Automatic discovery of aspect-based summarization

Abstract

A computer-implemented, machine learning method for aspect-based summarization. Aspect candidates are extracted from documents. The aspect candidates are filtered by a disambiguation technique that uses different aspect granularity levels. Informative contents are extracted based on the filtered aspect candidates. A summarization technique is implemented to generate summaries for a plurality of document-aspect pairs from the informative contents. An aspect- based summary is generated based on the summaries. The method has applications including, but not limited to, use cases in medicine / healthcare, public safety, cyberthreat intelligence and other artificial intelligence applications for document summarization.

Description

AUTOMATIC DISCOVERY OF ASPECT-BASED SUMMARIZATIONCROSS-REFERENCE TO PRIOR APPLICATION

[0001] Priority is claimed to U.S. Provisional Application Serial No. 63 / 522,447 filed on June 22, 2023, the entire contents of which is hereby incorporated by reference herein.FIELD

[0002] The present disclosure relates to Artificial Intelligence (Al) and machine learning, and in particular to a method, system and computer-readable medium for automatic discovery of aspect-based summarization.BACKGROUND

[0003] Summarization is a popular task in natural language processing (NLP), which generates a concise and coherent summary from a longer document or set of documents. The goal of summarization is to distill the most important information and main points, while preserving the overall meaning and context of the original text. In this way, automatic text summarization helps humans save time when confronted with large amounts of text.

[0004] NLP summarization may utilize, for example, large language models (LLMs), such as ChatGPT. LLMs are advanced deep learning models that have been trained on vast amounts of text data. They have the capability to understand and generate human-like language. These models utilize a transformer architecture, which allows them to process and generate text by considering contextual dependencies and capturing long-range dependencies in the input. SUMMARY

[0005] In an embodiment, the present invention provides a computer-implemented, machine learning method for aspect-based summarization. Aspect candidates are extracted from documents. The aspect candidates are fdtered by a disambiguation technique that uses different aspect granularity levels. Informative contents are extracted based on the fdtered aspect candidates. A summarization technique is implemented to generate summaries for a plurality of document-aspect pairs from the informative contents. An aspect-based summary is generated based on the summaries. The method has applications including, but not limited to, use cases in medicine / healthcare, public safety, cyberthreat intelligence and other artificial intelligence applications for document summarization.BRIEF DESCRIPTION OF THE DRAWINGS

[0006] Embodiments of the present disclosure will be described in even greater detail below based on the exemplary figures. The present disclosure is not limited to the exemplary embodiments. All features described and / or illustrated herein can be used alone or combined in different combinations in embodiments of the present disclosure. The features and advantages ofvarious embodiments of the present disclosure will become apparent by reading the following detailed description with reference to the attached drawings which illustrate the following:

[0007] FIG. 1 illustrates a comparison between a previous method for summarization to novel systems and methods for aspect-based summarization according to an embodiment of the present disclosure;

[0008] FIG. 2 schematically illustrates a method and system architecture according to an embodiment of the present disclosure;

[0009] FIG. 3 illustrates a flow diagram of a method for aspect-based summarization according to an embodiment of the present invention; and

[0010] FIG. 4 illustrates a block diagram of an exemplary processing system, which can be configured to perform any and all operations disclosed herein.DETAILED DESCRIPTION

[0011] Embodiments of the present disclosure are directed to providing improved NLP summarization, including by providing automatic discovery of aspect-based summarization.

[0012] According to existing technology, state-of-the-art NLP summarization techniques can have an issue that a static aggregated summary is provided based on multiple documents. To alleviate the static issue, aspect-based summarization generates different summaries based on a fixed (e.g., predefined) list of topics such as sections or subsections discussing a particular goal, such as a scientific method (also referred to as aspects). In embodiments aspects can refer to any topic of interest. For example, a document such as a news article or medical report can include multiple topics / aspects. However, this limits the technical capabilities when the documents are without clear segmentation of topics or vary from multiple resources with different formats. This also limits the capability of user needs, for example, when different user roles might consider different importance of aspects, while comparing multiple documents. In fact, even for a single user, the need for a summary might differ and depend on the current task at hand. Additionally, the time and effort required for comparing many documents can be alleviated by utilizing the generated aspect-based summaries for the creation of contrastive summaries, which show the differences and similarities between the inputs. While state-of-the-art NLP summarization techniques rely on predefined templates for the task, embodiments of the present disclosure allow the automatic discovery of aspects, making it easy to apply in different domains and for documents following various formats.

[0013] Accordingly, embodiments of the present disclosure enable automatically discovering different topics and adjusting summaries according to the user specified needs and requested granularity level. This helps to save time and increase efficiency because it allows the user to quickly identify topics and similarities and differences between the documents withoutthe need for extensive reading. Thereby, embodiments of the present disclosure simplify decision-making processes and streamline work processes.

[0014] Embodiments of the present disclosure provide:1) Relevance: Auto-discovery of aspects helps to select the most relevant aspects that are 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. Further, it helps improve accuracy, reliability and trustworthiness of the Al systems. Without a user providing input for selecting aspects of interest from the input documents the present system can use a combination of automated methods that extract headers, titles, and subtitles to register the different segments / aspects in the summary; and use a trained model that extracts additional aspects based on the topics of discussion in different paragraphs that may or may not match the headers.2) Efficiency: By automatically filtering the aspects, the process of summarization across multiple documents become more efficient. This helps significantly reduce the time and effort required to process and summarize the information across multiple documents, reduce the workload and streamline the process. This also provides for faster computation, conservation of computational resources and / or reduced compute time and / or power.3) Customization: Dynamic selection of aspects allow for a high degree of customization in generative summarization. It enables personalization of the summary based on the user’s specific preferences and domain interests, making the summary more relevant and useful to their needs.

[0015] In a first embodiment of the present disclosure, a method is provided for generating aspect-based summaries for multiple documents in a dynamic and customizable manner. The method includes:1) Selecting a document retrieval technique to find relevant documents and implement it.2) Selecting an aspect candidates extraction technique and implement it.3) Implementing the aspect disambiguation technique that considers different aspect granularity level.4) Extracting informative contents based on the filtered aspects. In this step, the parts that are most relevant for each aspect are extracted directly from the original text. For instance, if the aspect is the methodology of a scientific publication, the results is paragraphs extracted the paper that correspond to this topic.5) Selecting an extractive-then-abstractive or abstractive summarization technique and implement it.6) Selecting a summarization technique to generate cross-document aspect-based summaries; and7) Creating a user interface that links four steps (A, B, C, D - described below). The user interface would allow the user to select an initial document that they would like to summarize. Upon selecting the document and uploading it, the user would be asked to input any aspect that they would like to specifically focus on (such as, summarize the technical method, etc.). Once the user inputs his / her preference, relevant documents are retrieved and several different aspects are extracted, including (if applicable) the aspect specified by the user. The aspects are then filtered and disambiguated in order to prevent any repetitions, and the user is then shown a final list of aspects. In the backend, the model generated the summary for each aspect and displays the final output to the user shown in separate components, where the user can simply click on the aspect of interest in order to find the corresponding summary of the uploaded documented as well as the retrieved documents.

[0016] In a second embodiment, a non-transitory computer readable medium is provided storing instructions, which when executed by one or more processors, executes the method of the first embodiment.

[0017] In a third embodiment, a system is provided comprising one or more processors, which alone or together, are configured to execute the method of the first embodiment.

[0018] Embodiments of the present disclosure provide several advantages compared to current state-of-the-art. These advantages include:1. Existing work relies on a pre-defined list of topics whereas embodiments of the present disclosure have an auto-discovery step, making embodiments according to the present disclosure more flexible and adaptable to user needs, which in turn generates better summaries for each user query.2. Based on the different discovered aspects, embodiments according to the present disclosure can more efficiently contrast different documents by highlighting their similarities and differences, therefore reducing workload and streamlining processes.3. The dynamic selection of aspects allows for the creation of different summaries adapted to user needs.4. Embodiments according the present disclosure are not domain specific because the aspects are automatically detected, embodiments may be used for contracts, patents, resumes and case reports. In contract, previous work focuses on domain-specific use cases, such as nutrition reports.5. Embodiments of the present disclosure may use a combination of machine learning and rule-based methods in order to extract the aspects, this will allow for detection of aspects of new documents that have not been previously seen. In contrast, previous works that apply automatic detection of aspects can only extract those seen in training.6. Embodiments of the present disclosure provide for more efficient processing of natural language documents, including providing a more efficient use of compute resources (e.g., processing time).

[0019] Accordingly, embodiments implemented according to the present disclosure provide marked-improvements over the state-of-the-art in the field of computer implemented natural language processing, and specifically specialty computers configured to implement summarization tasks 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 documents. The aspect candidates are filtered by a disambiguation technique that uses different aspect granularity levels. Informative contents are extracted based on the filtered aspect candidates. A summarization technique is implemented to generate summaries for a plurality of document-aspect pairs from the informative contents. An aspect-based summary is generated based on the summaries.

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

[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 a semantic similarity computation or syntactic similarity computation.

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

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

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

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

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

[0028] In a ninth aspect, the present invention provides the method according to any of the first to eighth aspects, wherein the summarization technique includes an extractive-then-abstract summarization technique or an abstractive summarization technique, wherein the extractive- then-abstract summarization technique is based on a sentence or paragraph level.

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

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

[0031] In a twelfth aspect, the present invention provides the method according to any of the first to eleventh aspects, wherein the documents do not include explicit segmentation of sections.

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

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

[0034] In a fifteenth aspect, the present invention provides a tangible, non-transitory computer-readable medium for aspect-based summarization which, upon being executed by one or more hardware processors, provide for execution of a machine learning method according to any of the first to the thirteenth aspects.

[0035] FIG. 1 illustrates a comparison between a previous method for summarization 100 to a method for aspect-based summarization according to an embodiment of the present disclosure102. Previous methods for summarization 100 may rely on documents which include section segmentations. In comparison, the method for aspect-based summarization according to an embodiment of the present disclosure 102 described herein can generate document comparisons even in situations where section segmentation is lacking from the input documents. As illustrated in FIG. 1, previous methods for summarization 100 do not retrieve similar documents and include them and their features for generating aspect-based summaries. The method for aspect-based summarization according to an embodiment of the present disclosure 102 depicts at 104 flexible selection of aspects for generating summaries to compare different documents. FIG. 1 also depicts providing summaries of similar retrieved documents at 106. The method for aspect-based summarization according to an embodiment of the present disclosure 102 can include techniques for selecting or obtaining documents for analysis by using one or more document retrieval techniques that determine a similarity between sentences of each document from the documents.

[0036] FIG. 2 illustrates a system architecture according to an embodiment of the present disclosure. The system architecture illustrated in FIG. 2 implements a process 200 having four steps (A, B, C, D) aimed at generating summaries of multiple documents by automatic fdtered aspects. This process allows for a transparent and flexible manner of selecting focused aspects with domain-specific summarization use cases. The system is especially useful when input documents lack explicit segmentation of sections.Step A: To begin the process 200, an input document 202 is retrieved, and relevant documents are identified resulting in a set of documents. For this step, a method is used that can compute the similarity between sentence. For example, a module such as document retrieval module 204 may implement a semantic similarity computation (e.g., Sentence Transformers) or syntactic similarity computation (e.g., bag of words) may be used to compute the similarity between the sentences.Step B: The automatic selection of aspects is then introduced to discover common and unique aspects to focus on across documents 206. This is done by using a method that can automatically discover aspects. For example, a rule based method may be used that goes through the processing of the content, such as headers in the document or detecting keywords from a dictionary of general terminology such as “method”, “architecture”, etc. Methods or models for aspect candidate extraction can be implemented by aspect candidate extraction module 208. It can also be done through a model that will extract the aspects and is configured to automatically discover aspects (e.g. a clustering algorithm or a neural network that was trained on a supervised dataset). These extracted aspect candidates are then filtered to keep the common aspects detected. An aspectdisambiguation module 210 is the process of determining the specific meaning or interpretation of a term, phrase, or aspect in a given context, especially when that term or phrase has multiple possible meanings, and can be used to filter the aspect candidates to provide filtered aspects 214. Filtering can be executed by implementing word sense disambiguation and co-reference resolution, which aim to identify the most relevant sense for the specific context and identify when different aspects mentioned in the text actually refer to the same underlying concept. The techniques include dictionary-based or knowledge-based methods to disambiguate aspects based on their definitions or associations from dictionaries or knowledge bases, as well as supervised or unsupervised methods with either labeled data or clustering to group similar aspects together to detect which aspects indicate the same meaning. Flexibility control is also provided for determining aspect granularity through database input to aspect granularity determination module 212.Flexibility control is also provided for determining aspect granularity through database input to aspect granularity determination module 212. The aspect granularity determination is triggered along with the aspect disambiguation module 210, which involves the user input to specify how coarse and fine-grained information are of the user interest and lead to determine the granularity level of aspects to be filtered. Step C: The filtered aspects 214 are then utilized for document-aspect summarization, where each document-aspect pair results in a short summarization based on either an extract-then- abstract summarization module 216 or an abstractive summarization module 218 applying the respective summarization methods. In the case of the former, a term frequency (TF) - inverse dense frequency (IDF) method that is based on sentence or paragraph level rather than the typical token level may be used, as an example. This will allow determining the overall topic of a paragraph to control the summary focus for extraction. Following this, a large language model (LLM) 220 is used to improve the fluency of the extractive summary. In embodiments, the survey summarization of Step C may be performed by a module such as survey summarization module 224. This step generates a summary of each document that is in the form of separate summaries for each aspect extracted from the document.Step D: The overview of survey summarization triggers the process of flexible query or domain-specific needs to generate summarization for both selected aspects and documents by a cross-document aspect-based summarization module 222. This step aggregates the results from Step C and multiple documents and generates a single coherent summary comparing the documents, separated by aspect. The database 226represents an ability for users to select or otherwise interact with summaries from different aspects and documents.

[0037] Compared to standard aspect-based summarization technology, systems implemented according to embodiments of the present disclosure provide for the following improvements:1. Handling the documents comparisons even when sections segmentation is lacking.2. Allowing flexible selection of aspects for generating summaries to compare different documents.3. Enabling generation of summarizations tailored to the needs of the user. This means that the system can adapt to more versatile and customizable than standard aspect-based summarization methods.

[0038] Embodiments of the present invention thus provide for general improvements to computers in machine learning systems to provide for improved aspect-based document summarization. Moreover, embodiments of the present invention can be practically applied to use cases to effect further improvements in a number of technical fields including, but not limited to, medical (e.g., digital medicine, personalized healthcare, Al-assisted drug or vaccine development, cyberthreat detection, material 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, for example, to the field of summarization for knowledge work, such as those used for consulting support, contact center support, or cyberthreat intelligence. An exemplary use case includes, given at least one input document, generating a summary based on automatically discovered aspects. For example, given an input of at least one document, the at least one document(s) can be shortened and create an adaptable summary, which is based on: (1) the discovered aspects, (2) the aspects deemed most important, (3) similarities / differences between documents if more than one input document is supplied. The output of the exemplary use case may include a summary with different aspects. As one example of an automated action or resulting physical change (technicity) is the automatic generation of a document that includes a summary with different aspects. The summary generated by the features of the present disclosure can be used as input for further Al systems or displayed for further interaction by users.

[0040] Embodiments of the present invention can be applied, for example, to the field of medical Al or automated healthcare, for example to generate patient history summaries. An exemplary use case includes, given at least one patient report from the patient’s history, producing a summary where different aspects are considered. Aspects could for example be different categories such as prior surgeries, prior medication, prior conditions, etc. For example,given an input such as at least one patient report, the at least one patient report can be shortened and create an adapted summary, which is based on: (1) the discovered aspects, (2) the aspects deemed most important, (3) similarities / differences between documents if more than one input document is supplied. An output of such an exemplary use case could include a summary with different aspects and differences / similarities highlighted, which for example provides the basis for a diagnosis to be made. One example of an automated action or a physical change (technicity) is the selection of one or more medications or treatments by a healthcare professional based on the summary with the different aspects and differences / similarities highlighted.

[0041] Embodiments of the present invention can be applied, for example, to the field of automated public safety, for example for generating citizen report summaries. An exemplary use case includes, given at least one citizen report, producing an aspect based summary to a government worker (e.g. citizen could be seeking employment in an employment agency and government worker could be the citizen’s case worker). For example, given an input of at least one citizen report, the at least one citizen report can be shortened and create an adapted summary, which is based on: (1) the discovered aspects, (2) the aspects deemed most important, (3) similarities / differences between documents if more than one input document is supplied. An output of such an exemplary use case could include a summary with different aspects and differences / similarities highlighted, which provides the basis for the government worker makes a decision. An example of an automated action or a physical change (technicity) is the generation and execution of orders to fix a broken street sign / light based on the summary with different aspects and differences / similarities highlighted. The generation of such orders may include prioritizing certain areas of a city that include multiple and similar aspects which correspond to complaints for the certain areas.

[0042] Embodiments of the present invention can be applied, for example, to the field of automated research, for example, to generate research article or patent summaries. An exemplary use case includes, given at least one patent or patent application, produce an aspectbased summary. This saves the time of the patent reader and allows them to draw insights that the might have missed otherwise. For example, given an input of at least one patent or patent application, the at least one patent or patent application can be shortened and create an adaptable summary, which is based: on (1) the discovered aspects, (2) the aspects deemed most important, (3) similarities / differences between documents if more than one input document is supplied. An output of such an exemplary use case could include a summary with different aspects and differences / similarities highlighted, which provides the basis for the user to make a final decision (e.g. whether the patent or patent application is relevant with regards to another patent).In one example of an automated action or a physical change (technicity) is the creation of the patent application / document based on an analysis of whether the patent or patent application is relevant with regards to another patent, such as in a prior art comparison to potential novel ideas.

[0043] Embodiments of the present invention can be applied, for example, to the field of automated or Al -assisted law enforcement, for example for integration in a forensic tool or for generating suspect report summaries. An exemplary use case includes, given at least one suspect report, producing an aspect-based summary to a police worker. This saves the time of the police worker and allows them to draw insights that they might have missed otherwise. For example, given an input of at least one suspect report, the at least one suspect report can be shortened and create an adapted summary, which is based on: (1) the discovered aspects, (2) the aspects deemed most important, (3) similarities / differences between documents if more than one input document is supplied. An output of such an exemplary use case could include a summary with different aspects and differences / similarities highlighted, which provides the basis for the policer worker to make a final decision. One example of an automated action or a physical change (technicity) is the filing of legal actions and police investigations against suspects that include similarities across multiple documents (e.g., similar modus operand!). Automated actions could also include activating sensors or forensic tools, such as cameras or area-monitoring equipment. Other suspects may be disregarded as they do not include similar features across analyzed documents.

[0044] Embodiments of the present invention can be applied, for example, to the field of automated customer service or call handling, for example for generating call center summaries. An exemplary use case includes, a situation where after answering a call in a call center for problem shooting, the call handler has to write a summary. A system implemented according to aspects of the present disclosure can safely automate this process, therefore reducing the time it takes for the call handler to produce the summary. For example, given an input of a transcribed phone call, the transcribed phone call can be shortened and create an adapted summary, which is based on: (1) the discovered aspects, (2) the aspects deemed most important, (3) similarities / differences between documents if more than one input document is supplied. An output of such an exemplary use case could include a summary with different aspects and differences / similarities highlighted, which provides the basis for the user to make a final decision (e.g. whether the phone call is relevant with regards to others and what decision was made for similar instances). One example of an automated action or a physical change (technicity) is the automatic transfer of an incoming call to a correct destination for handling issues that are similar to other issues that were previously analyzed and directed to a similar destination using a transcription of the call compared to previously transcribed calls.

[0045] Embodiments of the present invention can be applied, for example, to the field of AI- assisted legal or contract enforcement, for example for control and compliance of contracts. An exemplary use case includes where lawyers manually check contracts in order to ensure that all information is present in the document according to guideline documents. A system implemented to aspects of the present disclosure can safely automate this process, therefore reducing the time it takes for the lawyer to produce a judgement on the contract. For example, given an input of contract documents, the contract documents can be shortened and create an adapted summary, which is based on: (1) the discovered aspects, (2) the aspects deemed most important, (3) similarities / differences between documents if more than one input document is supplied. An output of such an exemplary use case could include a summary with different aspects and differences / similarities highlighted, which provides the basis for the user to make a final decision (e.g. whether the contract is compliant with the guidelines and what decision was made for similar instances). One example of an automated action or physical change (technicity) is the automatic correction of any errors or omissions in documents based on the summary with different aspects by a computer.

[0046] Embodiments of the present invention can be applied, for example, to the field of AI- assisted human resources, for example to provide resume insights and analysis. An exemplary use case includes situations where while reviewing applicant resumes, a worker has to compare against multiple applicants to find the most suitable. A system implemented according to aspects of the present disclosure can safely automate this process, therefore reducing the time it takes for the call handler to produce a summary of each resume. For example, given an input of an applicant resume, the applicant resume can be shortened and create an adapted summary, which is based on: (1) the discovered aspects, (2) the aspects deemed most important, (3) similarities / differences between documents if more than one input document is supplied. An output of such an exemplary use case could include a summary with different aspects and differences / similarities highlighted, which provides the basis for the user to make a final decision (e.g. whether the applicant can move forward in the application process and what decision was made for similar instances). An example of an automated action or physical change (technicity) is the calling or otherwise advancing of certain applicants in the application process. For example, a computer may generate messages, emails, and hyperlinks which are transmitted to user devices of applicants for proceeding in the application process based on the summary with different aspects.

[0047] Fixed aspect-based summarization, which relies on a predefined list of aspects for generating summaries, limits the capability when dealing with documents lacking clear topic segmentation of topics or originating from multiple resources with varying formats.Additionally, fixed aspects may not adequately address the diverse needs of customers, as different customer roles might consider different importance of aspects when comparing multiple documents. In contrast, automatic discovery of aspects, enables the dynamic identification of relevant aspects, making it highly adaptable to a wide range of content types, accommodating different customer roles, and facilitating application across diverse domains and documents with varying formats.

[0048] FIG. 3 illustrates a flow diagram of a method for aspect-based summarization according to an embodiment of the present invention. Flow diagram 300 includes extracting aspect candidates from documents at 302. In embodiments the documents may be obtained using one or more retrieval techniques that determine a similarity between sentences of each document of the documents. The one or more retrieval techniques may include a semantic similarity computation or syntactic similarity computation. In accordance with an embodiment of the present disclosure, an aspect candidate extraction technique may be implemented to automatically determine the aspect candidates from the documents. The aspect candidate extraction technique may include a rule based method for processing content of the documents to identify headers in the documents or detecting keywords from a dictionary in the documents.

[0049] In FIG. 3, flow diagram 300 includes filtering the aspect candidates by a disambiguation technique that uses different aspect granularity levels at 304. Flow diagram 300 includes extracting informative contents based on the filtered aspect candidates at 306. The flow diagram 300 includes implementing a summarization technique to generate a plurality of document-aspect pair results from the informative contents at 308. In embodiments, the summarization technique includes an extractive-then-abstract summarization technique or an abstractive summarization technique. FIG. 3 includes generating cross-document aspect-based summaries for the plurality of document-aspect pair results at 310.

[0050] Referring to FIG. 4, a processing system 400 can include one or more processors 402, 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. Processing system 400 can be representative of each computing system disclosed herein.

[0051] Processors 402 can include one or more distinct processors, each having one or more cores. Each of the distinct processors can have the same or different structure. Processors 402 can include one or more central processing units (CPUs), one or more graphics processing units (GPUs), circuitry (e.g., application specific integrated circuits (ASICs)), digital signal processors (DSPs), and the like. Processors 402 can be mounted to a common substrate or to multiple different substrates.

[0052] Processors 402 are configured to perform a certain function, method, or operation (e.g., are configured to provide for performance of a function, method, or operation) at least when one of the one or more of the distinct processors is capable of performing operations embodying the function, method, or operation. Processors 402 can perform operations embodying the function, method, or operation by, for example, executing code (e.g., interpreting scripts) stored on memory 404 and / or trafficking data through one or more ASICs. Processors 402, and thus processing system 400, can be configured to perform, automatically, any and all functions, methods, and operations disclosed herein. Therefore, processing system 400 can be configured to implement any of (e.g., all of) the protocols, devices, mechanisms, systems, and methods described herein.

[0053] For example, when the present disclosure states that a method or device performs task “X” (or that task “X” is performed), such a statement should be understood to disclose that processing system 400 can be configured to perform task “X”. Processing system 400 is configured to perform a function, method, or operation at least when processors 402 are configured to do the same.

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

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

[0056] Input-output devices 406 can include any component for trafficking data such as ports, antennas (i.e., transceivers), printed conductive paths, and the like. Input-output devices 406 can enable wired communication via USB®, DisplayPort®, HDMI®, Ethernet, and the like. Input-output devices 406 can enable electronic, optical, magnetic, and holographic, communication with suitable memory 406. Input-output devices 406 can enable wireless communication via WiFi®, Bluetooth®, cellular (e.g., LTE®, CDMA®, GSM®, WiMax®, NFC®), GPS, and the like. Input-output devices 406 can include wired and / or wireless communication pathways.

[0057] Sensors 408 can capture physical measurements of environment and report the same to processors 402. User interface 410 can include displays, physical buttons, speakers, microphones, keyboards, and the like. Actuators 412 can enable processors 402 to control mechanical forces.

[0058] Processing system 400 can be distributed. For example, some components of processing system 400 can reside in a remote hosted network service (e.g., a cloud computing environment) while other components of processing system 400 can reside in a local computing system. Processing system 400 can have a modular design where certain modules include a plurality of the features / functions shown in FIG. 4. For example, I / O modules can include volatile memory and one or more processors. As another example, individual processor modules can include read-only-memory and / or local caches.

[0059] Tan, B., Qin, L., Xing, E., & Hu, Z. (2020, November). Summarizing Text on Any Aspects: A Knowledge-Informed Weakly-Supervised Approach. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP) (pp. 6301-6309), which is hereby incorporated by reference herein, describe abstractive aspect-based summarization using a weak-supervision approach that uses and extracts aspects from only a single document. In contrast, embodiments of the present invention provide for aspect discovery summarization that is versatile, and can accommodate both single and multiple documents. In particular, the aspect disambiguation step enables to extract aspects from diverse documents, subsequently generating summaries based on automatically identified aspects to enhance the decision-making process.

[0060] Wang, Y., Zhou, Y., Wang, M., Chen, Z., Cai, Z., Chen, J., & Leung, V. C. (2023). Multidocument Aspect Classification for Aspect-Based Abstractive Summarization. IEEE Transactions on Computational Social Systems, which is hereby incorporated by reference herein, describe a two-step pipeline, first aspect discovery which is presented as a classification task, and then a summary generator. In contrast to this approach, embodiments of the present invention provide that the input can be one or several documents, and provide for the additional step of retrieving relevant documents that may be helpful in providing a better overview and summary of the document. Moreover, in contrast to this approach, aspect discovery according to embodiments of the present invention is not purely ML based, but also provides to use both automated methods that rely on extracting headers, sections, etc., as well as NLP methods that extract aspects based on the discussed topic in a given section. Further, the later disambiguation steps according to embodiments of the present invention ensure uniformity among the summaries of all documents and enable to provide a comparable summary across the different inputs.

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

[0062] While subject matter of the present disclosure has been illustrated and described in detail in the drawings and foregoing description, such illustration and description are to be considered illustrative or exemplary and not restrictive. Any statement made herein characterizing the invention is also to be considered illustrative or exemplary and not restrictive as the invention is defined by the claims. It will be understood that changes and modifications may be made, by those of ordinary skill in the art, within the scope of the following claims, which may include any combination of features from different embodiments described above.

[0063] The terms used in the claims should be construed to have the broadest reasonable interpretation consistent with the foregoing description. For example, the use of the article “a” or “the” in introducing an element should not be interpreted as being exclusive of a plurality of elements. Likewise, the recitation of “or” should be interpreted as being inclusive, such that the recitation of “A or B” is not exclusive of “A and B,” unless it is clear from the context or the foregoing description that only one of A and B is intended. Further, the recitation of “at least one of A, B and C” should be interpreted as one or more of a group of elements consisting of A, B and C, and should not be interpreted as requiring at least one of each of the listed elements A, B and C, regardless of whether A, B and C are related as categories or otherwise. Moreover, the recitation of “A, B and / or C” or “at least one of A, B or C” should be interpreted as including any singular entity from the listed elements, e.g., A, any subset from the listed elements, e.g., A and B, or the entire list of elements A, B and C.

Claims

CLAIMSWhat is claimed is:

1. A computer-implemented method for aspect-based summarization, the computer- implemented method comprising: extracting aspect candidates from documents; fdtering the aspect candidates by a disambiguation technique that uses different aspect granularity levels; extracting informative contents based on the fdtered aspect candidates; implementing a summarization technique to generate summaries for a plurality of document-aspect pairs from the informative contents; and generating an aspect-based summary based on the summaries.

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

3. The computer-implemented method according to claim 2, wherein the one or more document retrieval techniques include a semantic similarity computation or syntactic similarity computation.

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

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

6. The computer-implemented method of any of the preceding claims, wherein the disambiguation technique filters the aspect candidates to determine common aspects from the aspect candidates.

7. The computer-implemented method of claim 6, wherein the disambiguation technique includes a dictionary-based method or a knowledge-based method.

8. The computer-implemented method of claim 6, wherein the different aspect granularity levels are obtained from a database.

9. The computer-implemented method according to any of the preceding claims, wherein the summarization technique includes an extractive-then-abstract summarization technique or an abstractive summarization technique, wherein the extractive-then-abstract summarization technique is based on a sentence or paragraph level.

10. The computer-implemented method of claim 9, wherein the extractive-then-abstract summarization technique includes a term frequency (TF) - inverse dense frequency (IDF) technique.

11. The computer-implemented method according to any of the preceding claims, wherein the summarization technique includes applying a large language model (LLM) to improve fluency of an extractive summary that corresponds to the plurality of document-aspect pairs.

12. The computer-implemented method according to any of the preceding claims, wherein the documents do not include explicit segmentation of sections.

13. The computer-implemented method according to any of the preceding claims, wherein the aspect-based summary compares the summaries of the plurality of document-aspect pairs.

14. A computer system for interpretable domain adaptation for aspect-based summarization, the computer system comprising one or more hardware processors which, alone or in combination, are configured to provide for execution of the following steps: extracting aspect candidates from documents; filtering the aspect candidates by a disambiguation technique that uses different aspect granularity levels; extracting informative contents based on the filtered aspect candidates; implementing a summarization technique to generate summaries for a plurality of document-aspect pairs from the informative contents; and generating an aspect-based summary based on the summaries.

15. A tangible, non-transitory computer-readable medium having instructions thereon which, upon being executed by one or more processors, provide for aspect-based summarization by execution of the following steps: extracting aspect candidates from documents; filtering the aspect candidates by a disambiguation technique that uses different aspect granularity levels; extracting informative contents based on the filtered aspect candidates; implementing a summarization technique to generate summaries for a plurality of document-aspect pairs from the informative contents; and generating an aspect-based summary based on the summaries.

Images