Machine-learning techniques for augmenting electronic documents with data-verification indicators

TECHNICAL FIELD

0001The present disclosure generally relates to machine-learning techniques for transforming or otherwise processing electronic data. More specifically, but not by way of limitation, the present disclosure relates to machine-learning techniques that generate data-verification indicators that are applied to electronic documents, such as techniques that use neural networks for retrieving and analyzing evidence regarding content of electronic documents and generating indicators of veracity to be displayed with that content via electronic reading software.

BACKGROUND

0002Online content has proliferated significantly. Identifying the veracity of statements (e.g., claims or assertions published online) is an important aspect of consuming content. With the bulk of information available online, it has become increasingly important to build technology that automatic determines the veracity of published statements. The wide necessity and applicability of such a technology has given rise to the research problem of Fact Extraction and Verification (FEVER), which has the objective of determining the veracity of a statement by identifying the supporting or refuting evidence contained in the set of documents. Generally, existing FEVER systems retrieve one or more sentences from the set of documents, which could serve as evidence when performing reasoning tasks, such as recognizing, via natural language processing, a textual entailment of the retrieved sentences with respect to a statement (e.g., reasoning over the selected sentences to determine the veracity of a statement). In a simplified example, a computing system that performs natural language processing on the statement “The sky was blue on November 2” could recognize a textual entailment of the statement with respect to a weather report indicating a lack of cloud cover on November 2 (e.g., the truth of the statement involving the blue sky follows logically from the weather report).

0003But existing FEVER systems do not provide feedback from the reasoning steps to the evidence extraction steps. For instance, in the example above in which the statement involves the blue sky on November 2, the reasoning step may be unable to classify the statement as supported or refuted, given the set of documents available (e.g., a classification of “not enough information” from the extracted evidence). In this situation, existing FEVER systems would not perform evidence extraction again to find additional content in the set of available document to support or refute the statement. Additionally, existing techniques are unable to handle the varying levels of complexity involved in combining the sentences retrieved across multiple documents to support or refute the veracity of the statement. For instance, in the example above in which the statement involves the blue sky, the statement may be “The sky was blue on November 2, but traffic was heavy on the Interstate 280.” In this situation, the reasoning step has to consider multiple pieces of evidence from multiple documents in the set of available documents (e.g., the weather report for November 2 and a traffic report for the same day). But, existing FEVER systems are unequipped to verify real-world statements, which may need evidence in the form of multiple sentences retrieved from multiple documents to determine the veracity of the statements.

SUMMARY

0004Certain aspects and features of the present disclosure relate to a computer-implemented method. The computer-implemented method includes receiving an input indicating a selection of a target text segment and determining a veracity metric of the target text segment. Determining the veracity metric includes extracting a first text segment from an electronic document and generating, using a trained reasoning model, an output. The output classifies the first text segment as lacking information to support or refute the target text segment. The computer-implemented method includes extracting a second text segment from the electronic document or another electronic document in response to the trained reasoning model classifying the first text segment as lacking information to support or refute the target text segment. The computer-implemented method include classifying, with the trained reasoning model, a combination of the first text segment and the second text segment according to a degree to which the trained reasoning model measures whether the combination supports the target text segment. The computer-implemented method also includes generating an augmentation for a graphical interface to display, adjacent to the target text segment, a verification indicator representing the determined veracity metric of the target text segment. Other embodiments of this aspect include corresponding computer systems, apparatus, and computer programs recorded on one or more computer storage devices, each configured to perform the actions of the methods.

0005This summary is not intended to identify key or essential features of the claimed subject matter, nor is it intended to be used in isolation to determine the scope of the claimed subject matter. The subject matter should be understood by reference to appropriate portions of the entire specification of this disclosure, any or all drawings and each claim.

BRIEF DESCRIPTION OF THE DRAWINGS

0006Features, implementations, and advantages of the present disclosure are better understood when the following Detailed Description is read with reference to the accompanying drawings.

0007<figref idref="DRAWINGS">FIG. 1</figref> depicts an example of a computing environment for determining the veracity of text segments, according to some aspects of the present disclosure.

0008<figref idref="DRAWINGS">FIG. 2</figref> depicts an example of a text verification system for determining the veracity of a target text segment, according to some aspects of the present disclosure.

0009<figref idref="DRAWINGS">FIG. 3</figref> depicts an example of multi-stage text retrieval model for selecting text segments from electronic documents to support or refute a target text segment, according to some aspects of the present disclosure.

0010<figref idref="DRAWINGS">FIG. 4</figref> depicts an example of a specialized reasoning model for generating inferences indicating whether text segments support or refute a target text segment, according to some aspects of the present disclosure.

0011<figref idref="DRAWINGS">FIG. 5</figref> depicts an example of using a text verification system to generate a verification indicator representing the inferred veracity of a target text segment, according to some aspects of the present disclosure.

0012<figref idref="DRAWINGS">FIG. 6</figref> depicts an example of a process for determining the veracity of a target text segment using a corpus of electronic documents, according to some aspects of the present disclosure.

0013<figref idref="DRAWINGS">FIG. 7</figref> depicts an example of an electronic document augmented to display multiple verification indicators, according to some aspects of the present disclosure.

0014<figref idref="DRAWINGS">FIG. 8</figref> depicts an example of a cloud computing system for implementing certain aspects described herein.

0015<figref idref="DRAWINGS">FIG. 9</figref> depicts an example of a computing system for implementing certain aspects described herein.

DETAILED DESCRIPTION

0016Certain aspects of the present disclosure relate to executing machine-learning models to automatically infer the veracity of a target text segment (e.g., a statement) using evidence (e.g., one or more sentences) retrieved from electronic documents included in a text corpus (e.g., a set of electronic documents). For example, a text verification system, which could be included in or accessed by electronic reading software that displays electronic documents on a display of a user device, executes the machine-learning models described herein to perform an evidence extraction and verification function relating to a target text segment included in a target electronic document.

0017The text verification system determines the veracity metric of a target text segment by retrieving evidence retrieved from a text corpus stored in an electronic document database. For instance, the target electronic document is any electronic document being displayed on a user device by the document display system, and the target text segment is any alphanumeric text, such as a sentence (or a portion thereof), included in a target electronic document. Examples of the target text segment include an assertion, claim, declaration, evidence, contention, rejection, or denial. Evidence is retrieved in the form of evidence text segments, such as sentences (or portions of sentences) that are retrieved from the text corpus as evidence to support or refute the target text segment.

0018The text verification system includes a document retrieval system, a two-stage evidence selection model, and a specialized reasoning model that operate in a network to retrieve evidence, in the form of alphanumeric text segments, from the text corpus for the purpose of supporting or refuting the target text segment. The document retrieval system searches the text corpus for electronic documents that are relevant to the target text segment. The two-stage evidence selection model executes a lazy retrieval protocol to retrieve evidence text segments from the relevant electronic documents. The lazy retrieval protocol includes a first stage and a conditional second stage. During the first stage, the two-stage evidence selection model analyzes the relevant documents, retrieves evidence text segments that are directly relevant to the target text segment (referred to herein as first-stage evidence), and transmits the first-stage evidence to the specialized reasoning model. An evidence text segment is directly relevant to the target text segment if the evidence text segment supports or refute the target text segment on its own (e.g., without needing any additional information).

0019The specialized reasoning model reasons over the first-stage evidence and the target text segment to classify the veracity of the target text segment as supported by the first-stage evidence, refuted by the first-stage evidence, or undetermined based on a lack of information in the first-stage evidence (this step is referred to herein as a first stage of reasoning). If the specialized reasoning model classifies the veracity of the target text segment as supported or refuted by the first-stage evidence, then the text verification system causes the electronic reading software to augment the target electronic document by displaying a verification indicator in association with (e.g., adjacent to) the target text segment. The verification indicator represents the classification.

0020If, however, the specialized reasoning model classifies the veracity of the target text segment as undetermined, then the lazy retrieval protocol proceeds to the conditional second stage, wherein the two-stage evidence selection model retrieves evidence text segments that are indirectly relevant to the target text segment (referred to herein as second-stage evidence). An evidence text segment is indirectly relevant to the target text segment if the evidence text segment cannot support or refute the target text segment on its own, but rather, needs to the context of another evidence text segment to support or refute the target text segment. The specialized reasoning model reasons over the target text segment and a combination of the first-stage evidence and the second-stage evidence (this step is referred to as the second stage of reasoning). The specialized reasoning model classifies the veracity of the target text segment as supported by the first-stage evidence and the second-stage evidence or refuted by the first-stage evidence and the second-stage evidence. The text verification system causes the electronic reading software to augment the target electronic document by displaying a verification indicator in association with (e.g., adjacent to) the target text segment.

0021As an illustrative example, a target electronic document includes the following target text segment: “Ryan Gosling has been to a country in Africa.” The text verification system receives a signal initiating the evidence extraction and verification function with respect to the target text segment. The document retrieval system retrieves relevant documents, such as documents relating to Ryan Gosling and documents relating to the country of Africa, from the text corpus. The two-stage evidence selection model analyzes the relevant documents to retrieve first-stage evidence. The first-stage evidence is a statement referring to Ryan Gosling: “Ryan Gosling has traveled to Chad, Uganda, and eastern Congo.” The specialized reasoning model classifies the veracity of the target text segment using the first-stage evidence as undetermined because the first-stage evidence does not associate Chad with the country of Africa. In response, the two-stage evidence selection model proceeds to the conditional second stage, where the two-stage evidence selection model retrieves second-stage evidence from the relevant documents. The second-stage evidence is a statement referring to Africa: “Chad is a land-locked country in Central Africa.” The specialized reasoning model reasons over the target text segment and a combination of the first-stage evidence (“Ryan Gosling has traveled to Chad, Uganda, and eastern Congo”) and the second-stage evidence (“Chad is a land-locked country in Central Africa”). Lastly, the specialized reasoning model classifies the veracity of the target text segment (“Ryan Gosling has been to a country in Africa”) as supported. The text verification system causes the electronic reading software to augment the target electronic document by displaying the verification indicator (e.g., “Verified”) next to the target text segment.

0022While existing FEVER systems can determine the veracity of complex statements involving multiple sentences of evidence from a text corpus, the existing FEVER systems are burdensome to network resources because these systems retrieve a large amount of evidence a reasoning model to process. For example, certain existing FEVER systems unconditionally perform evidence retrieval for direct evidence and indirect evidence to support or refute a statement. Unconditionally performing evidence retrieval for direct evidence and indirect evidence creates a large data set of sentences that increases the processing time needed to generate a result. The implementations described in the present disclosure, however, provide a technical improvement by reducing the amount of evidence and the processing time needed to determine the veracity of the target text segment. In particular, the lazy retrieval protocol, in which the two-stage evidence selection model retrieves indirect evidence (e.g., second-stage evidence) only if the specialized reasoning model is unable to classify the veracity of the target text segment given the first-stage evidence, provides a specific improvement to the operation of network resources by reducing the amount of indirect evidence that is processed, which in turn, improves the processing time for determining the veracity of a target text segment.

0023<figref idref="DRAWINGS">FIG. 1</figref> depicts an example of a computing environment for performing evidence retrieval and verification functionality, according to some aspects of the present disclosure. In this example, <figref idref="DRAWINGS">FIG. 1</figref> illustrates a document display system <b>100</b>. Document display system <b>100</b> includes any suitable computer system including, for example, server computer <b>805</b> of <figref idref="DRAWINGS">FIG. 8</figref> and/or computing device <b>900</b> of <figref idref="DRAWINGS">FIG. 9</figref>. User system <b>135</b> is any suitable computer system including, for example, any of user devices <b>825</b><i>a</i>-<i>c </i>of <figref idref="DRAWINGS">FIG. 8</figref> and/or computing device <b>900</b> of <figref idref="DRAWINGS">FIG. 9</figref>. A user may utilize user system <b>135</b> to access the document display system <b>100</b> via user interface (UI) subsystem <b>140</b>.

0024In certain implementations, the document display system <b>100</b> provides an electronic document reader <b>105</b> that incorporates a text verification system <b>110</b> for providing the evidence retrieval and verification functionality for target text segments included in target electronic documents. The evidence retrieval and verification functionality includes machine-learning techniques for displaying, transforming, or otherwise processing electronic documents. In certain implementations, document display system <b>100</b> provides users with electronic reading functionality, including providing the capability to display and edit electronic documents, for example, regarding the size, shape, presence of an object or text, color, or any other attribute of an electronic document or portion of an electronic document.

0025Electronic document reader <b>105</b> may be implemented using software, hardware, firmware, or any combination thereof. In some implementations, the electronic document reader <b>105</b> includes UI subsystem <b>140</b> that communicates with a user system <b>135</b> operated by a user. The electronic document reader <b>105</b> also includes the text verification system <b>110</b> for performing some or all of the electronic document reader <b>105</b> functionality (e.g., automatically determining the veracity of a target text segment, as described herein).

0026Text verification system <b>110</b> determines the veracity metric of a displayed text segment (e.g., a target text segment) by selecting one or more text segments displayed within the electronic document (e.g., the target text segment) and initiating a process for determining the veracity metric of the selected text segments. Examples of such evidence retrieval and verification functionality are described in more detail with respect to <figref idref="DRAWINGS">FIGS. 5 and 7</figref>.

0027In some implementations, the text verification system <b>110</b> includes a document retrieval system <b>115</b>, a two-stage evidence selection system <b>120</b>, and a specialized reasoning model <b>125</b>. The document retrieval system <b>115</b> identifies a subset of electronic documents from amongst the set of electronic documents included in the text corpus stored in the electronic document database <b>130</b>. For example, the document retrieval system <b>115</b> determines whether the electronic document is relevant to a given target text segment, and if so, then the document retrieval system <b>115</b> includes the electronic document in the subset of electronic documents (e.g., as documents relevant to the target text segment). In some implementations, the document retrieval system <b>115</b> performs a named entity recognition on the target text segment to extract one or more potential entities included in the target text segment. The document retrieval system <b>115</b> then performs an indexed search of each electronic document included in the text corpus using the one or more potential entities. The electronic documents that result from the indexed search are relevant to the target text segment. For example, an entity is a described concept, such as physical objects (e.g., a group of people, an individual person, or an inanimate object, such as a beach ball) or abstractions (e.g., feeling happy). Entities are represented in text by name, indicated by a noun or noun phrase, or represented by a pronoun.

0028The two-stage evidence selection system <b>120</b> performs evidence retrieval from the subset of electronic documents using the lazy retrieval protocol. The lazy retrieval protocol includes a first stage and a conditional second stage. During the first stage, the two-stage evidence selection system <b>120</b> evaluates the various text segments (e.g., sentences) included in each electronic document of the subset of electronic documents, and retrieves first-stage evidence that is directly relevant to the target text segment. The specialized reasoning model <b>125</b> classifies the veracity of the target text segment given the first-stage evidence. For example, the classifications are supported by the first-stage evidence, refuted by the first-stage evidence, or undetermined because the first-stage evidence lacks information to support or refute the target text segment. If the specialized reasoning model <b>125</b> classifies the veracity of the target text segment as supporting or refuting the target text segment, then the text verification system <b>110</b> generates a verification indicator representing the classification. The text verification system <b>110</b> also augments the displayed target electronic document to display the verification indicator in association with the displayed target text segment.

0029The lazy retrieval protocol proceeds to the conditional second stage of evidence retrieval if the specialized reasoning model <b>125</b> classifies the veracity of the target text segment, given the first-stage evidence, as undetermined. During the second stage, the two-stage evidence selection system <b>120</b> retrieves evidence text segments (from the subset of electronic documents) that are indirectly relevant to the target text segment. The two-stage evidence selection system <b>120</b> transmits the second-stage evidence to the specialized reasoning model <b>125</b>. The specialized reasoning model <b>125</b> performs textual entailment between the target text segment and a combination of the first-stage evidence and the second-stage evidence. The specialized reasoning model classifies the veracity of the target text segment based on the results of the textual entailment. For example, the groups of categories for classification include “supported” (e.g., by the second-stage evidence in context with the first-stage evidence), “refuted” (e.g., by the second-stage evidence in context with the first-stage evidence), or “undetermined” because the combination of the first-stage evidence and the second stage evidence lacks information to support or refute the target text segment.

0030To illustrate the text verification system <b>110</b> in use, and only as a non-limiting example, a user operates the user system <b>135</b> to load the electronic document reader <b>105</b>. The user selects a target text segment from a target electronic document (indicated by arrow <b>145</b>), and then triggers the evidence retrieval and verification functionality provided by the electronic document reader <b>105</b> (indicated by arrow <b>150</b>) using a user interface that is displayed or provided on user system <b>135</b>. Other communications may be transmitted or received indicated by arrow <b>150</b>. The UI subsystem <b>140</b> receives the selection of the target text segment and the indication to execute the evidence retrieval and verification functionality. The UI subsystem <b>140</b> transmits the target text segment to the text verification system <b>110</b> as an input. The document retrieval system <b>115</b> can select a subset of electronic documents from the set of electronic documents stored in electronic document database <b>130</b>, such that the subset of electronic documents are determined to be relevant to the target text segment. The document retrieval system <b>115</b> transmits the subset of electronic documents (or identifiers thereof) to the two-stage evidence selection system <b>120</b> (as indicated by arrow <b>170</b>). The two-stage evidence selection system <b>120</b> extracts first-stage text segments (e.g., text segments retrieved during the first stage of evidence retrieval) and transmits the first-stage text segments to the specialized reasoning model <b>125</b> (as indicated by arrow <b>180</b>). The specialized reasoning model <b>125</b> is unable to classify the first-stage text segments as supporting or refuting the target text segment. The classification of “not enough information” is transmitted from the specialized reasoning model <b>125</b> back to the two-stage evidence selection system <b>120</b> (as indicated by arrow <b>185</b>). The two-stage evidence selection system <b>120</b> then proceeds to the conditional second stage and extracts second-stage text segments based on the first-stage text segments. Extracting the second-stage text segments based on the first-stage text segments includes searching through the set of electronic documents using the first-stage text segments as a target and identifying electronic documents that may be relevant. The text segments from these documents are inputted back into the two-stage evidence selection system <b>120</b> to identify the second-stage text segments. To illustrate and only as a non-limiting example, the first-stage text segment of “He won two Emmy Awards for his performances in Winchell and Monk” is retrieved from Stanley Tucci's Wikipedia page, however, the second-stage text segment of “Monk is an American comedy-drama detective mystery television series” is retrieved from the Wikipedia page of the TV show “Monk”. The two-stage evidence selection system <b>120</b> transmits the second-stage text segments to the specialized reasoning model <b>125</b>. The specialized reasoning model <b>125</b> performs joint evidence reasoning to classify the first-stage text segments and second-stage text segments as supporting or refuting the target text segment. The specialized reasoning model <b>125</b> generates a verification indicator representing the classification and augments the target electronic document displayed using UI subsystem <b>140</b> (as indicated by arrow <b>195</b>).

0031While only three components are depicted in the text verification system of <figref idref="DRAWINGS">FIG. 1</figref> (e.g., the document retrieval system <b>115</b>, the two-stage evidence selection system <b>120</b>, and the specialized reasoning model <b>125</b>), the text verification system <b>110</b> includes any number of components or neural network layers in a pipeline.

0032<figref idref="DRAWINGS">FIG. 2</figref> depicts an example of the text verification system <b>110</b> for determining the veracity of a target text segment, according to some aspects of the present disclosure. The text verification system <b>110</b> in this example generates a verification indicator (e.g., verification indicator <b>215</b> or <b>220</b>) in response to receiving a target text segment <b>205</b> as an input. The verification indicators <b>215</b> or <b>220</b> represent the specialized reasoning model <b>125</b> inferring the veracity of the target text segment <b>205</b> given first-stage evidence or given a combination of first-stage evidence and second-stage evidence. Similar to the illustration of text verification system <b>110</b> in <figref idref="DRAWINGS">FIG. 1</figref>, the text verification system <b>110</b> shown in <figref idref="DRAWINGS">FIG. 2</figref> also includes the document retrieval system <b>115</b>, the two-stage evidence selection system <b>120</b>, and the specialized reasoning model <b>125</b>.

0033The target text segment is transmitted to the document retrieval system <b>115</b>. The document retrieval system <b>115</b> identifies a group of electronic documents (from the text corpus) that are relevant to the target text segment <b>205</b>. In some implementations, the document retrieval system <b>115</b> performs an entity-based recognition technique on the target text segment <b>205</b> to identify one or more potential entities described by the target text segment <b>205</b>. The document retrieval system <b>115</b> then performs an index-based search of the electronic documents of the text corpus to identify which electronic documents are “relevant” to the target text system <b>115</b>. An electronic document is determined to be relevant if the result of the index-based search indicates that the electronic document includes text representing the potential entities. The subset of electronic documents (or identifiers thereof) that the document retrieval system <b>115</b> identified as being relevant to the target text segment is transmitted to the two-stage evidence selection system <b>120</b> (as indicated by arrow <b>250</b>).

0034The two-stage evidence selection system <b>120</b> includes a stage-1 text retrieval model <b>225</b> and a stage-2 text retrieval model <b>230</b>. In some implementations, the stage-1 text retrieval model is a trained BERT model that receives as input each text segment of the subset of electronic documents and generates as an output a relevance score indicating a relevance between the text segment and the target text segment <b>205</b>. As a non-limiting example, the stage-1 text retrieval model <b>225</b> selects, for example, the text segments associated with the top five relevance scores as the first-stage evidence. Similarly, the stage-2 text retrieval model <b>230</b> is also a trained BERT model that generates a relevance score. The stage-1 text retrieval model <b>225</b> and the stage-2 text retrieval model <b>230</b> are each trained individually using different techniques, described below.

0035In some implementations, training the stage-1 text retrieval model <b>225</b> includes training a BERT model to identify text segments that could individually support or refute the target text segment (e.g., without needing the context of a different text segment, or in other words, text segments that support or refute the given target text segment on their own). Evidence text segments that directly support or refute a given target text segment are labeled as positive examples. After the positive examples are labeled, the remaining text segments of the subset of electronic documents are labeled as negative examples (e.g., negative examples are text segments that are not retrieved as evidence for supporting or refuting the target text segment). As a non-limiting example, the input to the stage-1 text retrieval model <b>225</b> for training purposes is “‘[CLS]’+claim+‘[SEP]’+page title+‘:’+evidence sentence+‘[SEP]’”. In this example, the term “[CLS]” is the BERT classification token. The term [SEP] informs the BERT model that a new sentence is starting. The BERT model is a pre-trained model using the format “[CLS] Sentence A [SEP] Sentence B [SEP].” The BERT model can detect whether there are any patterns or correlations between Sentence A and Sentence B. For example, in the training code line above, the BERT model can detect whether there are any correlations between the “claim” (e.g., the target text segment) and the “evidence sentence.” Further, a hinge loss function is used for sentence selection to maximize a margin between positive and negative examples.

0036To train the stage-2 text retrieval model <b>230</b>, a dataset of target text segments is created, such that each target text segment in the dataset requires multiple sentences of evidence to support or refute the target text segment (e.g., multiple sentences or phrases that could be used as evidence to support or refute the target text segment, as the “Ryan Gosling” example described above) is labeled as positive examples. Electronic documents of the text corpus from which first-stage evidence is retrieved are labeled as negative samples. Text segments that are not included in the evidence set (e.g., the text segments retrieved for the purpose of determining the veracity of a target text segment) are sampled from these documents. As a non-limiting example, the input to the second-stage BERT model (for the purpose of training the second-stage BERT model) is “‘[CLS]’+claim+page title_1+evidence_sentence_1+‘[SEP]’+page_title_2+‘:’+evidence sentence_2+‘[SEP]’”. Accordingly, in this example, the BERT model determines whether there are any patterns or correlations between “claim+page title_1+evidence_sentence_1” and “page_title_2+‘:’+evidence sentence_2.”

0037At the time of inference (e.g., after both the stage-1 text retrieval model <b>225</b> and the stage-2 text retrieval model <b>230</b> have been trained), the two-stage evidence selection system <b>120</b> passes each text segment included in the subset of electronic documents through the stage-1 text retrieval model <b>225</b>. The stage-1 text retrieval model <b>225</b> generates, as an output, a relevance score for each inputted text segment. The relevance score represents a degree to which the associated text segment is relevant to the target text segment. As a non-limiting example, the text segments associated with the top five relevance scores can be selected for retrieval. If the lazy retrieval protocol proceeds to the conditional second-stage of evidence selection, then the electronic documents from which the first-stage evidence was extracted are inputted into the stage-2 text retrieval model <b>230</b>. Similar to the stage-1 text retrieval model <b>225</b>, the stage-2 text retrieval model <b>230</b> generates, as an output, a relevance score for each inputted text segment. As a non-limiting example, the text segments associated with the top three relevance scores are selected for retrieval. These five text segments (in the case of first-stage only evidence retrieval) or eight text segments (if the second-stage evidence retrieval is triggered) are passed to the specialized reasoning model <b>125</b> for first- or second-stage reasoning, respectively.

0038The first-stage evidence is transmitted to the kernel node attention model <b>235</b> of the specialized reasoning model <b>125</b> (as indicated by arrow <b>255</b>). During the first stage of reasoning, the specialized reasoning model <b>125</b> reasons over the first-stage evidence and the target text segment to infer the veracity of the target text segment, given the first-stage evidence. As described in greater detail with respect to <figref idref="DRAWINGS">FIG. 3</figref>, the kernel node attention model <b>235</b> is a sub-tree attention network that generates an evidence graph to represent the first-stage evidence or the second-stage evidence, and then reasons over the evidence graph and the target text segment to classify the veracity of the target text segment. If the kernel node attention model <b>235</b> classifies the veracity of the target text segment as supporting or refuting the target text segment <b>205</b>, then the text verification system <b>110</b> generates the verification indicator <b>215</b> to represent the classification. The verification indicator <b>215</b> is used to augment the target electronic document to display the inferred veracity of the target text segment <b>205</b>. If, however, the kernel node attention model <b>235</b> classifies the veracity of the target text as undetermined given a lack of information in the first-stage evidence to either support or refute the target text segment <b>205</b>, then the specialized reasoning model <b>125</b> transmits the classification to the stage-2 text retrieval model <b>230</b> for retrieving second-stage evidence (as indicated by arrow <b>260</b>). The joint reasoning graph model <b>240</b> reasons over target text segment <b>205</b> and the combined first-stage evidence and the second-stage evidence to generate an inference as to the veracity of the target text segment <b>205</b>. The joint reasoning graph model <b>240</b> is described in greater detail with respect to <figref idref="DRAWINGS">FIG. 4</figref>.

0039<figref idref="DRAWINGS">FIG. 3</figref> depicts an example of the two-stage evidence selection system <b>120</b> for selecting text segments from electronic documents to support or refute a target text segment, according to some aspects of the present disclosure. The two-stage evidence selection system <b>120</b> in this example retrieves first-stage evidence from a text corpus, and conditionally retrieves second-stage evidence from certain electronic documents of the text corpus. As illustrated in <figref idref="DRAWINGS">FIG. 3</figref>, the two-stage evidence selection system <b>120</b> includes a stage-1 text retrieval model <b>225</b> and a stage-2 text retrieval model <b>230</b>.

0040In some implementations, stage-1 text retrieval model <b>225</b> includes first-stage BERT model <b>310</b> and evidence retrieval system <b>320</b>. The first-stage BERT model <b>310</b> is a pre-trained machine-learning network for performing natural language processing (NLP) tasks. The first-stage BERT model <b>310</b> is trained to detect the meaning of unlabeled text by using nearby text to determine contextual information. For example, the first-stage BERT model <b>310</b> is pre-trained using text from WIKIPEDIA and can be further fine-tuned. The first-stage BERT model <b>310</b> is trained using the training techniques described with respect to <figref idref="DRAWINGS">FIG. 2</figref>.

0041In some implementations, training the first-stage BERT model <b>310</b> includes labeling a set of target text segments (e.g., claims) as positive examples. Each target text segment of the set of target text segments requires retrieving a single text segment from the text corpus to support or refute the target text segment. The sentences that are not used to directly support or refute each target text segment of the set of target text segments are labeled as negative examples (e.g., negative examples are text segments that are not retrieved as evidence for supporting or refuting the target text segment). As a non-limiting example, during training, the input to the first-stage BERT model <b>310</b> is “‘[CLS]’+claim+‘[SEP]’+page title+‘:’+evidence sentence+‘[SEP]’”. Further, a hinge loss function is used for sentence selection to maximize a margin between positive and negative examples.

0042At the time of first-stage evidence retrieval in production (e.g., after the first-stage BERT model <b>310</b> has been trained), all of the text segments of the subset of electronic documents (yielded by the document retrieval system <b>115</b>) stored in the electronic document database <b>130</b> are individually inputted into the first-stage BERT model <b>310</b>. The first-stage BERT model <b>310</b> generates, as an output, a relevance score for each inputted text segment <b>305</b>. The relevance score represents a degree to which the associated text segment is relevant to the target text segment <b>205</b>. The evidence retrieval system <b>320</b> orders the text segments according to their associated relevance scores. As a non-limiting example, the evidence retrieval system <b>320</b> ranks all of the text segments included in the subset of electronic documents by their relevance score, identifies the top relevance scores and the associated text segments, and then retrieves or extracts the top five text segments as first-stage evidence (as indicated by arrow <b>315</b>). The first-stage evidence is transmitted to the specialized reasoning model <b>325</b> for performing the first stage of reasoning, which generates a classification for each of the top five text segments included in the first-stage evidence. The classifications include supporting the target text segment <b>205</b>, refuting the target text segment <b>205</b>, or neither supporting nor refuting the target text segment <b>205</b> (e.g., a classification of “not enough information” to support or refute the target text segment <b>205</b>). The classification is based on the combined score computed as the weighted combination of each of the five text segments. If the classification is supporting or refuting, then the two-stage evidence selection system <b>120</b> outputs the classification and does not use the stage-2 text retrieval model <b>230</b>. If, however, the stage-1 text retrieval model <b>225</b> is unable to classify at least one first-stage text segment as either supporting the target text segment <b>205</b> or refuting the target text segment <b>205</b> (e.g., a classification of “not enough information” as indicated by arrow <b>345</b>), then the two-stage evidence selection system <b>120</b> uses stage-2 text retrieval model <b>230</b> to retrieve second-stage evidence (e.g., sentences that are indirectly relevant to the target text segment <b>205</b>), according to the lazy retrieval protocol.

0043In some implementations, the stage-2 text retrieval model <b>230</b> includes a second-stage BERT model <b>330</b>. For training the second-stage BERT model <b>330</b>, a dataset of target text segments is created, such that each target text segment requires multiple text segments (e.g., multiple sentences or phrases that serve as evidence) to support or refute the target text segment. Each target text segment included in the dataset is labeled as a positive example. Electronic documents of the text corpus from which the first-stage evidence was retrieved are labeled as negative samples. Training the stage-2 retrieval model <b>230</b> further includes sampling text segments that are not included in the evidence set (e.g., the text segments retrieved for the purpose of determining the veracity of a target text segment). As a non-limiting example, the input to the second-stage BERT model <b>230</b> (for the purpose of training the second-stage BERT model) is “‘[CLS]’+claim+page title_1+evidence_sentence_1+‘[SEP]’+page_title_2+‘:’+evidence sentence_2+‘[SEP]’”.

0044After the training of the stage-2 text retrieval model <b>230</b> is complete, if the lazy retrieval protocol proceeds to the conditional second-stage of evidence selection, then the electronic documents from which the first-stage text segments were extracted <b>335</b> are inputted into the stage-2 BERT model <b>330</b> (as indicated by arrow <b>330</b>). The stage-2 BERT model <b>330</b> generates, as an output, a relevance score for each inputted text segment. As a non-limiting example, the text segments associated with the top three relevance scores are selected for retrieval. These five first-stage text segments (in the case of first-stage only evidence retrieval) and eight second-stage text segments (if the second-stage evidence retrieval is triggered) are passed to the specialized reasoning model <b>125</b> for performing the first and second stage reasoning, respectively.

0045The present disclosure is not limited to the use of two trained BERT models; namely the first-stage BERT model <b>310</b> and the stage-2 BERT model <b>330</b>. Any number of trained BERT models may be implemented in as a multi-stage evidence selection system. The present disclosure is not limited to BERT models, and thus, other language models can be used.

0046<figref idref="DRAWINGS">FIG. 4</figref> depicts an example of a specialized reasoning model <b>125</b> configured to classify text segments as supporting or refuting a target text segment <b>405</b>, according to some aspects of the present disclosure. The specialized reasoning model <b>125</b> in this example is configured with two parts: a graph construction system <b>410</b> and a kernel node attention-based reasoning model <b>425</b>. The graph construction system <b>410</b> constructs a stage-1 evidence graph <b>415</b> for the first stage of reasoning and a stage-2 evidence graph <b>420</b> for second stage reasoning. The first stage of reasoning refers to the stage at which the specialized reasoning model <b>125</b> reasons over the first-stage evidence. The second stage of reasoning refers to the stage at which the specialized reasoning model <b>125</b> reasons over the first-stage evidence and the second-stage evidence.

0047For the first stage of reasoning, the graph construction system <b>410</b> generates evidence graph <b>415</b>. Evidence graph <b>415</b> includes a plurality of disconnected nodes. Each node represents the information associated with a text segment retrieved during the first-stage evidence retrieval. The nodes of evidence graph <b>415</b> are disconnected from each other, which means that during the first stage of reasoning when the kernel node attention-based reasoning model <b>425</b> reasons over the target text segment <b>405</b> and the first-stage evidence <b>335</b>. Thus, during the first stage of reasoning, the specialized reasoning model <b>125</b> does not collate information from multiple text segments when reasoning over the target text segment. The evidence graph <b>415</b> does not include any connections between nodes because each evidence text segment included in the first-stage evidence could potentially support or refute the target text segment <b>405</b>. Thus, each evidence text segment included in the first-stage evidence is selected so as to not need the context of another evidence text segment (e.g., another node) to support or refute the target text segment <b>405</b>.

0048During the second stage of reasoning (e.g., when the kernel node attention-based reasoning model <b>425</b> reasons over first-stage evidence and the second-stage evidence combined), the graph construction system <b>410</b> generates the evidence graph <b>420</b>, which is a partially connected set of nodes. The second-stage evidence includes evidence text segments that are indirectly relevant to the target text segment <b>405</b>. Thus, each evidence text segment included in the second-stage evidence needs the context of another evidence text segment to support or refute the target text segment <b>405</b>. Referring back to the “Ryan Gosling” example described above, the second-stage evidence included the sentence “Chad is a land-locked county in Central Africa,” which is indirectly relevant to the target text segment (“Ryan Gosling has been to a country in Africa”). In this example, the sentence “Chad is a land-locked country in Central Africa” needs the context of the first-stage evidence “Ryan Gosling has traveled to Chad, Uganda, and eastern Congo” to support or refute the target text segment.

0049Connecting two or more nodes together in evidence graph <b>420</b> enables the kernel node attention-based reasoning model <b>425</b> to collate information between the connected nodes during the second stage of reasoning (which jointly reasons over the first-stage evidence and the second-stage evidence).

0050Connecting all of the nodes together collates all of the first-stage evidence and the second-stage evidence together when the kernel node attention-based reasoning model <b>425</b> reasons over the evidence graph <b>420</b>. If all of the nodes are connected, however, the first-stage evidence and the second-stage evidence constitute all of the entities from the target text segment. Thus, kernel node attention-based reasoning model <b>425</b> will be more likely to classify the veracity of the target text segment as supported or refuted, even when the first-stage evidence and the second-stage evidence is insufficient to support or refute the target text segment. That is, when the kernel node attention-based reasoning model <b>425</b> reasons over a fully connected evidence graph, the entities included in each of the evidence text segments of the retrieved evidence are aggregated together, which overlooks the intricacies of individual combinations of evidence text segments and leads to the potential for misclassification as supported or refuted.

0051Certain implementations of the present disclosure address the potential for misclassification in a fully-connected graph by partially and selectively connecting nodes of the stage-2 evidence graph <b>420</b>, rather than connecting all of the nodes together. In some implementations, determining which nodes to connect together, the graph construction system <b>410</b> identifies entities present in each node (e.g., each retrieved first-stage and second-stage text segment) of the stage-2 evidence graph <b>420</b>, and then connects the nodes that have common entities between them. As a technical improvement, evidence graph <b>420</b> is partially connected, in that two or more nodes of the set of nodes are connected to each other, but not all nodes are connected to another node. Certain nodes may still be disconnected, which provides a technical improvement by reducing the processing time needed for the kernel node attention-based reasoning model <b>425</b> to perform the second stage reasoning.

0052Referring back to the “Ryan Gosling” example, the graph construction system <b>410</b> identifies that the first-stage evidence of “Ryan Gosling has traveled to Chad, Uganda, and eastern Congo” has the entity “Chad,” and the second-stage evidence of “Chad is a land-locked country in Central Africa” also has the entity “Chad.” The graph construction system <b>410</b> connects the node representing the first-stage evidence of “Ryan Gosling has traveled to Chad, Uganda, and eastern Congo” with the node representing the second-stage evidence of “Chad is a land-locked country in Central Africa.” Thus, when the kernel node attention-based reasoning model <b>425</b> reasons over the partially connected evidence graph in this example, the kernel node attention-based reasoning model <b>425</b> will infer that the target text segment of “Ryan Gosling has been to a country in Africa” is supported by the combined first-stage and second-stage evidence.

0053During the first stage of reasoning, the kernel node attention-based reasoning model <b>425</b> performs a reasoning task over the stage-1 evidence graph <b>415</b> and the target text segment <b>405</b>. During the second stage of reasoning, the kernel node attention-based reasoning model <b>425</b> performs the reasoning task over the partially connected stage-2 evidence graph <b>420</b>. Further, the kernel node attention-based reasoning model <b>425</b> generates as an output a verification indicator <b>430</b>, which represents the classification of the first-stage text segments or the joint classification of the first-stage and second-stage text segments.

0054The kernel node attention-based reasoning model <b>425</b> executes a reasoning task when reasoning over evidence graph <b>415</b> or evidence graph <b>420</b>. Executing the reasoning task includes executing the following equations. The reasoning task involves building sub-trees of an evidence graph (e.g., evidence graph <b>415</b> or <b>420</b>), where each node of the evidence graph is a root of the sub-tree. For instance, for a given node p of the evidence graph, the kernel node attention-based reasoning model <b>425</b> extracts one-hop nodes having a direct edge to that node p. The kernel node attention-based reasoning model <b>425</b> forms a sub-tree G^pfrom the node p and the extracted one-hop nodes. The kernel node attention-based reasoning model <b>425</b> generates a node encoding n^p, which numerically represents the evidence text segment associated with node p. n^pis generated by passing the input “[CLS]+claim+[SEP]+evidence text segment associated with node p+[SEP]” through the trained BERT model (e.g., the first-stage BERT model for first-stage reasoning and the second-stage BERT model for second-stage reasoning). The output is used as the value for n^p.

0055The kernel node attention-based reasoning model <b>425</b> then generates an information propagation term, z^q→p, by computing an attention-based weighted average over the nodes, n^q, that are one-hop away from node p. Then, the kernel node attention-based reasoning model <b>425</b> can generate a sub-tree representation, a^p, to numerically represent each sub-tree G. The following equation is used to generate the sub-tree representation a^p:

0000<br /><i>a</i>^p=Σ_q,p∈ξβ^q→p<i>*z</i>^q→p,  (Equation 1),

0000where ξ represents the set of edges in the evidence graph, z^q→prepresents the information that is propagated from n^qto n^p, and β^q→pis the attention weight of n^qwith respect to n^p.

0056The kernel node attention-based reasoning model <b>425</b> reasons over the evidence graph by identifying a current node of the evidence graph and generating an updated representation for the current node, v^p, by concatenating a^pwith the initial root representation, z^p, using the following equation:

0000<br /><i>v</i>^p<i>=a</i>^p<i>∘z</i>^p,  (Equation 2),

0000where “[CLS]” is used as the initial root representation z^p.

0057Using the updated representation, v^p, for each node in the evidence graph, the kernel node attention-based reasoning model <b>425</b> computes a probability distribution P(y|n^p,G^p) over various labels y given G^pand n^p. For instance, the kernel node attention-based reasoning model <b>425</b> computes P(y|n^p,G^p) as

0000<br /><i>P</i>(<i>y|n</i>^p<i>,G</i>^p)=softmax_y(Linear(<i>v</i>^p)),  (Equation 3)

0058The kernel node attention-based reasoning model <b>425</b> generates a composite representation, ϕ(n^p), which is used determine the importance of node p as compared to other nodes, such as node q, of the evidence graph. To illustrate, for the target text element “Stanley Tucci performed in a television series,” the first-stage evidence of “Monk is an American comedy-drama detective mystery television series” may be ranked low if only the updated representation, v^pis considered. However, the composite representation, ϕ(n^p), contains additional information about node p's relationship with another node q, “He won two Emmy Awards for his performances in Winchell and Monk”, which is used to interpret the importance of node p. The composite representation is calculated using the following equation

0000<br />ϕ(<i>n</i>^p)=<i>n</i>^p<i>∘a</i>^p,  (Equation 4)

0059The kernel node attention-based reasoning model <b>425</b> assigns an importance order to the nodes in the presence of the context information of the current node with respect to its connected nodes, as opposed to taking the context solely from the target text segment. For instance, the kernel node attention-based reasoning model <b>425</b> calculates, with the following equation, a probability distribution P(n^p|G) of a particular node n^pfor a tree G:

0000<br /><i>P</i>(<i>n</i>^p<i>|G</i>)=softmax_y(<i>MLP</i>(ϕ(<i>n</i>^p))),  (Equation 5)

0000The kernel node attention-based reasoning model <b>425</b> determines the importance order from the probability distribution P(n^p|G) by ranking the resulting probability distribution for each node p of the evidence graph.

0060The kernel node attention-based reasoning model <b>425</b> also performs a joint graph reasoning of the first-stage evidence and the second-stage evidence. For example, the specialized reasoning model <b>125</b> generates a combined label prediction probability P for a particular label y as given below,

0000<br /><i>P</i>(<i>y|G</i>)=Σ_p=1^l<i>p</i>(<i>y|n</i>^p<i>,G</i>^p)<i>P</i>(<i>n</i>^p<i>|G</i>),  (Equation 6)

0061Equations 1 through 6 are calculated in order to determine a probability of selecting each node in an evidence graph. Selecting a node refers to selecting an evidence text segment from amongst the top five evidence text segments retrieved during the first stage of retrieval or from amongst the top eight during the second stage of retrieval. Further, text verification system <b>110</b> selects the node by minimizing a cross entropy loss (e.g., the node associated with the minimum cross entropy loss is selected as the evidence text segments that supports or refutes the target text segment). The kernel node attention-based reasoning model <b>425</b> calculates the cross entropy loss associated with each node p, using the following equation:

0000<br /><img file="US2022171935A1_D0001.tif" />=CrossEntropy(<i>y*P</i>(<i>y|G</i>))+BinaryCrossEntropy(<i>n*,p</i>*),  (Equation 7)

0062In Equation 7, a cross entropy loss CrossEntropy indicates a difference between a predicted label y and a ground truth label y*, where y* is the one-hot three-dimensional vector for the ground truth label and P(y|G) is the predicted probability over the three classes (e.g., supported, refuted, and “not enough information”). Furthermore, a binary cross entropy loss BinaryCrossEntropy indicates a difference between the predicted evidence p* and the ground truth evidence n*. The ground truth evidence term n* is a multi-dimensional vector (e.g., eight dimensions) with elements {0,1}, where 1 represents a true evidence, and 0 otherwise. The multi-dimensional vector can have multiple elements representing a true evidence. As an illustrative example, both “He won two Emmy Awards for his performances in Winchell and Monk” and “Monk is an American comedy-drama detective mystery television series” will have 1 in the ground truth multi-dimensional vector. The predicted evidence p* is computed as:

0000<br /><i>p*={P</i>(<i>n</i>_i<i>*|G</i>)|<i>i∈{</i>1,8}},  (Equation 8)

0063In Equation 8, P(n_i|G) is the probability of node n_ibeing an evidence node. The binary cross entropy loss term in the loss function <img file="US2022171935A1_D0002.tif" /> provides a feedback on correct evidence, which improves the importance probability hence increasing the weight of label prediction during combining predictions from all nodes, and consequently enhances the label accuracy.

0064<figref idref="DRAWINGS">FIG. 5</figref> depicts an example <b>500</b> of using a text verification system to generate a verification indicator representing the inferred veracity of a target text segment, according to some aspects of the present disclosure. The target text segment <b>505</b> is “Stanley Tucci performed in a mystery television series.” That is, a user operates the electronic document reader <b>105</b> by selecting the target text segment <b>505</b> displayed on a target electronic document and triggering the evidence retrieval and verification functionality provided by the electronic document reader <b>105</b>.

0065The document retrieval system <b>115</b> receives the target text segment <b>505</b> and performed an entity-based search to identify the potential entities described in the target text segment <b>505</b>. The document retrieval system <b>115</b> performs an index-based search of all of the electronic documents stored in electronic document database <b>130</b> to identify selected documents <b>510</b>, from which the evidence will be retrieved. The document retrieval system <b>115</b> determines the selected documents <b>510</b> as being relevant to the target text segment <b>505</b>. The document retrieval system <b>115</b> transmits the selected documents <b>510</b> to the two-stage evidence selection system <b>120</b> (as indicated by arrow <b>525</b>).

0066During the first stage of evidence retrieval, the first-stage BERT model of the two-stage evidence selection system <b>120</b> individually receives as input each text segment from the selected documents <b>510</b>. The first-stage BERT model generates a relevance score for each inputted text segment. The text segment <b>515</b> is selected for first-stage evidence retrieval because the text segment had the highest relevance score, as outputted by the first-stage BERT model of the two-stage evidence selection system <b>120</b>. For example, the text segment <b>515</b> includes the text “Stanley Tucci won two Emmy Awards for his performances in Winchell and Monk.” The two-stage evidence selection system transmits the text segment <b>515</b> to the specialized reasoning model <b>125</b> to perform the reasoning task (as indicated by arrow <b>530</b>).

0067The specialized reasoning model <b>125</b> generates a stage-1 evidence graph for the text segment <b>515</b> and any other text segments that were retrieved during the first stage of evidence retrieval. The stage-1 evidence graph is a disconnected graph, in that none of the nodes of the evidence graph is connected to another node. One of the nodes of the stage-1 evidence graph represents the text segment <b>515</b>. The kernel node attention-based reasoning model of the specialized reasoning model <b>125</b> evaluates the stage-1 evidence graph and the target text segment <b>505</b> to classify the text segment <b>515</b> as supporting the target text segment <b>505</b>, refuting the target text segment <b>505</b>, or not providing sufficient information to support or refute the target text segment <b>505</b>. While text segment <b>515</b> uses the term “Stanley Tucci” and “Monk,” text segment <b>515</b> on its own does not support or refute the target text segment <b>505</b> because text segment <b>515</b> does not describe the television show “Monk” as being a “mystery television series.” Thus, the specialized reasoning model <b>125</b> classifies the text segment <b>515</b> as not providing sufficient information to support or refute the target text segment <b>505</b> (as indicated by arrow <b>535</b>).

0068Given that the specialized reasoning model <b>125</b> was unable to classify the text segment <b>515</b> as supporting the target text segment <b>505</b> or refuting the target text segment <b>505</b>, the two-stage evidence selection system <b>120</b> proceeds to the conditional second-stage evidence retrieval due to the lazy retrieval protocol. Extracting the second-stage text segments based on the first-stage text segments includes searching through the set of electronic documents using the first-stage text segments as a target and identifying electronic documents that may be relevant. The text segments from these documents are inputted into the second-stage BERT model to generate a relevance score used to select the second-stage text segment. The second-stage BERT model generates a relevance score for each inputted text segment. The text segment <b>520</b> is identified as being associated with the highest relevance score as outputted by the second-stage BERT model, and thus, text segment <b>520</b> is retrieved during the second stage of evidence retrieval. The text segment <b>520</b> is a sentence that states “Monk is an American comedy-drama detective mystery television series.” The two-stage evidence selection system <b>120</b> transmits the text segment <b>520</b> to the specialized reasoning model <b>125</b> to be jointly reasoned over together with text segment <b>515</b>.

0069The specialized reasoning model <b>125</b> generates a stage-2 evidence graph representing both the text segment <b>515</b> and the text segment <b>520</b>. The stage-2 evidence graph is a partially connected graph, such that at least two nodes are connected to collate the information from the two nodes. The specialized reasoning model <b>125</b> determines that the node representing the text segment <b>515</b> is to be connected with the node representing the text segment <b>520</b> because both nodes include the same entity (e.g., the entity of “Monk.”). The kernel node attention-based reasoning model reasons over the stage-2 evidence graph and the target text segment to infer whether the text segment <b>515</b> and the text segment <b>520</b> together are classified as supporting or refuting the target text segment <b>505</b>. The kernel node attention-based reasoning model automatically determines that the text segment <b>515</b> together with the text segment <b>520</b> is classified as supporting the target text segment. For example, the fact that Stanley Tucci won an Emmy for his performance in Monk, combined with the fact that Monk is a mystery television series supports the target text segment <b>505</b> that read “Stanley Tucci performed in a mystery television series.” Therefore, the specialized reasoning model <b>125</b> generates a verification indicator <b>545</b>, indicating that the veracity of the target text segment <b>505</b> has been verified.

0070<figref idref="DRAWINGS">FIG. 6</figref> depicts an example of a process <b>600</b> for using a text verification system to automatically infer the veracity of a target text segment, according to some aspects of the present disclosure. Process <b>600</b> is performed at least in part by any of the hardware-based computing devices illustrated in <figref idref="DRAWINGS">FIGS. 1-5</figref> or <figref idref="DRAWINGS">FIGS. 8-9</figref>. For example, process <b>600</b> is performed by one or more servers included in the document display system <b>100</b>, the electronic document reader <b>105</b>, or the text verification system <b>110</b>. As a further example, the text verification system <b>110</b> performs process <b>600</b> as part of fact extraction and verification functionality, such as determining the veracity of a target text segment selected based on input received from a user device.

0071At block <b>605</b>, the text verification system <b>110</b> receives input (e.g., from user system <b>135</b>) corresponding to a selection of a target text segment. The target text segment is included in a target electronic document that is displayed by electronic document reader <b>105</b>. For example, the target text segment is selected if a user who is viewing the target electronic document seeks to determine the veracity of the target text segment. The target text segment is any text displayed by the electronic document reader <b>105</b>. An example of implementing block <b>605</b> is described above with respect to <figref idref="DRAWINGS">FIG. 1</figref>.

0072At block <b>610</b>, the two-stage evidence selection system <b>120</b> performs a first stage of evidence retrieval, which includes retrieving first-stage evidence from the text corpus stored in electronic document database <b>130</b>. The first-stage evidence refers to evidence text segments that are directly relevant to the target text segment. The two-stage evidence selection system <b>120</b> generates a relevance score for each text segment included in the text corpus. For example, the relevance score represents the relevance between the text segment that is being evaluated for retrieval and the target text segment. In some implementations, the text segments that are associated with the top five relevance scores are retrieved as first-stage evidence. An example of implementing block <b>610</b> is described above with respect to <figref idref="DRAWINGS">FIGS. 1, 2, and 3</figref>.

0073At block <b>615</b>, the specialized reasoning model <b>125</b> classifies the veracity of the target text segment using the first-stage evidence retrieved at block <b>610</b> as supported by the first-stage evidence, refuted by the first-stage evidence, or undetermined. In some implementations, the specialized reasoning model <b>125</b> includes constructing an evidence graph, which graphically represents the first-stage evidence retrieved at block <b>610</b>. The evidence graph includes a plurality of nodes. Each node represents, for example, one of the five text segments included in the first-stage evidence retrieved at block <b>610</b>. Further, each node of the plurality of nodes is disconnected from any other node in the evidence graph during the first stage of reasoning. The evidence graph is disconnected at this first stage because the specialized reasoning model <b>125</b> determines whether each text segment of the first-stage evidence individually provides direct support (or not) for the target text segment. Additionally, the specialized reasoning model includes a kernel node attention-based reasoning model that is trained to classify the veracity of the target text segment. An example of implementing block <b>615</b> is described above with respect to <figref idref="DRAWINGS">FIGS. 4-5</figref>.

0074At block <b>620</b>, the specialized reasoning model <b>125</b> classifies the veracity of the target text segment as undetermined. In response, the two-stage evidence selection system <b>120</b> proceeds to the second stage of evidence retrieval according to the lazy retrieval protocol. The second stage of evidence retrieval identifies text segments in the text corpus that are indirectly relevant to the target text segment. Indirect relevance indicates that the specialized reasoning model <b>125</b> infers that the second-stage evidence supports or refutes the target text segment only with the context of the first-stage evidence. An example of implementing block <b>620</b> is described above with respect to <figref idref="DRAWINGS">FIGS. 4-5</figref>.

0075At block <b>625</b>, the specialized reasoning model <b>125</b> evaluates the target text segment using the first-stage evidence and the second-stage evidence. The second-stage evidence is retrieved based on the first-stage evidence because the first-stage evidence alone does not provide sufficient information to enable the specialized reasoning model <b>125</b> to classify the target text segment as “supported” or refuted.” The specialized reasoning model <b>125</b> modifies the evidence graph by selectively connecting two or more nodes in the evidence graph. In some implementations, the two-stage evidence selection system <b>120</b> identifies the entities present in each retrieved text segment, and then connects the nodes that have common entities between them. The connected nodes are used to learn and update the representation of each node in the evidence graph. The output of the specialized reasoning model at stage two is a classification of whether the first and second text segments together support or refute the target text segment. An example of implementing block <b>625</b> is described above with respect to <figref idref="DRAWINGS">FIGS. 4-5</figref>.

0076At block <b>630</b>, the text verification system <b>110</b> generates a verification indicator representing the classification of block <b>620</b>. The text verification system <b>110</b> can augment the target electronic document by adding the output of the specialized reasoning model (e.g., “supported” or “refuted”) to the target electronic document adjacent or otherwise in association with the displayed target text. <figref idref="DRAWINGS">FIG. 7</figref>, for example, illustrates displaying the target text segment in association with the classification outputted by the specialized reasoning model. An example of implementing block <b>630</b> is described below with respect to <figref idref="DRAWINGS">FIG. 7</figref>.

0077<figref idref="DRAWINGS">FIG. 7</figref> depicts an example of an output of process <b>600</b> as described with respect to <figref idref="DRAWINGS">FIG. 6</figref>. Electronic reading software (e.g., ADOBE READER) displays electronic documents on a display screen of a user device. The electronic reading software is further configured to include a text verification system (e.g., text verification system <b>110</b> as illustrated in <figref idref="DRAWINGS">FIG. 1</figref>) that is usable to augment a displayed electronic document with verification indicators. For example, a verification indicator represents an inferred veracity of a target text segment included in the electronic document.

0078To illustrate and only as a non-limiting example, the electronic reading software receives input (e.g., from a user device) corresponding to a selection of a target text segment, for which the veracity is to be determined. As illustrated in <figref idref="DRAWINGS">FIG. 7</figref>, target text segment <b>705</b> is selected based on input received from a user device. The electronic reading software receives another input, instructing the electronic reading software to execute the text verification system infer the veracity of target text segment <b>705</b> using the content of electronic documents stored in electronic document database <b>130</b>. In response, the text verification system automatically retrieves various text segments from the content of the electronic documents stored in electronic documents database <b>130</b> and determines the veracity metric of target text segment <b>705</b> based on the retrieved text segments. The text verification system classifies target text segment <b>705</b> as being supported or refuted and augments target electronic document <b>700</b> with verification indicator <b>710</b>. For example, as illustrated in <figref idref="DRAWINGS">FIG. 7</figref>, verification indicator <b>710</b> indicates that the text verification system classified target text segment <b>705</b> as “refuted” (e.g., “false” based on the content included in the electronic documents stored in electronic documents database <b>130</b>).

0079Additionally, electronic reading software receives input corresponding to a selection of target text segment <b>715</b>. The electronic reading software receives an additional input, instructing the electronic reading software to execute the text verification system to infer the veracity of text segment <b>715</b> using the electronic documents stored in electronic document database <b>130</b>. In response, the text verification system determines the veracity of target text segment <b>715</b> and augments target electronic document <b>700</b> with verification indicator <b>720</b>. For example, as illustrated in <figref idref="DRAWINGS">FIG. 7</figref>, verification indicator <b>720</b> indicates that the text verification system classified target text segment <b>715</b> as “supported” (e.g., “true” based on the content included in the electronic documents stored in electronic documents database <b>130</b>).

0080Examples of Computing Environments for Implementing Certain Implementations

0081Any suitable computing system or group of computing systems can be used for performing the operations described herein. For example, <figref idref="DRAWINGS">FIG. 9</figref> depicts an example of computing device <b>900</b> that may be at least a portion of document display system <b>100</b>. The implementation of the computing device <b>900</b> could be used for one or more of the electronic document reader <b>105</b> or the user system <b>135</b>. In an implementation, a single document display system <b>100</b> having devices similar to those depicted in <figref idref="DRAWINGS">FIG. 9</figref> (e.g., a processor, a memory, etc.) combines the one or more operations and data stores depicted as separate subsystems in <figref idref="DRAWINGS">FIG. 1</figref>. Further, <figref idref="DRAWINGS">FIG. 8</figref> illustrates a cloud computing system <b>800</b> by which at least a portion of the document display system <b>100</b> may be offered.

0082In some implementations, the functionality provided by the document display system <b>100</b> may be offered as cloud services by a cloud service provider. For example, <figref idref="DRAWINGS">FIG. 8</figref> depicts an example of a cloud computing system <b>800</b> offering an image editing service that can be used by a number of user subscribers using user devices <b>825</b><i>a</i>, <b>825</b><i>b</i>, and <b>25</b><i>c </i>across a data network <b>820</b>. In the example, the image editing service may be offered under a Software as a Service (SaaS) model. One or more users may subscribe to the image editing service, and the cloud computing system performs the processing to provide the image editing service to subscribers. The cloud computing system may include one or more remote server computers <b>805</b>.

0083The remote server computers <b>805</b> include any suitable non-transitory computer-readable medium for storing program code (e.g., a document display system <b>100</b>) and program data <b>810</b>, or both, which is used by the cloud computing system <b>800</b> for providing the cloud services. A computer-readable medium can include any electronic, optical, magnetic, or other storage device capable of providing a processor with computer-readable instructions or other program code. Non-limiting examples of a computer-readable medium include a magnetic disk, a memory chip, a ROM, a RAM, an ASIC, optical storage, magnetic tape or other magnetic storage, or any other medium from which a processing device can read instructions. The instructions may include processor-specific instructions generated by a compiler or an interpreter from code written in any suitable computer-programming language, including, for example, C, C++, C#, Visual Basic, Java, Python, Perl, JavaScript, and ActionScript. In various examples, the server computers <b>805</b> can include volatile memory, non-volatile memory, or a combination thereof.

0084One or more of the servers <b>805</b> execute the program code <b>810</b> that configures one or more processors of the server computers <b>805</b> to perform one or more of the operations that provide image editing services, including the ability to utilize the text verification system <b>110</b> to perform shot-matching and other image editing techniques. As depicted in the implementation in <figref idref="DRAWINGS">FIG. 8</figref>, the one or more servers providing the services to perform image editing via the text verification system <b>110</b> may include access to the models of the text verification system <b>110</b> including the document retrieval system <b>115</b>, the two-stage evidence selection system <b>120</b>, and the specialized reasoning model <b>125</b>. Any other suitable systems or subsystems that perform one or more operations described herein (e.g., one or more development systems for configuring an interactive user interface) can also be implemented by the cloud computing system <b>800</b>.

0085In certain implementations, the cloud computing system <b>800</b> may implement the services by executing program code and/or using program data <b>810</b>, which may be resident in a memory device of the server computers <b>805</b> or any suitable computer-readable medium and may be executed by the processors of the server computers <b>805</b> or any other suitable processor.

0086In some implementations, the program data <b>810</b> includes one or more datasets and models described herein. Examples of these datasets include image data, new image content, image energy data, etc. In some implementations, one or more of data sets, models, and functions are stored in the same memory device. In additional or alternative implementations, one or more of the programs, data sets, models, and functions described herein are stored in different memory devices accessible via the data network <b>815</b>.

0087The cloud computing system <b>800</b> also includes a network interface device <b>815</b> that enable communications to and from cloud computing system <b>800</b>. In certain implementations, the network interface device <b>815</b> includes any device or group of devices suitable for establishing a wired or wireless data connection to the data networks <b>820</b>. Non-limiting examples of the network interface device <b>815</b> include an Ethernet network adapter, a modem, and/or the like. The document display system <b>100</b> is able to communicate with the user devices <b>825</b><i>a</i>, <b>825</b><i>b</i>, and <b>825</b><i>c </i>via the data network <b>820</b> using the network interface device <b>815</b>.

0088<figref idref="DRAWINGS">FIG. 9</figref> illustrates a block diagram of an example computer system <b>900</b>. Computer system <b>900</b> can be any of the described computers herein including, for example, electronic document reader <b>105</b>, user system <b>135</b>, or server computer <b>805</b>. The computing device <b>900</b> can be or include, for example, a laptop computer, desktop computer, tablet, server, or other electronic device.

0089The computing device <b>900</b> can include a processor <b>935</b> interfaced with other hardware via a bus <b>905</b>. A memory <b>910</b>, which can include any suitable tangible (and non-transitory) computer readable medium, such as RAM, ROM, EEPROM, or the like, can embody program components (e.g., program code <b>915</b>) that configure operation of the computing device <b>800</b>. Memory <b>910</b> can store the program code <b>915</b>, program data <b>917</b>, or both. In some examples, the computing device <b>900</b> can include input/output (“I/O”) interface components <b>925</b> (e.g., for interfacing with a display <b>940</b>, keyboard, mouse, and the like) and additional storage <b>930</b>.

0090The computing device <b>900</b> executes program code <b>915</b> that configures the processor <b>935</b> to perform one or more of the operations described herein. Examples of the program code <b>915</b> include, in various implementations, the text verification system <b>110</b> including the document retrieval system <b>115</b>, the two-stage evidence selection system <b>120</b>, and the specialized reasoning model <b>125</b>, the image editing function, or any other suitable systems or subsystems that perform one or more operations described herein (e.g., one or more development systems for configuring an interactive user interface). The program code <b>915</b> may be resident in the memory <b>910</b> or any suitable computer-readable medium and may be executed by the processor <b>940</b> or any other suitable processor.

0091The computing device <b>900</b> may generate or receive program data <b>917</b> by virtue of executing the program code <b>915</b>. For example, the source image and modified source image are all examples of program data <b>917</b> that may be used by the computing device <b>900</b> during execution of the program code <b>915</b>.

0092The computing device <b>900</b> can include network components <b>920</b>. Network components <b>920</b> can represent one or more of any components that facilitate a network connection. In some examples, the network components <b>920</b> can facilitate a wireless connection and include wireless interfaces such as IEEE 802.11, Bluetooth, or radio interfaces for accessing cellular telephone networks (e.g., a transceiver/antenna for accessing CDMA, GSM, UMTS, or other mobile communications network). In other examples, the network components <b>920</b> can be wired and can include interfaces such as Ethernet, USB, or IEEE 1394.

0093Although <figref idref="DRAWINGS">FIG. 9</figref> depicts a single computing device <b>900</b> with a single processor <b>935</b>, the system can include any number of computing devices <b>900</b> and any number of processors <b>935</b>. For example, multiple computing devices <b>900</b> or multiple processors <b>935</b> can be distributed over a wired or wireless network (e.g., a Wide Area Network, Local Area Network, or the Internet). The multiple computing devices <b>900</b> or multiple processors <b>935</b> can perform any of the steps of the present disclosure individually or in coordination with one another.

0094Numerous specific details are set forth herein to provide a thorough understanding of the claimed subject matter. However, those skilled in the art will understand that the claimed subject matter may be practiced without these specific details. In other instances, methods, apparatuses, or systems that would be known by one of ordinary skill have not been described in detail so as not to obscure claimed subject matter.

0095Unless specifically stated otherwise, it is appreciated that throughout this specification discussions utilizing terms such as “processing,” “computing,” “calculating,” “determining,” and “identifying” or the like refer to actions or processes of a computing device, such as one or more computers or a similar electronic computing device or devices, that manipulate or transform data represented as physical electronic or magnetic quantities within memories, registers, or other information storage devices, transmission devices, or display devices of the computing platform.

0096The system or systems discussed herein are not limited to any particular hardware architecture or configuration. A computing device can include any suitable arrangement of components that provide a result conditioned on one or more inputs. Suitable computing devices include multi-purpose microprocessor-based computer systems accessing stored software that programs or configures the computing system from a general purpose computing apparatus to a specialized computing apparatus implementing one or more implementations of the present subject matter. Any suitable programming, scripting, or other type of language or combinations of languages may be used to implement the teachings contained herein in software to be used in programming or configuring a computing device.

0097Embodiments of the methods disclosed herein may be performed in the operation of such computing devices. The order of the blocks presented in the examples above can be varied—for example, blocks can be re-ordered, combined, and/or broken into sub-blocks. Certain blocks or processes can be performed in parallel.

0098The use of “adapted to” or “configured to” herein is meant as open and inclusive language that does not foreclose devices adapted to or configured to perform additional tasks or steps. Additionally, the use of “based on” is meant to be open and inclusive, in that a process, step, calculation, or other action “based on” one or more recited conditions or values may, in practice, be based on additional conditions or values beyond those recited. Headings, lists, and numbering included herein are for ease of explanation only and are not meant to be limiting.

0099While the present subject matter has been described in detail with respect to specific implementations thereof, it will be appreciated that those skilled in the art, upon attaining an understanding of the foregoing, may readily produce alterations to, variations of, and equivalents to such implementations. Accordingly, it should be understood that the present disclosure has been presented for purposes of example rather than limitation, and does not preclude the inclusion of such modifications, variations, and/or additions to the present subject matter as would be readily apparent to one of ordinary skill in the art.