Computer implementation methods, computer systems, and programs

Abstract

A machine learning method for a computer implementation of aspect-based summarization. It extracts aspect candidates from a document. These aspect candidates are filtered by a deambiguation technique using different aspect granularity levels. Informational content is extracted based on the filtered aspect candidates. Summarization techniques are implemented to generate summaries of multiple document-aspect pairs from the informational content. The aspect-based summary is generated based on the summaries. Use cases for this method include, but are not limited to, medical / healthcare, public safety, cyber threat intelligence, and other artificial intelligence applications for document summarization.

Description

Technical Field

[0001] Cross-reference to Prior Applications Priority is claimed to U.S. Provisional Application No. 63 / 522,447, filed Jun. 22, 2023, the entire content of which is incorporated herein by reference.

[0002] This disclosure relates to Artificial Intelligence (AI) and machine learning, and more particularly, to methods, systems, and computer-readable media for the automated discovery of aspect-based summaries.

Background Art

[0003] Summarization is a common task in natural language processing (NLP) that produces a concise and consistent summary from a longer document or set of documents. The purpose of summarization is to extract the most important information and key points while maintaining the overall meaning and context of the original text. In this way, automatic text summarization helps humans save time when faced with large amounts of text.

[0004] NLP summarization can utilize, for example, large language models (LLMs) such as ChatGPT. LLMs are advanced deep learning models trained on vast amounts of text data. LLMs have the ability to understand and generate human-like language. These models utilize a transformer architecture, which enables them to process and generate text by capturing long-range dependencies within the input while considering context-dependence.

Summary of the Invention

[0005] In one embodiment, the present invention provides a computer-implemented machine learning method for aspect-based summarization. Aspect candidates are extracted from a document. The aspect candidates are filtered by a deambiguation technique using different aspect granularity levels. Informational content is extracted based on the filtered aspect candidates. Summarization techniques are implemented to generate summaries of multiple document-aspect pairs from the informational content. Aspect-based summaries are generated based on these summaries. Use cases for this method include, but are not limited to, medical / healthcare, public safety, cyber threat intelligence, and other artificial intelligence applications for document summarization.

[0006] Embodiments of this disclosure are described in further detail below with reference to exemplary drawings. However, this disclosure is not limited to exemplary embodiments. All features described and / or illustrated herein may be used individually or in different combinations in embodiments of this disclosure. The features and advantages of various embodiments of this disclosure will become apparent by reading the following detailed description with reference to the accompanying drawings shown below. [Brief explanation of the drawing]

[0007] [Figure 1] This figure shows a comparison between a novel aspect-based summarization system and method according to one embodiment of the present disclosure and a conventional summarization method.

[0008] [Figure 2] This figure schematically illustrates a method and system architecture according to one embodiment of the present disclosure.

[0009] [Figure 3] This is a flowchart of an aspect-based summarization method according to one embodiment of the present invention.

[0010] [Figure 4]A block diagram of an exemplary processing system that can be configured to perform any and all of the operations disclosed herein is shown. [Modes for carrying out the invention]

[0011] Embodiments of this disclosure aim to provide improved NLP summarization, including providing automated discovery of aspect-based summaries.

[0012] According to existing technologies, state-of-the-art NLP summarization techniques may have the problem of providing statically aggregated summaries based on multiple documents. To mitigate this static problem, aspect-based summarization generates different summaries based on a fixed (e.g., predefined) list of topics, such as sections or subsections that discuss specific goals, such as scientific methods (also called aspects). In embodiments, aspects may refer to any topic of interest. For example, a document such as a news article or a medical report may contain multiple topics / aspects. However, this limits the technical capabilities when a document does not have a clear division of topics or spans multiple resources in different formats. This also limits the ability to meet user needs, for example, when different user roles may consider different importance of aspects while comparing multiple documents. In fact, even a single user may have different summarization needs and may depend on the current task at hand. Furthermore, by utilizing generated aspect-based summarization to create contrasting summaries that highlight the differences and similarities between inputs, the time and effort required to compare many documents can be reduced. While state-of-the-art NLP summarization techniques rely on predefined templates for tasks, embodiments of this disclosure enable automated aspect discovery and facilitate application to documents that conform to various domains and formats.

[0013] Accordingly, embodiments of the present disclosure enable the automatic discovery of different topics and the adjustment of summaries according to user-specified needs and the required level of granularity. This allows users to quickly identify topics, similarities, and differences between documents without having to read extensive documents, thus saving time and improving efficiency. Accordingly, embodiments of the present disclosure simplify decision-making processes and streamline work processes.

[0014] Embodiments of this disclosure provide the following: 1) Relevance: Automatic aspect discovery helps select the most relevant aspects common across multiple documents. This ensures that the summary focuses on the most important aspects and provides a clear and concise summary of the main topics. Furthermore, it helps improve the accuracy, reliability, and credibility of the AI ​​system. Without requiring user input to select aspects of interest from input documents, the system can register different segments / aspects in the summary using a combination of automated methods to extract headers, titles, and subtitles, and can use a trained model to extract additional aspects based on the topics of discussion in different paragraphs, which may or may not match the headers. 2) Efficiency: By automatically filtering aspects, the process of summarizing across multiple documents becomes more efficient. This significantly reduces the time and effort required to process and summarize information across multiple documents, helping to lower the workload and streamline the process. This also speeds up calculations, saves computing resources, and / or reduces computation time and / or power consumption. 3) Customization: Dynamic aspect selection enables advanced customization in generative summarization. This allows summaries to be personalized based on the user's specific preferences and domain interests, increasing the relevance of the summary to the user's needs and serving those needs better.

[0015] A first embodiment of this disclosure provides a method for generating aspect-based summaries of multiple documents in a dynamic and customizable manner. The method includes: 1) Select and implement a document search technique to find relevant documents. 2) Select and implement an aspect candidate extraction technique. 3) Implement aspect ambiguity resolution technology that takes into account different aspect granularity levels. 4) Extract informational content based on the filtered aspects. In this step, the most relevant parts for each aspect are extracted directly from the original text. For example, if the aspect is methodology in scientific publications, the result will be the extracted paragraphs from papers corresponding to this topic. 5) Select and implement either an extracted abstraction or an abstraction summarization technique. 6) Select a summarization technique for generating inter-document aspect-based summaries. 7) Create a user interface that links the four steps (A, B, C, and D described below). The user interface will allow the user to select an initial document to summarize. After selecting and uploading the document, the user will be prompted to enter any aspects they particularly want to focus on (e.g., a summary of technical methods). Once the user enters their preferences, relevant documents will be searched and several different aspects will be extracted, including (if applicable) the aspects specified by the user. The aspects will then be filtered to prevent any repetition and to resolve any ambiguity, and the user will be shown a final list of aspects. In the backend, the model will generate a summary for each aspect and display the final output to the user, shown in separate components, allowing the user to find summaries corresponding to the uploaded and searched documents simply by clicking on the aspects of interest.

[0016] In a second embodiment, a non-temporary computer-readable medium is provided that, when executed by one or more processors, stores instructions for performing the method of the first embodiment.

[0017] In a third embodiment, a system is provided that includes one or more processors configured to perform the method of the first embodiment alone or in combination.

[0018] Embodiments of the present disclosure provide several advantages compared to the current state of the art. These advantages include the following. 1. Existing work depends on a predefined list of topics, but embodiments of the present disclosure have an automatic discovery step, making the embodiments according to the present disclosure more flexible and adaptable to the needs of users, thereby generating better summaries for each user query. 2. Based on various discovered aspects, embodiments according to the present disclosure can more efficiently contrast various documents by highlighting similarities and differences, thus reducing the workload and rationalizing the process. 3. By dynamically selecting aspects, it becomes possible to create different summaries tailored to the needs of the user. 4. Embodiments according to the present disclosure are not domain-specific because aspects are automatically detected, and the embodiments can be used for contracts, patents, resumes, and case reports. In contracts, previous work has focused on domain-specific use cases such as nutritional reports. 5. Embodiments of the present disclosure can use a combination of machine learning and rule-based methods to extract aspects, thereby enabling the detection of aspects of new documents that were not previously seen. In contrast, previous work applying automatic aspect detection can only extract what was seen during training. 6. Embodiments of the present disclosure provide more efficient processing of natural language documents, including providing more efficient use of computing resources (such as processing time).

[0019] Thus, embodiments implemented in accordance with the present disclosure provide a significant improvement over the state of the art in computer-implemented natural language processing and, in particular, in the field of special computers configured to implement the summarization task for natural language processing.

[0020] In a first aspect, the present invention provides a computer-implemented machine learning method for aspect-based summarization. Aspect candidates are extracted from a document. The aspect candidates are filtered by an ambiguity resolution technique using different aspect granularity levels. Information content is extracted based on the filtered aspect candidates. A summarization technique is implemented to generate a summary for a plurality of document-aspect pairs from the information content. An aspect-based summary is generated based on the summary.

[0021] In a second aspect, the present invention provides the method according to the first aspect, wherein the document is obtained using one or more document retrieval techniques that determine similarities between sentences of each document of the document.

[0022] In a third aspect, the present invention provides the method according to the first or second aspect, wherein the one or more document retrieval techniques include semantic similarity calculation or syntactic similarity calculation.

[0023] In a fourth aspect, the present invention provides the method according to any one of the first to third aspects, which further includes implementing an aspect candidate extraction technique from one or more aspect candidate extraction techniques to automatically determine aspect candidates from a document, and the one or more aspect candidate extraction techniques include a rule-based method for processing the content of the document to identify headers in the document or for detecting keywords from a dictionary in the document.

[0024] In a fifth aspect, the present invention provides a method according to any one of the first to fourth aspects, wherein one or more aspect candidate extraction techniques include a clustering algorithm or neural network configured to automatically discover aspect candidates, the clustering algorithm and neural network being trained on a training dataset.

[0025] In a sixth aspect, the present invention provides a method according to any one of the first to fifth aspects, wherein the deambiguation technique filters aspect candidates to determine a common aspect from the aspect candidates.

[0026] In a seventh aspect, the present invention provides a method according to any one of the first to sixth aspects, wherein the disambiguation technique includes a dictionary-based method or a knowledge-based method.

[0027] In the eighth aspect, the present invention provides a method according to any one of the first to seventh aspects, wherein different aspect granularity levels are obtained from a database.

[0028] In the ninth aspect, the present invention provides a method according to any one of the first to eighth aspects, wherein the summarization technique includes a post-extraction abstraction technique or an abstraction technique, the post-extraction abstraction technique being based at the sentence or paragraph level.

[0029] In the tenth aspect, the present invention provides a method according to any one of the first to ninth aspects, wherein the post-extraction abstraction summarization technique includes a term frequency (TF)-inverse dense frequency (IDF) technique.

[0030] In the eleventh aspect, the present invention provides a method according to any one of the first to tenth aspects, wherein the summarization technique includes applying a large language model (LLM) to improve the fluency of extractive summaries corresponding to multiple document-aspect pairs.

[0031] In the twelfth aspect, the present invention provides a method according to any one of the first to eleventh aspects, wherein the document does not include explicit section divisions.

[0032] In the thirteenth aspect, the present invention provides a method according to any one of the first to twelfth aspects, wherein aspect-based summarization compares summaries for multiple document-aspect pairs.

[0033] In a fourteenth aspect, the present invention provides a computer system for aspect-based summarization comprising one or more processors configured to perform, alone or in combination, a machine learning method for aspect-based summarization described in any one of the first to thirteenth aspects.

[0034] In a 15th aspect, the present invention provides a tangible, non-temporary, computer-readable medium for aspect-based summarization, which, when executed by one or more hardware processors, provides execution of the machine learning method described in any one of the first to 13th aspects.

[0035] Figure 1 shows a comparison between a conventional summarization method 100 and a method for aspect-based summarization according to one embodiment 102 of the present disclosure. The conventional summarization method 100 may depend on documents that include section divisions. In contrast, the method for aspect-based summarization according to one embodiment of the present disclosure 102 described herein can generate document comparisons even when section divisions are missing from the input documents. As shown in Figure 1, the conventional summarization method 100 does not search for similar documents and does not include similar documents and their features for generating an aspect-based summary. The method for aspect-based summarization according to one embodiment of the present disclosure shows a flexible selection of aspects for generating a summary to compare different documents in 104. Figure 1 also shows providing a summary of similarly searched documents in 106. The method for aspect-based summarization according to one embodiment of the present disclosure 102 may include techniques for selecting or retrieving documents for analysis by using one or more document search techniques to determine the similarity between sentences in each document from the documents.

[0036] Figure 2 shows a system architecture according to one embodiment of the present disclosure. The system architecture shown in Figure 2 implements a process 200 having four steps (A, B, C, D) aimed at generating summaries of multiple documents by automatically filtered aspects. This process enables a transparent and flexible method for selecting aspects of interest in domain-specific summarization use cases. This system is particularly useful when the input documents do not have explicit section divisions. Step A: To initiate process 200, the input document 202 is searched, relevant documents are identified, and a set of documents is obtained. In this step, a method is used that can calculate the similarity between sentences. For example, a module such as the document search module 204 can implement semantic similarity calculation (e.g., Sentence Transformers), or it can calculate the similarity between sentences using syntactic similarity calculation (e.g., Bag of Words). Step B: Next, automatic aspect selection is introduced to discover common and unique aspects to focus on across multiple documents 206. This is done by using a method that can automatically discover aspects. For example, a rule-based method may be used after processing content such as headers in the documents, or after detecting keywords from a dictionary of common terms such as “method” or “architecture.” A method or model for extracting aspect candidates may be implemented by the aspect candidate extraction module 208. This may be done via a model (e.g., a clustering algorithm or neural network trained on a supervised dataset) configured to extract aspects and automatically discover aspects. These extracted aspect candidates are then filtered to retain the discovered common aspects. The aspect ambiguity disambiguation module 210 is a process of determining a specific meaning or interpretation of a term, phrase, or aspect in a given context, particularly when the term or phrase has multiple possible meanings, and may be used to filter aspect candidates to provide filtered aspects 214. Filtering can be performed by implementing word semantic disambiguation and coreference analysis, which aims to identify the most relevant meaning for a given context and to identify when different aspects mentioned in the text actually refer to the same underlying concept. These techniques include dictionary-based or knowledge-based methods that disambiguate aspects based on definitions or associations from a dictionary or knowledge base, as well as supervised methods using labeled data to group similar aspects together to detect which aspects have the same meaning, or unsupervised methods using clustering. Flexibility control is also provided to determine aspect granularity via database input to the aspect granularity determination module 212. Flexibility control is also provided to determine aspect granularity via database input to the aspect granularity determination module 212. Aspect granularity determination is triggered in conjunction with the aspect disambiguation module 210, which includes user input to specify how coarse and fine-grained the user's interest is and to guide the determination of the granularity level of the aspects to be filtered. Step C: The filtered aspects 214 are then used for document-aspect summarization, and each document-aspect pair generates a short summary based on either the post-extraction abstract summarization module 216 or the abstract summarization module 218, which apply their respective summarization methods. In the former case, as an example, the Term Frequency (TF)-Inverse Dense Frequency (IDF) method can be used, based on sentence and paragraph levels rather than general token levels. This allows determining the overall topic of the paragraphs to control the focus of the summary to be extracted. This is followed by improving the fluency of the extraction summarization using the large language model (LLM) 220. In this embodiment, the survey summarization in step C may be performed by a module such as the survey summarization module 224. This step generates summaries for each document in the form of separate summaries for each aspect extracted from the document. Step D: The survey summary overview triggers a flexible query or domain-specific needs process to generate summaries on both the aspects and documents selected by the inter-document aspect-based summarization module 222. This step aggregates the results from Step C and multiple documents to generate a single coherent summary comparing the documents, segmented by aspect. Database 226 represents the user's ability to select summaries from different aspects and documents, or the user's ability to interact with those summaries.

[0037] Compared to standard aspect-based summarization techniques, the systems implemented according to the embodiments of this disclosure offer the following improvements: 1. Process document comparison even if there is no section division. 2. Generate summaries for comparing different documents, allowing for flexible selection of aspects. 3. Enables the generation of summaries tailored to user needs. This means the system can adapt more versatility and customizability than standard aspect-based summarization methods.

[0038] Accordingly, embodiments of the present invention provide a general improvement to computers in machine learning systems to provide improved aspect-based document summarization. Furthermore, embodiments of the present invention can be practically applied to use cases to achieve further improvements in many technological fields, including, but not limited to, healthcare (e.g., digital medicine, personalized healthcare, AI-assisted drug or vaccine development, cyber threat detection, materials development, patents, public safety, and smart cities (e.g., automated traffic or vehicle control, smart districts, smart buildings, smart industrial plants, smart agriculture, energy management, etc.)).

[0039] Embodiments of the present invention can be applied to the field of summarizing knowledge work, such as for use in consulting support, contact center support, or cyber threat intelligence. An exemplary use case involves being given at least one input document and automatically generating a summary based on aspects discovered. For example, given at least one document as input, at least one document may be shortened to create an adaptable summary, the summary based on (1) aspects discovered, (2) aspects considered most important, and (3) similarities / differences between documents when two or more input documents are supplied. The output of the exemplary use case may include summaries of different aspects. An example of an automated action and the resulting physical change (technical aspect) is the automated generation of a document containing summaries with different aspects. Summaries generated by the features of this disclosure may be used as input for further AI systems or displayed for further interaction by a user.

[0040] Embodiments of the present invention can be applied, for example, to the field of medical AI or automated healthcare, and can generate, for example, a patient's medical history summary. An exemplary use case involves being given at least one patient report from a patient's medical history and generating a summary that takes different aspects into consideration. The aspects may be different categories, for example, previous surgeries, previous medications, previous medical conditions, etc. Given input such as at least one patient report, at least one patient report can be shortened and adapted to create a summary, the summary based on (1) aspects found, (2) aspects considered most important, and (3) similarities / differences between documents, if two or more input documents are supplied. The output of such an exemplary use case may include a summary that highlights different aspects and differences / similarities, which provides, for example, a basis for making a diagnosis. An example of an automated action and physical change (technical aspect) is the selection of one or more drug therapies or treatments by a healthcare professional based on the summary that highlights different aspects and differences / similarities.

[0041] Embodiments of the present invention may be applied, for example, to the field of automated public safety, such as generating citizen report summaries. An exemplary use case involves being given at least one citizen report and creating an aspect-based summary for an administrative officer (for example, a citizen is seeking employment at an employment agency and the administrative officer may be the citizen's caseworker). For example, given at least one citizen report as input, at least one citizen report can be shortened and adapted to create a summary, the summary based on (1) aspects found, (2) aspects considered most important, and (3) similarities / differences between documents, if two or more input documents are provided. The output of such an exemplary use case may include a summary that highlights various aspects and differences / similarities, which provides an basis for administrative officers to make decisions. An example of an automated action or physical change (technical) is the generation and execution of an instruction to repair a broken road sign / lighting based on a summary that highlights various aspects and differences / similarities. The generation of such orders may include prioritizing certain areas of the city, including multiple similar aspects that address complaints in a particular area.

[0042] Embodiments of the present invention may be applied, for example, to the field of automated research, for example, to generate research papers or patent abstracts. An exemplary use case involves being given at least one patent or patent application and creating an aspect-based abstract. This saves patent readers time and allows them to extract insights they might otherwise overlook. For example, given input of at least one patent or patent application, at least one patent or patent application can be shortened and adapted to create an abstract, which is based on (1) aspects found, (2) aspects considered most important, and (3) similarities / differences between documents, if two or more input documents are provided. The output of such an exemplary use case may include an abstract that highlights various aspects and differences / similarities, which provides the user with a basis for making a final decision (e.g., whether the patent or patent application relates to another patent). An example of automated actions and physical changes (technical aspects) is the creation of patent applications / documents based on an analysis of whether a patent or patent application is relevant to another patent, for example, a comparison of prior art with potential novel ideas.

[0043] Embodiments of the present invention can be applied to the field of automated or AI-assisted law enforcement, for example, with respect to integration into forensic tools or generation of summaries of suspect reports. An exemplary use case involves being given at least one suspect report and creating an aspect-based summary for an officer. This saves the officer time and allows for the extraction of insights that might otherwise be missed. For example, being input with at least one suspect report and creating a shortened and adapted summary of at least one suspect report, the summary based on (1) aspects found, (2) aspects considered most important, and (3) similarities / differences between documents, if two or more input documents are provided. The output of such an exemplary use case may include a summary highlighting various aspects and differences / similarities, which provides the officer with a basis for making final decisions. An example of automated actions and physical changes (technical aspects) is the filing of a complaint for legal action and police investigation against a suspect, including similarities (e.g., similar methods) across multiple documents. Automated actions may also include activating sensors or forensic tools, such as cameras or area surveillance equipment. Other suspects may be disregarded because they do not exhibit similar characteristics across the analyzed documents.

[0044] Embodiments of the present invention may be applied to the field of automated customer service or call processing, for example, with respect to the generation of call center summaries. An exemplary use case involves a situation in which a call handler must write a summary after answering a call in a call center to resolve a problem. A system implemented according to aspects of this disclosure can securely automate this process and thus reduce the time spent by the call handler in generating the summary. For example, given a transcribed call input, the transcribed call can be shortened and a fitted summary can be created, the summary based on (1) aspects found, (2) aspects considered most important, and (3) similarities / differences between documents, if two or more input documents are provided. The output of such an exemplary use case may include a summary highlighting various aspects and differences / similarities, which provides the user with a basis for making a final decision (e.g., whether the call relates to another person, and what decision was made in the case of similarities). An example of an automated action or physical change (technical) is to automatically forward an incoming call to the correct destination to handle a similar issue to other issues that have been previously analyzed and directed to similar destinations, using a transcript of that call compared to a previously transcribed call.

[0045] Embodiments of the present invention may be applied, for example, to the field of AI-assisted legal or contract enforcement for contract management and compliance. An exemplary use case involves a lawyer manually checking a contract to ensure that all information is present in the document in accordance with guideline documents. A system implemented in aspects of this disclosure can securely automate this process, thus reducing the time a lawyer spends making decisions about a contract. For example, given input of a contract document, the contract document can be shortened and a tailored summary can be created, the summary based on (1) aspects found, (2) aspects considered most important, and (3) similarities / differences between documents if two or more input documents are provided. The output of such an exemplary use case may include a summary highlighting various aspects and differences / similarities, which provides the user with a basis for making a final decision (e.g., whether the contract complies with guidelines and what decisions have been made in similar cases). An example of automated actions and physical changes (technical aspects) is the automated correction of errors or omissions in a document based on a computer-generated summary of various aspects.

[0046] Embodiments of the present invention may be applied, for example, to the field of AI-assisted human resources to provide resume insights and analysis. An exemplary use case involves a situation in which an employee must evaluate an applicant's resume and compare it with multiple applicants to find the best fit. A system implemented according to aspects of this disclosure can securely automate this process and thus reduce the time a call handler spends generating summaries of individual resumes. For example, an applicant's resume can be input, the applicant's resume can be shortened, and a tailored summary can be created, based on (1) aspects discovered, (2) aspects considered most important, and (3) similarities / differences between documents, if two or more input documents are provided. The output of such an exemplary use case may include a summary highlighting various aspects and differences / similarities, which provides the user with a basis for making a final decision (e.g., whether the applicant can proceed with the application process and what decisions were made regarding similar cases). An example of automated actions and physical changes (technical aspects) is the progress of a specific applicant in the application process by telephone or other means. For example, a computer can generate messages, emails, and hyperlinks that are sent to an applicant's user device to advance the application process, based on summaries of different aspects.

[0047] Fixed aspect-based summarization, which relies on a predefined list of aspects to generate summaries, has limitations when dealing with documents that lack a clear topical division or documents generated from multiple resources in various formats. Furthermore, when comparing multiple documents, fixed aspects may not adequately address diverse customer needs, as the importance of aspects may differ depending on the customer's role. In contrast, automated aspect discovery enables the dynamic identification of relevant aspects, allowing it to adapt to a wide range of content types, accommodate different customer roles, and facilitate application across diverse domains and documents in various formats.

[0048] Figure 3 shows a flowchart of a method for aspect-based summarization according to one embodiment of the present invention. Flowchart 300 includes extracting aspect candidates from a document in 302. In the embodiment, documents can be retrieved using one or more search techniques that determine the similarity between sentences in each document. One or more search techniques may include semantic similarity calculation or syntactic similarity calculation. According to one embodiment of the present disclosure, an aspect candidate extraction technique can be implemented to automatically determine aspect candidates from a document. The aspect candidate extraction technique may include rule-based methods for processing the content of a document to identify headers within the document or for detecting keywords from a dictionary within the document.

[0049] In Figure 3, the flowchart 300 includes filtering aspect candidates by a deambiguation technique using different aspect granularity levels in 304. The flowchart 300 includes extracting information content based on the filtered aspect candidates in 306. The flowchart 300 includes implementing a summarization technique in 308 to generate multiple document-aspect pair results from the information content. In embodiments, the summarization technique includes a post-extraction type summarization technique or an abstraction type summarization technique. Figure 3 includes generating an inter-document aspect-based summary of the multiple document-aspect pair results in 310.

[0050] Referring to Figure 4, the processing system 400 may include one or more processors 402, a memory 404, one or more input / output devices 406, one or more sensors 408, one or more user interfaces 410, and one or more actuators 412. The processing system 400 may represent each of the computing systems disclosed herein.

[0051] The processor 402 may include one or more separate processors, each having one or more cores. Each of the separate processors may have the same or different architecture. The processor 402 may include one or more central processing units (CPUs), one or more graphics processing units (GPUs), circuits (such as application-specific integrated circuits (ASICs)), digital signal processors (DSPs), etc. The processor 402 may be mounted on a common board or on multiple different boards.

[0052] Processor 402 is configured to perform a specific function, method, or operation (for example, to provide the performance of a function, method, or operation) if one or more of the separate processors can perform the operation that embodies the function, method, or operation. Processor 402 can perform the operation that embodies the function, method, or operation, for example, by executing code stored in memory 404 (for example, interpreting a script) and / or by trafficing data through one or more ASICs. Processor 402, and therefore the processing system 400, can be configured to automatically perform any and all of the functions, methods, and operations disclosed herein. Accordingly, the processing system 400 can be configured to implement any (for example, all) of the protocols, devices, mechanisms, systems, and methods described herein.

[0053] For example, where this disclosure states that a method or device performs task "X" (or that task "X" is performed), such statement should be understood to disclose that the processing system 400 may be configured to perform task "X". At the very least, if the processor 402 is configured to perform a function, method, or operation, the processing system 400 is configured to do so.

[0054] Memory 404 may include volatile memory, non-volatile memory, and any other medium capable of storing data. Each of volatile memory, non-volatile memory, and any other type of memory may include multiple different memory devices, which are located in multiple different locations and each having a different structure. Memory 404 may include remotely hosted storage (e.g., cloud storage).

[0055] Examples of memory 404 include non-temporary computer-readable media, such as RAM, ROM, flash memory, EEPROM, and any type of optical storage disk, such as DVD, Blu-ray® discs, and magnetic storage, holographic storage, HDD, SSD, and any medium that can be used to store program code in the form of instructions or data structures. Any and all of the methods, functions, and operations described herein can be fully embodied in the form of tangible and / or non-temporary machine-readable code (e.g., interpretable scripts) stored in memory 404.

[0056] The input / output device 406 may include any components for trafficking data, such as ports, antennas (i.e., transceivers), and printed conductive paths. The input / output device 406 is capable of wired communication such as USB®, DisplayPort®, HDMI®, and Ethernet. The input / output device 406 can enable electronic, optical, magnetic, and holographic communication with appropriate memory 406. The input / output device 406 is capable of wireless communication via WiFi®, Bluetooth®, cellular (e.g., LTE®, CDMA®, GSM®, WiMAX®, NFC®), GPS, and the like. The input / output device 406 may include wired and / or wireless communication paths.

[0057] Sensor 408 can capture physical measurements of the environment and report them to processor 402. User interface 410 may include a display, physical buttons, speaker, microphone, keyboard, etc. Actuator 412 may enable mechanical force control by processor 402.

[0058] The processing system 400 may be distributed. For example, some components of the processing system 400 may reside in a remote hosted network service (e.g., a cloud computing environment), while other components of the processing system 400 may reside in a local computing system. The processing system 400 may have a modular design in which a particular module includes multiple features / functions as shown in Figure 4. For example, an I / O module may include volatile memory and one or more processors. As another example, individual processor modules may include read-only memory and / or a local cache.

[0059] B. Tan, L. Qin, E. Xing, Z. Hu (November 2020). Summarizing Text on Any Aspects: A Knowledge-Informed Weakly-Supervised Approach. The proceedings of the 2020 Empirical Methods in Natural Language Processing (EMNLP) conference (pages 6301-6309), incorporated herein by reference, describes an abstract aspect-based summarization using a weakly supervised approach that extracts and uses aspects from only a single document. In contrast, embodiments of the present invention provide a versatile aspect-discovery summarization that can handle both single and multiple documents. In particular, an aspect deambiguation step makes it possible to extract aspects from diverse documents and then generate a summarization based on the automatically identified aspects to enhance the decision-making process.

[0060] Y. Wang, Y. Zhou, M. Wang, Z. Chen, Z. Cai, J. Chen, VCLeung (2023). Multidocument Aspect Classification for Aspect-Based Abstractive Summarization. Incorporated herein by reference, IEEE Transactions on Computational Social Systems describes a two-step pipeline: a first aspect discovery presented as a classification task, followed by a summarization generator. In contrast to this approach, embodiments of the present invention present that the input may be one or more documents and provide an additional step of searching for relevant documents which may help provide a better overview and summary of the documents. Furthermore, in contrast to this approach, aspect discovery according to embodiments of the present invention also provides the use of both an automated method that relies on extracting headers, sections, etc., rather than being purely ML-based, and an NLP method that extracts aspects based on topics discussed in a given section. Furthermore, a later disambiguation step according to embodiments of the present invention ensures uniformity among summaries of all documents and makes it possible to provide equivalent summaries across different inputs.

[0061] X. Yang, Y. Li, X. Zhang, H. Chen, W. Cheng (2023). Exploring the limits of ChatGPT for query or aspect-based text summarization.arXiv preprint arXiv:2302.08081, incorporated herein by reference, describes evaluating ChatGPT for aspect-based summarization that requires a written query from the user to indicate the desired aspect. In contrast to this approach, embodiments of the present invention enable the automatic extraction of aspects and do not require user input.

[0062] The subject matter of this disclosure is illustrated and described in detail in the drawings and the foregoing description, but such illustrations and descriptions should be considered illustrative or illustrative and not limiting. Any descriptions made herein characterizing the invention should also be considered illustrative or illustrative and not limiting, since the invention is defined by the claims. Those skilled in the art will understand that modifications and alterations may be made within the following claims, which may include any combination of features from the different embodiments described above.

[0063] The terms used in the claims should be interpreted to have the broadest reasonable interpretation consistent with the foregoing description. For example, the use of the article "a" or "the" when introducing elements should not be interpreted as excluding multiple elements. Similarly, the statement "or" should be interpreted as inclusive, and the statement "A or B" does not exclude "A and B" unless it is clear from the context or the foregoing description that only one of A and B is intended. Furthermore, the statement "at least one of A, B, and C" should be interpreted as one or more of the group of elements consisting of A, B, and C, and should not be interpreted as requiring at least one of each of the enumerated elements A, B, and C, regardless of whether A, B, and C are related as categories. Furthermore, the statement "A, B, and / or C" or "at least one of A, B, or C" should be interpreted to include any single entity from the enumerated elements, such as A, or any subset from the enumerated elements, such as A and B, or the entire enumeration of elements A, B, and C.

Claims

1. A computer implementation method for aspect-based summarization, Extract aspect candidates from the document, The aspect candidates are filtered by a de-ambiguation technique that uses different aspect granularity levels. Information content is extracted based on the filtered aspect candidates. Summarization technology is implemented to generate summaries of multiple document-aspect pairs from the aforementioned information content. Based on the above summary, an aspect-based summary is generated. A computer implementation method that includes the following.

2. The computer implementation method according to claim 1, wherein the documents are obtained using one or more document retrieval techniques that determine the similarity between sentences in each document.

3. The computer implementation method according to claim 2, wherein the one or more document retrieval techniques include semantic similarity calculation or syntactic similarity calculation.

4. A computer implementation method according to any one of claims 1 to 3, further comprising implementing one aspect candidate extraction technique from one or more aspect candidate extraction techniques to automatically determine the aspect candidates from the document, wherein the one or more aspect candidate extraction techniques include a rule-based method for processing the content of the document to identify headers in the document or for detecting keywords from a dictionary in the document.

5. The computer implementation method according to claim 4, wherein the one or more aspect candidate extraction techniques include a clustering algorithm or a neural network configured to automatically discover the aspect candidates, and the clustering algorithm and the neural network are trained on a supervised dataset.

6. The computer implementation method according to any one of claims 1 to 5, wherein the ambiguity resolution technique filters the aspect candidates in order to determine a common aspect from among the aspect candidates.

7. The computer implementation method according to claim 6, wherein the ambiguity resolution technique includes a dictionary-based method or a knowledge-based method.

8. The computer implementation method according to claim 6, wherein the different aspect granularity levels are obtained from a database.

9. The summarization technique includes a post-extraction abstraction technique or an abstraction technique, wherein the post-extraction abstraction technique is based on the sentence or paragraph level, according to the computer implementation method of any one of claims 1 to 8.

10. The computer implementation method according to claim 9, wherein the post-extraction abstraction summarization technique includes a term frequency (TF) - inverse dense frequency (IDF) technique.

11. The computer implementation method according to any one of claims 1 to 10, wherein the summarization technique includes applying a large language model (LLM) to improve the fluency of the extractive summaries corresponding to the plurality of document-aspect pairs.

12. The computer implementation method according to any one of claims 1 to 11, wherein the document does not include an explicit division of sections.

13. The computer implementation method according to any one of claims 1 to 12, wherein the aspect-based summaries compare the summaries for the plurality of document-aspects.

14. An interpretable domain-adaptive computer system for aspect-based summarization, comprising one or more hardware processors, the hardware processors, individually or in combination, perform the following steps, namely: Extract aspect candidates from the document, The aspect candidates are filtered by a de-ambiguation technique that uses different aspect granularity levels. Information content is extracted based on the filtered aspect candidates. Summarization technology is implemented to generate summaries of multiple document-aspect pairs from the aforementioned information content. A computer system configured to perform the task of generating an aspect-based summary based on the aforementioned summary.

15. A tangible, non-temporary computer-readable medium in which, when an instruction is executed by one or more processors, the following steps occur: Extract aspect candidates from the document, The aspect candidates are filtered by a de-ambiguation technique that uses different aspect granularity levels. Information content is extracted based on the filtered aspect candidates. Summarization technology is implemented to generate summaries of multiple document-aspect pairs from the aforementioned information content. A tangible, non-temporary, computer-readable medium having instructions for providing an aspect-based summary by performing the following: generating an aspect-based summary based on the aforementioned summary.

Images