXXX." The text identifier module 128 may employ a rule that searches for such language (e.g., "according to") and may identify text following such language as text associated with the reference.

[0027] The memory 110 also includes a classifier 130, which is configured to: 1) receive the text and content used by the generation model 120 to generate the text; and 2) output a prediction of whether the content supports the text. Therefore, the classifier 130 can receive the text generated by the generation model 120 and the content of the web page cited by the generation model 120 as supporting the text. The classifier 130 can output a prediction about whether the content supports the text. In an example, the classifier 130 is a binary classifier that outputs a value of 0 when the content does not support the text, and outputs a value of 1 when the content supports the text. In another example, when predicting whether the content supports the text, the classifier 130 outputs a probability value indicating the confidence of the classifier 130.

[0028] With reference to determining whether the text generated by the generation model 120 is supported by the content, when the text includes facts or statements that are also expressed in the content, the text is supported by the content. For example, when the text includes an incorrect statement of a fact that is not included in the content, the text generated by the generation model 120 is not supported by the content. In another example, when the text includes a statement that contradicts a statement included in the content, the text generated by the generation model 120 is not supported by the content.

[0029] Memory 110 also optionally includes a sorter module 132 that sorts different content included in the prompt with respect to the text generated by generative model 120, where the content is extracted from different sources. It has been observed that when the text generated by generative model 120 is in fact supported by different sources, generative model 120 generates text and then generates a reference to the source (e.g., a webpage) as support for such text. Therefore, when classifier 130 predicts that the text generated by generative model 120 is not supported by the content corresponding to the reference to the text, sorter module 132 can be combined with determining whether generative model 120 should reference different sources as support for the text to sort the content provided to generative model 120 in the prompt with respect to the text. In the example, sorter module 132 treats the text generated by generative model 120 as a query and treats the content extracted from the source by content extractor module 124 as potential search results. Sorter module 132 therefore calculates a ranking score for each content. When the first score for the referenced content by the reference is lower than the second score for the second content, the ranker module 132 may provide the second content and the text generated by the generative model 120 to the classifier 130. The classifier 130 may then predict whether the second content supports the text generated by the generative model 120. When the classifier 130 predicts that the text generated by the generative model 120 is supported by the second content, the reference to the content may be replaced with a reference to the second content.

[0030] The memory 110 also includes a text rewriting model 134 that rewrites text generated by the generation model 120, wherein the text has been identified by the classifier 130 as not supported by a source identified in a reference generated by the generation model 120, and wherein the reference is also associated with the text. More specifically, the text rewriting model 134 can be provided with text generated by the generation model 120 and content extracted from a source (e.g., a web page, an instant answer, an entity card, etc.) identified in a reference to the text (where the reference is generated by the generation model 120). The text rewriting model 134 is trained to rewrite the text so that the content provided to the text rewriting model 134 supports the rewriting of the text. For example, the text rewriting model 134 is trained based on text pairs, wherein each pair includes: 1) text generated by the generation model 120; and 2) content corresponding to a reference to the text. Each pair can be marked to indicate whether the text in the pair is supported by the content in the pair. Thus, when the text rewriting model 134 is provided with text and content, the text rewriting model 134 can rewrite the text to ensure that such text is supported by the content. The text rewriting model 134 can cause the rewriting of the text to be transmitted to the client computing device 102 for presentation in response to user input received from the client computing device 102.

[0031] Now refer to Figure 2The operation of computing system 100 is described below. Figure 2 2 is a communication diagram 200 depicting the communication occurring between the client computing device 102, the generative model 120, the search engine 122, the content extractor module 124, the text identifier module 128, the classifier 130, and the text rewriting model 134. At 202, the generative model 120 receives user input from the client computing device 102. The user input may include text, queries, statements, etc. formulated in a conversational manner. Further, the client computing device 102 may receive user input such as voice input, text input via a keyboard, etc.

[0032] Upon receiving the user input, the generation model 120 may optionally generate a query based on the user input. For example, the user input may be relatively long and / or complex, and the generation model 120 may generate a query based on the user input that is well-suited for providing to the search engine 122. At 204, the generation model 120 provides the user input and / or the query generated by the generation model 120 to the search engine 122. The search engine 122 searches the web index 114, the instant answers 116, and the knowledge graph 118 based on the user input and / or the query. In the example, the search engine 122 identifies the first web page 104 as including content related to the user input and / or the query. The search engine 122 may retrieve the first web page 104, including markup language associated with the first web page 104, plain text included in the first web page 104, metadata of the first web page 104, and the like.

[0033] At 206, the search engine 122 provides the first web page 104 to the content extractor module 124. The content extractor module 124 extracts content (web page content 126) from the first web page 104, wherein the web page content 126 is identified by the content extractor module 124 as relevant to the user input and / or the query. For example, the first web page 104 can be relatively dense and can include a large amount of text. The content extractor module 124 is configured to identify a portion of the first web page 104 that is relevant to the user input and / or the query, wherein the portion is sized such that the generative model 120 can receive the web page content 126 as input. At 208, the content extractor module 124 provides the web page content 126 to the generative model 120, wherein the web page content 126 is included in a prompt used by the generative model 120 to generate an output.

[0034] Upon receiving the prompt, the generative model 120 begins generating output (tokens in a sequence). At 210, and optionally as the generative model 120 generates such output, the generative model 120 provides the output to the client computing device 102. As noted above, the output can include text generated by the generative model 120 based on information in the prompt (user input, query, web page content 126, etc.). In addition, the output includes a reference generated by the generative model 120. For example, the reference identifies the first web page 104 as supporting the text in the output (e.g., a sentence in the output).

[0035] At 212, and substantially simultaneously with the transmission of the output to the client computing device 102, the output generated by the generative model 120 is provided to the text identifier module 128. The text identifier module 128 parses the output for references. When the text identifier module 128 identifies a reference in the output, the text identifier module 128 may identify the text corresponding to the reference generated by the generative model 120. For example, the text identifier module 128 may employ a rule or set of rules to identify the text corresponding to the reference.

[0036] At 214, in response to the text identifier module 128 identifying a reference in the output generated by the generative model 120 that identifies the first web page 104, and also in response to identifying text generated by the generative model 120 that is associated with such a reference, the text identifier module 128 provides the text and the web page content 126 (included in the hint) to the classifier 130. In another example, the text identifier module 128 may provide the entirety of the first web page 104 to the classifier 130.

[0037] As described above, the classifier 130 predicts whether the text generated by the generative model 120 is supported by the web page content 126. When the classifier 130 predicts that the text generated by the generative model 120 is supported by the web page content 126, the text generated by the generative model 120 is not modified. However, when the classifier 130 predicts that the text generated by the generative model 120 is not supported by the web page content 126 extracted from the first web page 104, then at 216, the text rewriting model 134 is provided with the text generated by the generative model 120 and the web page content 126. In addition, and optionally, the text rewriting model 134 can be provided with further information and can use this information to rewrite the text. For example, the text rewriting model 134 can be provided with the value of the logarithm probability (logit) of the generative model 120 corresponding to the text generated by the generative model 120. In another example, the text rewriting model 134 is provided with the value of the hidden layer of the generative model 120 corresponding to the text generated by the generative model 120. In yet another example, the text rewriting model 134 is provided with user input received by the generative model 120 from the client computing device 102. In yet another example, the text rewriting model 134 is provided with a query generated by the generative model 120. In yet another example, the text rewriting model 134 is provided with information from the classifier 130.

[0038] Further, although the text rewriting model 134 and the classifier 130 are illustrated as different models, in the example, the text rewriting model 134 itself includes a classification function. For example, the text rewriting model 134 (which includes a classification function) can include n hidden layers, and the text rewriting model 134 receives text / content pairs. The node value at the hidden layer m in the text rewriting model 134 can be used to determine whether the text should be written (therefore potentially avoiding the propagation of information from the layer mn of the text rewriting model). Accordingly, the representation from the mth layer (where m is a relatively small number compared to n) is used to determine whether text rewriting is needed. Therefore, when rewriting is not needed, the text rewriting model 134 is exited after layer m, and if text rewriting is needed, information is propagated through the remaining layers to create a representation for rewriting. In this embodiment, the text rewriting model 134 can have an encoder-decoder architecture similar to the T5 model.

[0039] Although the communication diagram 200 depicts the classifier 130 as providing this information to the text rewriting model 134, other communication flows are also contemplated. For example, the classifier 130 may provide the text rewriting model 134 with the text generated by the generative model 120 and the web page content 126. The text rewriting model 134 may then query the generative model 120 for log-probability values, values ​​of hidden layers, user input, etc.

[0040] The text rewriting model 134 rewrites the text generated by the generative model 120 based on the text generated by the generative model 120 and the web page content 126. Further, as indicated above, the text rewriting model 134 can optionally rewrite the text based on the log probability values ​​of the generative model 120, the values ​​of the hidden layers of the generative model 120, user input provided to the generative model 120, and / or other information.

[0041] At 218, the text rewrite model 134 provides the rewrite of the text to the client computing device 102. In an example, the rewrite of the text replaces the text generated by the generative model 120 as presented at the client computing device 102. In another example, a graphical indicator is presented at the client computing device 102 indicating that the rewrite of the text is available upon receiving the user input. In yet another example, the text generated by the generative model 120 and the rewrite of the text generated by the text rewrite model 134 are displayed simultaneously (and graphically distinguished) at the client computing device 102 so that a user of the client computing device 102 can identify the original text generated by the generative model 120 and the rewrite of the text.

[0042] Optionally, when the text rewriting model 134 generates a rewrite of the text, the text rewriting model 134 can provide the rewrite of the text to the classifier 130. The classifier 130 can then predict whether the web page content 126 supports the rewrite of the text. This can be repeated until the text rewriting model 134 generates a rewrite of the text that the web page content 126 supports.

[0043] The techniques described herein exhibit various advantages over conventional generative models that generate citations of text generated by the generative model. Specifically, the techniques described herein can identify hallucinations output by the generative model and correct such hallucinations, thereby improving the output presented to the end user.

[0044] In another example, the text rewriting model 134 can be configured to rewrite the entire output generated by the generation model 120 (rather than the sentence or phrase corresponding to the reference), wherein the text rewriting model 134 uses all the prompts adopted by the generation model 120 to rewrite the text to generate the output. The disadvantage of this method is that the text rewriting model 134 cannot rewrite the text generated by the generation model 120 until the generation model 120 has generated the entire output. On the contrary, the method described above allows the text rewriting model 134 to rewrite the part (e.g., phrase or sentence) of the output generated by the generation model 120 before the generation model 120 completes the output. Another disadvantage of configuring the text rewriting model 134 to rewrite the entire output of the generation model 120 is that the text rewriting model 134 may unnecessarily rewrite the part of the output, thereby consuming additional processing resources. On the contrary, the communication diagram 200 indicates that the text is provided to the text rewriting model 134 only when the classifier 130 has predicted that the text is not supported by the content of the source (e.g., webpage) identified in the text reference. Furthermore, providing a relatively small amount of text (e.g., one sentence instead of several sentences) to the text rewriting model 134 results in faster end-to-end rewriting time because more time (and processing resources) are required to build the query-key-value cache for longer input sequences, and each new generation time step will require more resources as the text rewriting model 134 processes longer context lengths (when the text rewriting model 134 is a transformer language model).

[0045] Figure 2 Another advantage of the illustrated approach is that the memory 110 can include multiple instances of the text rewriting model 134, thereby allowing several different text portions to be rewritten in parallel (thereby further speeding up the end-to-end rewriting process). For example, while a first unsupported text is being rewritten, a second unsupported text can be sent to another instance of the text rewriting model 134. Alternatively, if two different text portions are identified as unsupported in very short succession, the two text portions can be provided as a batch to a single instance of the text rewriting model 134.

[0046] Figure 2 Another advantage of the method illustrated in is that this method allows the already generated initial labeled text sequence (which is expensive to produce in terms of computing resources, especially in the case of long label lengths) to remain intact within the range of supported text, thereby saving processing resources.

[0047] Now refer to Figures 3 to 7 , depicts an example graphical user interface (GUI) corresponding to the generative model. Figure 3, the GUI 300 includes a text input field 302 in which a user has stated that "How fast were personal computers in 1975?" was entered into the text input field 302. As described above, the generative model 120 can provide this input to the search engine 122 (and / or generate a query based on this input and provide the query to the search engine), and the search engine 122 identifies a source (such as a web page) based on the user input and / or the query. The content extractor module 124 extracts content from the web page and provides the extracted content to the generative model 120 as part of a prompt for the generative model 120 to generate output. The graphical user interface 300 includes an output field 304 that depicts the output generated by the generative model 120. As described above, the generative model 120 can provide this input to the search engine 122 (and / or generate a query based on this input and provide the query to the search engine), and the search engine 122 identifies a source (such as a web page) based on the user input and / or the query. The content extractor module 124 extracts content from the web page and provides the extracted content to the generative model 120 as part of a prompt for the generative model 120 to generate output. Figure 3 As illustrated, the output includes two sentences (and will include more sentences as the generative model 120 continues to generate output), where each sentence has a reference associated with it that identifies a corresponding web page that supports the statement set forth in the sentence. In the example, the sentence "Computer XXXX has a clock speed of 4 MHz, meaning that it can execute four million instructions per second" is not supported by the content of the second web page identified in the reference associated with the sentence. Accordingly, the classifier 130 can predict that the content of the second web page does not support at least some of such sentences, and the text rewriting model 134 can rewrite the sentence.

[0048] Figure 4 GUI 300 is depicted in which output field 304 is updated to indicate to the user of client computing device 102 that the second sentence has been rewritten (while generative model 120 continues to generate output). For example, the second sentence generated by generative model 120 is displayed as having been deleted, and the rewriting of the text generated by text rewriting model 134 is displayed as underlined. Figure 4 The entire statement is depicted as struck through, but in other examples, the rewritten portion of the statement that differs from the original generated statement may be highlighted to the user.

[0049] Now go to Figure 5 , GUI 300 is presented in which the unsupported statements in output field 304 are highlighted in output field 304 (e.g., to indicate that the unsupported statements are selectable). For example, the unsupported statements may be bolded, assigned a particular color, have a hyperlink assigned to them, etc., to inform the user that classifier 130 has predicted that the highlighted statement is not supported by the source associated with the statement identified in the reference. Further, although not illustrated, the unsupported statements may have a probability value calculated by classifier 130 and displayed therewith, thereby providing the user with additional information regarding the likelihood that the determined statement should be rewritten. Now referring to Figure 6, GUI 300 is shown in which the user has instructed the text rewriting model 134 to rewrite Figure 5 . For example, a user may select an unsupported statement, and when the text rewriting model 134 receives an indication that the user has selected an unsupported statement, the text rewriting model 134 may generate a rewrite of the statement and replace the unsupported statement with it. In another embodiment, the text rewriting model 134 may rewrite the unsupported statement without user input based on a confidence value associated with the unsupported text (indicating the probability that the unsupported statement should be rewritten). In an example, a user may set a probability score threshold (e.g., in the user's profile), and the text rewriting model 134 may rewrite the text based on a probability score calculated for the text relative to the probability score threshold set by the user.

[0050] refer to Figure 7 , GUI 300 is shown where output field 304 includes additional text generated by generative model 120. Figure 7 To illustrate, the text rewriting model 134 can rewrite a portion of the output generated by the generative model 120 while the generative model 120 continues to generate output. That is, the text rewriting model 134 can operate in parallel with the generative model 120.

[0051] Figure 8 and Figure 9 The method related to rewriting the text generated by the generative model is illustrated. Although these methods are shown and described as a series of actions performed in sequence, it should be understood that these methods are not limited by the order of the sequence. For example, some actions can occur in an order different from the order described herein. In addition, an action can occur simultaneously with another action. Further, in some instances, not all actions are required to implement the methods described herein.

[0052] Furthermore, the actions described herein may be computer-executable instructions that can be implemented by one or more processors and / or stored on one or more computer-readable media. Computer-executable instructions may include routines, subroutines, programs, execution threads, etc. Furthermore, the results of the actions of these methods may be stored on computer-readable media, displayed on a display device, etc.

[0053] Now refer to Figure 8, a flow chart illustrating a method 800 for rewriting text is presented. Method 800 begins at 802, and at 804, text generated by a generative model and content of a content source (such as a web page) is provided to a computer-implemented text rewriting model. The generative model generates text based on user input received from a client computing device. In addition, the generative model generates a reference to the source to indicate that the text generated by the generative model is supported by the content of the source.

[0054] At 806, a computer-implemented text rewriting model generates a rewrite of the text. The text rewriting model generates the rewrite of the text based on the text generated by the generative model and the content of the source. At 808, the rewrite of the text is transmitted to the client computing device for presentation in response to user input received from the client computing device. Method 800 completes at 810.

[0055] Now go to Figure 9 , presents a method 900 for rewriting text generated by a generative model, wherein the method 900 is performed by a computing system. The method 900 begins at 902 and receives user input from a client computing device in network communication with the computing system at 904. At 906, based on the user input, a webpage including content related to the user input is identified. As described above, a search engine can identify a webpage based on the user input received from the client computing device.

[0056] At 908 , the generation model generates text and a reference identifying the web page, wherein the generation model generates the text and the reference based on the user input received at 904 and the content of the web page identified at 906 .

[0057] At 910, the text and the content of the web page are provided to a classifier that has been trained to output an indication (prediction) of whether the web page content provided to the generative model when generating the text supports the text generated by the generative model. At 912, a determination is made as to whether the text generated by the generative model is supported by the content of the web page identified in the reference.

[0058] When it is determined at 912 that the text generated by the generative model is not supported by the content of the webpage, at 914, the text and the content are provided to a computer-implemented text rewriting model. At 916, the text rewriting model generates a rewrite of the text. The text rewriting model generates a rewrite of the text based on the text generated by the generative model and the content of the webpage. At 918, the rewrite of the text is transmitted to the client computing device in response to the user input received from the client computing device. Method 900 ends at 920.

[0059] Now refer to Figure 10, illustrates a high-level diagram of an exemplary computing device 1000 that can be used in accordance with the systems and methods disclosed herein. For example, the computing device 1000 can be used in a system for rewriting text generated by a generative model. As another example, the computing device 1000 can be used in a system for predicting whether text generated by a generative model is supported by a source associated with the text identified in a reference. The computing device 1000 includes at least one processor 1002 that executes instructions stored in a memory 1004. The instructions can be, for example, instructions for implementing functions described as being performed by one or more components discussed above, or instructions for implementing one or more of the methods described above. The processor 1002 can access the memory 1004 via a system bus 1006. In addition to storing executable instructions, the memory 1004 can also store text, images, web page content, and the like.

[0060] Computing device 1000 additionally includes a data store 1008 that is accessible by processor 1002 via system bus 1006. Data store 1008 may include executable instructions, web indexes, instant answers, and the like. Computing device 1000 also includes an input interface 1010 that allows external devices to communicate with computing device 1000. For example, input interface 1010 may be used to receive instructions from an external computer device, from a user, and the like. Computing device 1000 also includes an output interface 1012 that interfaces computing device 1000 with one or more external devices. For example, computing device 1000 may display text, images, and the like via output interface 1012.

[0061] It is contemplated that external devices that communicate with the computing device 1000 via the input interface 1010 and the output interface 1012 may be included in an environment that provides substantially any type of user interface with which a user can interact. Examples of user interface types include graphical user interfaces, natural user interfaces, and the like. For example, a graphical user interface may accept input from a user employing (a plurality of) input devices (such as a keyboard, a mouse, a remote control, and the like) and provide output on an output device such as a display. Further, a natural user interface may enable a user to interact with the computing device 1000 in a manner that is not constrained by the constraints imposed by input devices (such as a keyboard, a mouse, a remote control, and the like). In contrast, a natural user interface may rely on speech recognition, touch and stylus recognition, gesture recognition on and near the screen, mid-air gestures, head and eye tracking, sound and voice, vision, touch, gestures, machine intelligence, and the like.

[0062] Additionally, although illustrated as a single system, it should be understood that the computing device 1000 may be a distributed system. Thus, for example, several devices may communicate via a network connection and may collectively perform the tasks described as being performed by the computing device 1000.

[0063] Various techniques are described herein with reference to at least the following examples.

[0064] (A1) In one aspect, a method performed by a computing system includes receiving user input from a client computing device in network communication with the computing system. The method also includes identifying a webpage that includes content related to the user input, wherein the webpage is identified based on the user input. The method also includes generating text and references related to the content using a generative model, wherein the generative model generates the text and references based on the user input and the content of the webpage. The method additionally includes providing the text and the content of the webpage to a classifier, wherein the classifier has been trained to output an indication of whether the webpage content provided to the generative model when generating the text supports the text generated by the generative model. The method also includes obtaining an indication from the classifier that the webpage content does not support the text. The method also includes providing the text and the content of the webpage to a computer-implemented text rewriting model, wherein the text and the content of the webpage are provided to the text rewriting model based on an indication from the classifier that the text is not supported by the content of the webpage. The method additionally includes generating a rewrite of the text using the computer-implemented text rewriting model, wherein the computer-implemented text rewriting model generates the rewrite of the text based on the text generated by the generative model and the content of the webpage. The method also includes transmitting the rewriting of the text to the client computing device in response to the user input received from the client computing device.

[0065] (A2) In some embodiments of the method of (A1), the method further comprises providing a value of a log probability of the generative model to a computer-implemented text rewriting model, wherein the value of the log probability corresponds to text generated by the generative model, and wherein the computer-implemented text rewriting model further generates a rewrite of the text based on the value of the log probability.

[0066] (A3) In some embodiments of the method of at least one of (A1) to (A2), the method further comprises providing a value of a hidden layer of the generative model to a computer-implemented text rewriting model, wherein the value of the hidden layer corresponds to text generated by the generative model, and wherein the computer-implemented text rewriting model also generates a rewrite of the text based on the value of the hidden layer.

[0067] (A4) In some embodiments of the method of at least one of (A1) to (A3), the method further includes determining that the reference is associated with the text, wherein the text and content are provided to the classifier based on determining that the reference is associated with the text. Further, determining that the reference is associated with the text occurs before the text and the reference are provided to the classifier.

[0068] (A5) In some embodiments of the method of (A4), the generative model generates additional text based on user input, wherein the additional text follows the text, and wherein the text and the content of the web page are also provided to the computer-implemented text rewriting model before the generative model completes the additional text.

[0069] (A6) In some embodiments of the method of at least one of (A1) to (A5), the method further includes transmitting the text and the reference to the client computing device before transmitting the rewrite of the text to the client computing device.

[0070] (A7) In some embodiments of the method of (A6), the method further includes causing a graphical indicator to be displayed at the client computing device to indicate that the text is not supported by the content of the web page, wherein the graphical indicator is displayed in response to obtaining an indication from the classifier that the text is not supported by the content of the web page.

[0071] (A8) In some embodiments of the method of at least one of (A1) to (A7), the method further includes providing user input to a computer-implemented text rewriting model, wherein the computer-implemented text rewriting model generates a rewrite of the text based on the user input.

[0072] (B1) In another aspect, a method performed by a computing system includes providing text generated by a generative model and content of a webpage to a computer-implemented text rewriting model, wherein the generative model generates the text based on user input received from a client computing device, and wherein the generative model also generates a reference to the webpage to indicate that the text generated by the generative model is supported by the content of the webpage. The method also includes generating a rewrite of the text by the computer-implemented text rewriting model, wherein the computer-implemented text rewriting model generates the rewrite of the text based on: 1) the text generated by the generative model; and 2) the content of the webpage. The method also includes transmitting the rewrite of the text to the client computing device for presentation in response to the user input received from the client computing device.

[0073] (B2) In some embodiments of the method of (B1), the method further includes providing the text and the content of the web page to a computer-implemented classifier, wherein the computer-implemented classifier is trained to determine whether the text generated by the generative model is supported by the content of the web page to which the generative model refers for the text, and wherein the text and the content of the web page generated by the generative model are also provided to the computer-implemented classifier before being provided to the computer-implemented text rewriting model.

[0074] (B3) In some embodiments of the method of (B2), the output of the generative model includes text and references to web pages. The method further includes, when the generative model generates the output, identifying that the references to web pages correspond to the text in the output. The method further includes, in response to identifying that the references to web pages correspond to the text in the output, providing the text and the content of the web page to a computer-implemented classifier.

[0075] (B4) In some embodiments of the method of at least one of (B1) to (B3), the method further includes providing instructions along with the text and the content of the web page to a computer-implemented text rewriting model, wherein the instructions instruct the text rewriting model to ensure that the rewriting of the text is supported by the content of the web page.

[0076] (B5) In some embodiments of the method of at least one of (B1) to (B4), the method further includes: before transmitting the rewrite of the text to the client computing device, transmitting the text and a reference to the web page to the client computing device for presentation as an initial response to the user input received from the client computing device.

[0077] (B6) In some embodiments of the method of (B5), when the client computing device receives the rewrite of the text, the text is replaced by the rewrite of the text.

[0078] (B7) In some embodiments of the method of (B5), the method further includes: before transmitting the rewrite of the text to the client computing device and after transmitting the text and the reference to the webpage to the client computing device, causing at least one of the text or the reference to the webpage to be highlighted to indicate that the text is not supported by content of the webpage. The method additionally includes receiving a request from the client computing device to provide the rewrite of the text to the client computing device, wherein the rewrite of the text is transmitted to the client computing device in response to receiving the request.

[0079] (B8) In some embodiments of the method of at least one of (B1) to (B7), the method further includes obtaining a logarithmic probability value of the generative model corresponding to the text generated by the generative model, wherein the computer-implemented text rewriting model also generates a rewrite of the text based on the logarithmic probability value.

[0080] (B9) In some embodiments of the method of at least one of (B1) to (B8), the method further includes obtaining a value of a hidden layer of the generative model corresponding to the text generated by the generative model, wherein the computer-implemented text rewriting model also generates a rewrite of the text based on the value of the hidden layer.

[0081] (B10) In some embodiments of the method of at least one of (B1) to (B9), the computer-implemented text rewriting model further generates the rewriting of the text based on user input.

[0082] (B11) In some embodiments of the method of at least one of (B1) to (B10), the method further includes providing the rewrite of the text and the content of the webpage to a computer-implemented text rewrite model. The method further includes generating, by the computer-implemented text rewrite model, a further rewrite of the text, wherein the computer-implemented text rewrite model generates the further rewrite of the text based on: 1) the rewrite of the text generated by the computer-implemented text rewrite model; and 2) the content of the webpage. The method additionally includes transmitting the further rewrite of the text to the client computing device for presentation as another response to user input received from the client computing device.

[0083] (C1) In another aspect, a computing system includes a processor and a memory, wherein the memory stores instructions that, when executed by the processor, cause the processor to perform at least one of the methods disclosed herein (e.g., any one of (A1) to (A8) or (B1) to (B11)).

[0084] (D1) In another aspect, a computer-readable storage medium includes instructions that, when executed by a processor, cause the processor to perform at least one of the methods disclosed herein (e.g., any one of (A1) to (A8) or (B1) to (B11)).

[0085] The various functions described herein can be implemented with hardware, software, or any combination thereof. If implemented with software, these functions can be stored on a computer-readable medium or transmitted on a computer-readable medium as one or more instructions or codes. Computer-readable media include computer-readable storage media. Computer-readable storage media can be any available storage medium that can be accessed by a computer. As an example and not limitation, such computer-readable storage media can include RAM, ROM, EEPROM, CD-ROM or other optical disk storage devices, magnetic disk storage devices or other magnetic storage devices, or any other medium that can be used to carry or store the desired program code in the form of instructions or data structures and can be accessed by a computer. Disks and optical disks as used herein include compact disks (CDs), laser disks, optical disks, digital versatile disks (DVDs), floppy disks, and Blu-ray disks (BDs), where disks typically reproduce data magnetically, while optical disks typically reproduce data optically using lasers. Further, propagation signals are not included within the scope of computer-readable storage media. Computer-readable media also include communication media, which include any media that facilitates the transfer of a computer program from one place to another. For example, a connection can be a communication medium. For example, if the software is transmitted from a website, server, or other remote source using a coaxial cable, fiber optic cable, twisted pair, digital subscriber line (DSL), or wireless technologies (such as infrared, radio, and microwave), then the coaxial cable, fiber optic cable, twisted pair, DSL, or wireless technologies (such as infrared, radio, and microwave) are included in the definition of communications media. Combinations of the above should also be included within the scope of computer-readable media.

[0086] Alternatively or additionally, the functions described herein may be performed, at least in part, by one or more hardware logic components. For example, but not limited to, illustrative types of hardware logic components that may be used include field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), systems on chips (SOCs), complex programmable logic devices (CPLDs), and the like.

[0087] What has been described above includes examples of one or more embodiments. Of course, for the purposes of describing the aforementioned aspects, it is not possible to describe every conceivable modification and alteration of the apparatus or method above, but those skilled in the art will recognize that many further modifications and permutations of the various aspects are possible. Accordingly, the described aspects are intended to encompass all such alterations, modifications, and variations that fall within the spirit and scope of the appended claims. Furthermore, to the extent the term "includes" is used in a specific embodiment or in a claim, such term is intended to be inclusive in a manner similar to how the term "comprising" is interpreted when employed as a transitional word in a claim.

Claims

1. A computing system comprising: processor; as well as a memory storing instructions that, when executed by the processor, cause the processor to perform actions including: providing text generated by a generative model and content of a web page to a computer-implemented text rewriting model, wherein the generative model generates the text based on user input received from a client computing device, and wherein the generative model also generates a reference to the web page to indicate that the text generated by the generative model is supported by the content of the web page; A rewrite of the text is generated by the computer-implemented text rewrite model, wherein the computer-implemented text rewrite model generates the rewrite of the text based on: the text generated by the generative model; and the content of the web page; and The rewriting of the text is transmitted to the client computing device for presentation in response to the user input received from the client computing device.

2. The computing system of claim 1 , wherein the actions further comprise: providing the text generated by the generative model and the content of the webpage to a computer-implemented classifier before providing the text and the content of the webpage to the computer-implemented text rewriting model, wherein the computer-implemented classifier is trained to determine whether the text generated by the generative model is supported by the content of the webpage to which the generative model refers for the text; as well as An indication is obtained from the computer-implemented classifier that the text generated by the generative model is not supported by the content of the webpage, wherein 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. The computing system of claim 2, wherein the output of the generative model comprises the text and the reference to the web page, the actions further comprising: As the generative model generates the output, identifying that the reference to the webpage corresponds to the text in the output; as well as In response to identifying that the reference to the web page corresponds to the text in the output, providing the text and the content of the web page to the computer-implemented classifier.

4. The computing system according to at least one of claims 1 to 3, wherein the actions further comprise: Instructions are provided to the computer-implemented text rewriting model along with the text and the content of the web page, wherein the instructions direct the text rewriting model to ensure that the rewriting of the text is supported by the content of the web page.

5. The computing system according to at least one of claims 1 to 4, wherein the actions further comprise: Prior to transmitting the rewriting of the text to the client computing device, the text and the reference to the webpage are transmitted to the client computing device for presentation as an initial response to the user input received from the client computing device. 6 . The computing system of claim 5 , wherein the text is replaced by the rewrite of the text when the client computing device receives the rewrite of the text.

7. The computing system of claim 5, wherein the actions further comprise: prior to transmitting the rewriting of the text to the client computing device and after transmitting the text and the reference to the webpage to the client computing device, causing at least one of the text or the reference to the webpage to be highlighted to indicate that the text is not supported by the content of the webpage; as well as A request is received 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. The computing system of at least one of claims 1 to 7, the actions further comprising: A value of a log probability of the generative model corresponding to the text generated by the generative model is obtained, wherein the computer-implemented text rewriting model further generates the rewriting of the text based on the value of the log probability.

9. The computing system of at least one of claims 1 to 8, wherein the actions further comprise: A value of a hidden layer of the generative model corresponding to the text generated by the generative model is obtained, wherein the computer-implemented text rewriting model also generates the rewriting of the text based on the value of the hidden layer.

10. The computing system of at least one of claims 1 to 9, wherein the computer-implemented text rewriting model generates the rewriting of the text further based on the user input.

11. The computing system of at least one of claims 1 to 10, wherein the actions further comprise: providing the rewriting of the text and the content of the web page to the computer-implemented text rewriting model; as well as A further rewrite of the text is generated by the computer-implemented text rewrite model, wherein the computer-implemented text rewrite model generates the further rewrite of the text based on: the rewriting of the text generated by the computer-implemented text rewriting model; and the content of the web page; as well as The further rewrite of the text is transmitted to the client computing device for presentation as another response to the user input received from the client computing device.

12. A method performed by a computing system, the method comprising: receiving user input from a client computing device in network communication with the computing system; Based on the user input, identifying a webpage that includes content related to the user input; generating text and quotations related to the content by a generative model, wherein the generative model generates the text and the quotations based on the user input and the content of the webpage; providing the text and the content of the webpage to a classifier, wherein the classifier has been trained to output an indication of whether the text generated by the generative model when generating the text is supported by the webpage content provided to the generative model; obtaining, from the classifier, an indication that the text is not supported by the content of the webpage; providing the text and the content of the web page to a computer-implemented text rewriting model based on the indication from the classifier that the text is not supported by the content of the web page; generating, by the computer-implemented text rewriting model, a rewriting of the text, wherein the computer-implemented text rewriting model generates the rewriting of the text based on the text generated by the generative model and the content of the webpage; as well as The rewriting of the text is transmitted to the client computing device in response to the user input received from the client computing device.

13. The method of claim 12, further comprising providing a value of the log probability of the generative model to the computer-implemented text rewriting model, wherein the value of the log probability corresponds to the text generated by the generative model, and wherein the computer-implemented text rewriting model also generates the rewriting of the text based on the value of the log probability.

14. The method according to at least one of claims 12 to 13 further comprises providing the values ​​of the hidden layer of the generative model to the computer-implemented text rewriting model, wherein the values ​​of the hidden layer correspond to the text generated by the generative model, and wherein the computer-implemented text rewriting model also generates the rewriting of the text based on the values ​​of the hidden layer.

15. The method according to at least one of claims 12 to 14, further comprising: Prior to providing the text and the reference to the classifier, a determination is made that the reference is related to the text, wherein the text and the content are provided to the classifier based on the determination that the reference is related to the text.

Images