JP2023507881A - Domain-based text extraction method and system - Google Patents

Abstract

The present disclosure relates to methods and systems for extracting information from the content of input files. The method includes identifying text data from an input file, receiving text input from a user to identify related text entities from a plurality of text entities, and automatically generating search patterns corresponding to the text input. May include doing. The method may further include determining patterns associated with each of the plurality of textual entities, and matching search patterns corresponding to the text input with the patterns associated with the plurality of textual entities. The method includes identifying one or more matching patterns from patterns associated with a plurality of text entities based on the correspondence; and extracting from.
[Selection drawing] Fig. 2

Description

TECHNICAL FIELD This disclosure relates generally to data extraction, and more particularly to methods and systems for extracting textual information from the content of input files using one or more data extraction approaches.

Recently, text extraction techniques have gained importance. For example, extraction techniques such as Optical Character Recognition (OCR) may allow users to extract text data from files such as images or Portable Document Format (PDF) files. Additionally, it may be desirable to extract relevant information and generate a representation using the extracted relevant information.

According to some available techniques, textual information is extracted from the input file if the input file contains tags or if patterns can be identified in the identified text, and the extracted information is used. It may be possible to generate a representation using However, in complex documents where tags may not exist or patterns cannot be identified, it is difficult to extract relevant information or generate representations.

A method and system according to the present invention extract information from the content of an input file to identify text data from the input file and text data from a user to identify related text entities from a plurality of text entities. receiving input; automatically generating a search pattern corresponding to the text input; determining a pattern associated with each of a plurality of text entities; matching patterns associated with a plurality of textual entities; identifying one or more matching patterns from the patterns associated with the plurality of textual entities based on the matching; and one or more matching patterns. extracting related text entities corresponding to from the plurality of text entities. The identified text data includes multiple text entities.

In addition, the problems disclosed by the present application and their solutions will be clarified by the description in the section of the description of the mode for carrying out the invention and the drawings.

The accompanying drawings, which are incorporated in and constitute a part of this disclosure, illustrate exemplary embodiments and, together with the description, serve to explain the principles of the disclosure.

1 is a functional block diagram of an exemplary system for extracting information from content of an input file, in accordance with some embodiments of the present disclosure; FIG. 2 is a functional block diagram of a Natural Language Processing (NLP) metadata extraction framework, according to some embodiments of the present disclosure; FIG. 1 is a block diagram of a spell checking system, according to some embodiments of the present disclosure; FIG. 1 is a block diagram of a recommendation system, according to some embodiments of the present disclosure; FIG. 1 is a block diagram of a metadata update system, according to some embodiments of the present disclosure; FIG. 4 is a flowchart of a method for extracting information from the contents of an input file, according to various embodiments of the present disclosure; 4 is a flowchart of a method for extracting information from the contents of an input file, according to various embodiments of the present disclosure; 4 is a flowchart of a method for extracting information from the contents of an input file, according to various embodiments of the present disclosure; 4 is a snapshot of an exemplary input file with noisy entities from which information is to be extracted, according to some embodiments of the present disclosure; 5 is a snapshot of another exemplary input file from which information is extracted, in accordance with various embodiments of the present disclosure;

Exemplary embodiments are described with reference to the accompanying drawings. Where convenient, the same reference numbers will be used throughout the drawings to designate the same or like parts. While examples and features of the disclosed principles are described herein, modifications, variations, and other implementations are possible without departing from the spirit and scope of the disclosed embodiments. It is intended that the following detailed description be considered exemplary only, with a true scope and spirit being indicated by the following claims.

Referring now to FIG. 1, illustrated is a functional block diagram of an exemplary system 100 for extracting information from content of an input file, according to some embodiments of the present disclosure. System 100 may include a text identification device 102 . In some embodiments, text identification device 102 may identify text data from the content of input files. For example, the input file may be an image or a Portable Document Format (PDF) file. The identified text data may include multiple text entities. Accordingly, in some embodiments, the text identification device 102 may employ Optical Character Recognition (OCR) techniques to identify text data from input files. In alternative embodiments, text identification device 102 may use any other technique known in the art to identify text data from input files.

System 100 may further include an information extraction device 104 in communication with text identification device 102 . For example, the information extraction device 104 may be configured to extract relevant information from multiple textual entities identified from the input file by the text identification device 102 . In some embodiments, to extract information, the information extraction device 104 receives text input from a user to identify related text entities from a plurality of text entities and automatically generates search patterns corresponding to the text input. may be generated. The information extraction device 104 may further determine patterns associated with each of the plurality of textual entities, and match search patterns corresponding to the text input with the patterns associated with the plurality of textual entities. The information extraction device 104 further identifies one or more matching patterns from patterns associated with the plurality of text entities based on the correspondence, and extracts associated text entities corresponding to the one or more matching patterns from the plurality of texts. May be extracted from an entity.

System 100 may further include display 114 . In some embodiments, information extraction device 104 may include one or more processors 110 and computer-readable media (eg, memory) 112 . Computer readable storage medium 112 contains instructions that, when executed by one or more processors 110, cause one or more processors 110 to generate a representation based on the contents of an input file, in accordance with aspects of this disclosure. can be stored. Computer-readable storage medium 112 also stores various data (eg, identified text data, value entities, data, pattern entity data, etc.) that may be captured, processed, and/or required by system 100. good. System 100 may interact with a user via user interface 116 accessible via display 114 . System 100 may also interact with one or more external devices 106 via communication network 108 to send or receive various data. External device 106 may include, but is not limited to, a remote server, digital device, or another computing system.

Referring now to FIG. 2, a functional block diagram of a Natural Language Processing (NLP) metadata extraction framework 200 according to some embodiments of the present disclosure is shown. The NLP metadata extraction framework 200 may be implemented in the system 100 and specifically in the information extraction device 104 . The NLP metadata extraction framework 200 may include various modules that perform various functions to extract information from the content of input files. In some embodiments, extraction engine 202 may extract/identify text data from the content of input file 204 . The input file 204 may be a flat file 206, or a PDF file 208, or a database file 210, or an OCR output 212.

Extraction engine 202 may also receive one or more domain rules from domain rules database 214 . Further, in some embodiments, the domain rules database 214 may be based on "if-then" rules (as opposed to traditional procedural code). Note that domain rules may be used to determine the domain-based name associated with each of multiple textual entities. The domain-based name may be determined based on associating each of the plurality of textual entities with a list of domain-based names stored in a data repository (not shown in FIG. 1). For example, if the identified data includes text entities belonging to a PIN code for a city in a country, the domain-based database may, for example, associate the PIN code with a list of domain-based names stored in the data repository. , the domain name, ie the name of the city that has that PIN code.

Extraction engine 202 may also receive one or more location rules from location rules database 216 . Note that in some embodiments, location rules may be used to determine locations associated with entities from text data in an input file based on OCR output. The position may be determined when first identifying the type associated with the input file and determining the position of one or more associated textual entities in the input file based on the type associated with the input file. Once the locations of the associated text entities are determined, the associated text entities may be extracted from the input file based on location. The location rule(s) may be more relevant for input data that includes tables, i.e. input data that is arranged in tabular form, but the location rule(s) may be any Note that it may be equally applicable to documents with other types of input data, eg free text.

As an example, the input file type may be of an "Aadhar" card. It should further be appreciated that in a document such as an "Aadhar" card, data may exist in a particular format. Accordingly, the location of the relevant textual entity may be predetermined based on a template stored in the data repository and corresponding to the type associated with the input file. For example, the location of the textual entities "name" and "address" (of the person of interest) may be in the middle of the document, and this information may be available and stored in the system 100 database. In particular, the position may be based on (x,y) coordinates. In the example above, the location of the text entities "name" and "address" may be within the region defined by (x,y) coordinates (5,10) to (15,15). Thus, when the type of document is identified as an "Aadhar" card, system 100 may access the location and extract the relevant textual entities (ie, "name" and "address") from the location.

Extraction engine 202 may also receive one or more first NLP rules from first NLP rule database 218 . Note that the first NLP rules database 218 may create a Regular Expression (Regex) based on the NLP rules provided for the correct value/attribute. As those skilled in the art will appreciate, a Regex may be a special string to describe a search pattern.

Extraction engine 202 may also receive one or more second NLP rules from second NLP rule database 220 . Note that in some embodiments, the second NLP rules database 220 may define parts of speech (POS) for correct values/attributes. For example, POS associated with multiple textual entities may be identified. As will be appreciated by those skilled in the art, POS may be one of nouns, pronouns, verbs, adverbs, adjectives, conjunctions, prepositions, and interjections. Upon identifying POS, one or more unique text entities may be selected from multiple text entities for which POS is a noun. One or more (selected) unique textual entities where POS is identified as a noun may be extracted as relevant data.

For example, a payslip document (input data) may contain a number of text entities including the name of the person of interest. Extracting names may be the purpose of the extraction engine 202 . Further, it will be appreciated that names belong to the POS category of nouns. Therefore, textual entities where POS is a noun will be selected. Therefore, the required name will be extracted from the selected entity.

In some embodiments, when an attempt is made to identify POS associated with multiple textual entities, one or more unique textual entities for which no POS is identified may be identified. An entity type associated with each of the one or more unique textual entities may be determined. An entity type may be a value entity or a pattern entity. Each value entity may have an associated value and each pattern entity has an associated pattern.

For example, an identified textual entity may consist of a combination of text characters that indicate a codename such as an email ID or oil field name (eg, "john@gmail.com" or "AB-123", respectively). Therefore, POS may not be identified for such text entities. Accordingly, such text entities may be identified as either value entities or pattern entities. Thus, the email ID ("john@gmail.com") may be determined as the value entity, where the value is the email ID. This may be because email ID patterns are usually typical and known patterns, and all email IDs may share the same or similar patterns. On the other hand, a text entity such as "AB-123" corresponding to an oil field name may be determined as a pattern entity. This is because while field names may share the same pattern, such patterns may not be commonly known, i.e., if the NLP metadata extraction framework 200 does not pre-store such patterns. This is because Upon identifying a value entity, text entities in the input file that have an associated value that is the same as the value entity's associated value may be automatically identified. In other words, all email ID text entities may be identified and extracted.

Upon identifying a pattern entity, search patterns corresponding to patterns (Regex) associated with the pattern entity may be automatically generated. Additionally, patterns associated with each of the plurality of text entities in the input file may be determined. A search pattern (Regex) corresponding to a pattern entity may be matched with the pattern associated with the multiple text entities, and based on the matching, one or more of the patterns from the patterns associated with the multiple text entities. A matching pattern may be determined. Matching text entities corresponding to one or more matching patterns may be extracted from multiple text entities.

For example, for the text entity "AB-123" corresponding to an oil field, the search pattern (Regex) "apd" (where "a" means one or more alphabets and "p" means one or more punctuation marks, and "d" means one or more digits) may be automatically generated. Additionally, patterns associated with other textual entities in the input file may be determined, the search pattern "apd" may be matched with patterns associated with other textual entities, and matching patterns may be identified. All text entities corresponding to the matching pattern are then extracted. This may extract all oil field names in the input file.

In some embodiments, once text data from an input file is identified, a first NLP rule receives text input from a user to identify related text entities from a plurality of text entities, and responds to the text input. automatically generate a search pattern for The first NLP rule may further cause a pattern associated with each of the plurality of text entities to be determined and a search pattern corresponding to the text input to be matched with the patterns associated with the plurality of text entities. The first NLP rule further identifies one or more matching patterns from the patterns associated with the plurality of text entities based on the matching, and identifies the associated text entities corresponding to the one or more matching patterns as a plurality of matching patterns. Extracting from text entities may be done.

For example, a user may provide a text input "AB-123" to identify a related text entity from multiple text entities extracted from an input file. A search pattern (Regex) “apd” may be identified for the entered text entity “AB-123”. Additionally, patterns associated with other textual entities in the input file may be determined, a search pattern (Regex) "apd" may be matched with the patterns associated with the other textual entities, and matching patterns may be identified. be. All text entities corresponding to the matching pattern may then be extracted. This may extract all oil field names in the input file.

Note that in some embodiments, text input from a user may be received in the form of entries in a Microsoft Excel® sheet. One or more rules may be defined that may automatically generate search patterns (Regex) that correspond to text input. One or more rules then determine patterns associated with each of the plurality of text entities of the input file, match search patterns corresponding to the text input with patterns associated with the plurality of text entities, and One or more matching patterns may be identified from the patterns associated with the plurality of textual entities based on the attachment, and related textual entities corresponding to the one or more matching patterns may be extracted from the plurality of textual entities.

Thus, in some embodiments, Microsoft Excel® data may be received from one or more Microsoft Excel® sheets 222 . Various rules (domain rules, NLP rules, location rules, etc.) may be applied to obtain extracted data 226 . In some embodiments, a "Microsoft Excel®" macro may be used for extraction. Further, in some embodiments, the same macros may be embedded in Python (language; registered trademark). Here, extracted data 226 may refer to related text entities extracted from multiple text entities or representations that may be generated based on related text entities extracted using various rules. . In some additional embodiments, when the above approaches, i.e., domain-based, location-based, POS-based, and Regex-based approaches fail to provide text extraction, Machine Learning (ML)-based approach may be used. An artificial intelligence (AI) model (or ML model) 224 may be used for ML-based approaches.

Referring now to FIG. 3, shown is a block diagram of a spell checking system 300 according to some embodiments of the present disclosure. In some embodiments, textual entities extracted by extraction module 302 (corresponding to text identification device 102 ) may be received by spell checking system 300 . These expressions generated by extraction module 302 may serve as input text 304 to spell checking system 300 . Input text 304 may be preprocessed by preprocessing module 306 .

Once pre-processing is performed, cosine similarity analysis may be performed on the pre-processed data by cosine similarity module 308 . As an example, cosine similarity module 308 may perform cosine similarity analysis using a threshold. A minimum edit distance module 310 may perform a minimum edit distance analysis on data received from cosine similarity module 308 . In some embodiments, minimum edit distance analysis may be performed on filtered words.

A maximum cosine similarity module 312 may perform a maximum cosine similarity analysis on the data received from the minimum edit distance module 310 . Note that the maximum cosine similarity analysis may be based on words within the minimum edit distance words. A replacement module 314 may replace incorrect words with corrected words based on the analysis performed by the maximum cosine similarity module 312 .

Referring now to FIG. 4, a block diagram of a recommendation system 400 is shown in accordance with some embodiments of the present disclosure. For example, recommendation system 400 may be used to identify entity types, ie, value entities or pattern entities. In some embodiments, recommendation system 400 may use classification-based machine learning (ML) models 402 . The recommendation system 400 may receive input data (NLP extraction and spell check data) 406 . Note that the input data can be NLP and spell checked data. Recommendation system 400 may include a config file 408 . The config file 408 may trigger determining whether values or patterns may be recommended for the input data.

Recommendation system 400 may provide input data to ML model 402 . ML model 402 may use historical data from archive database 404 . Based on historical data, ML model 402 may provide predictive data 410 or may provide recommended data 412 . Thus, recommended values 414 may be generated based on prediction data 410 or recommended patterns 416 may be generated based on recommended data 412 .

Referring now to FIG. 5, a block diagram of metadata update system 500 is shown in accordance with some embodiments of the present disclosure. In some embodiments, metadata update system 500 may receive NLP extracted data 502 . Metadata update system 500 may include system date and time module 504 . Metadata update system 500 may further include confidence module 506 and probability module 508 . Confidence module 506 may determine a confidence score associated with the accuracy of the extracted data and the representations generated by the NLP rules. Probability module 508 may determine accuracy scores associated with the accuracy of the extracted data and representations generated by the NLP rules based on the confidence scores. In other words, the probability module 508 may determine a probability of how accurate/useful the extracted data or the representation generated using the NLP rules is.

If the probability determined by the probability module 508 is high enough (ie, greater than a threshold), the extracted data and the representation generated using the NLP rules, ie, the rule-based value 514, may be provided. On the other hand, if the probability determined by the probability module 508 is not sufficiently high (i.e., less than the threshold), a representation generated using the extracted data and the ML model, i.e., machine learning (ML) value 516 is provided. you can As previously mentioned, ML values may be generated by ML models using historical data that may be stored in archive database 510 . The extracted data and generated representations may be stored in final database 512 .

Referring now to FIG. 6, a flowchart of a method 600 for extracting information from the content of an input file is shown, according to one embodiment of the present disclosure. As will be appreciated, method 600 may provide a Regex-based method of text extraction. At step 602, text data may be identified from the input file. The identified text data may include multiple text entities. Note that in some examples, the input files may include at least one of image files and Portable Document Format (PDF) files.

At step 604, text input may be received from a user to identify related text entities from a plurality of text entities. For example, a user may provide text as an entry in a Microsoft Excel® sheet. At step 606, a search pattern (Regex) corresponding to the text input may be automatically generated. At step 608, patterns associated with each of the plurality of text entities may be determined. At step 610, a search pattern corresponding to a text input may be matched with patterns associated with a plurality of text entities. At step 612, one or more matching patterns from patterns associated with multiple text entities may be identified based on the correspondence. At step 614, related text entities corresponding to one or more matching patterns may be extracted from the plurality of text entities.

Additionally, in some embodiments, upon extracting related text entities corresponding to one or more matching patterns from the plurality of text entities, a representation may be generated using the extracted related text entities.

Additionally, a machine learning (ML) model may be used to extract related entities, for example, if the method 600 fails to extract related entities. In other words, if the data (relevant textual entities) cannot be extracted, the ML-based approach may start.

Referring now to FIG. 7, shown is a flowchart of a method 700 for extracting information from the content of an input file, according to another embodiment of the present disclosure. At step 702, text data may be identified from the input file. The identified text data may include multiple text entities. Note that the input files may include at least one of image files and Portable Document Format (PDF) files.

At step 704, domain-based names associated with each of the plurality of textual entities may be determined. The domain-based name may be determined based on associating each of the plurality of textual entities with a list of domain-based names stored in the data repository. At step 706, a type associated with the input file may be identified. At step 708, the location of one or more related textual entities in the input file may be determined based on the type associated with the input file. At step 710, one or more relevant textual entities may be extracted from the input file based on location.

At step 712, parts of speech (POS) associated with the plurality of text entities may be identified. It should be appreciated that POS may be one of nouns, pronouns, verbs, adverbs, adjectives, conjunctions, prepositions, and interjections. At step 714, one or more textual entities for which POS is identified as a noun may be determined.

At step 716, one or more unique textual entities for which no POS is identified may be selected. At step 718, an entity type associated with each of the one or more unique text entities may be determined. An entity type may be one of a value entity and a pattern entity. Furthermore, each value entity may have an associated value and each pattern entity has an associated pattern. As described in connection with FIG. 4, a (ML-based) recommendation system 400 may be used to determine the entity type associated with each of the one or more unique textual entities.

At step 720, for each value entity, one or more text entities having an associated value that is the same as the value entity's associated value may be automatically identified. Steps 722-728 may be performed for each pattern entity. Accordingly, at step 722, a search pattern (Regex) corresponding to the pattern associated with the pattern entity may be automatically generated. At step 724, a search pattern corresponding to the pattern entity may be matched with patterns associated with multiple text entities. At step 726, one or more matching patterns may be determined from the patterns associated with the plurality of text entities based on the matching. At step 728, matching text entities corresponding to one or more matching patterns may be extracted from the plurality of text entities.

Note that method 700 may be performed in combination with method 600. In an exemplary scenario, method 600 may begin after step 710 of method 700 is completed but before step 712 of method 700 begins.

Referring now to FIG. 8, shown is a flowchart of a method 800 for extracting information from the content of an input file, according to another embodiment of the present disclosure. At step 802, text data may be extracted from the contents of the input file. The extracted text data may include multiple text entities.

At step 804, domain-based names associated with each of the plurality of textual entities may be determined. The domain-based name may be determined based on associating each of the plurality of textual entities with a list of domain-based names stored in the data repository. For example, if the PIN code is retrievable, the city is determined from the PIN code using a domain-based approach, e.g., by matching the PIN code with a list (of PIN codes and associated city names) stored in a knowledge base. may be Similarly, if a city name (eg Delhi) is identified, a country name (eg India) may be determined. In some embodiments, a tag may be assigned to each of the second plurality of entities accordingly.

At step 806, a check may be made to see if the domain-based name was successfully resolved. If the domain-based name is successfully determined, the method may proceed to step 822 (“Yes” path), at which time the identified textual entity may be extracted. Additionally, a representation may be generated using the extracted text entities. In step 806, if the domain-based name is not determined, the method may proceed to step 808 (“No” path). In other words, if the domain-based name cannot be identified, the method may proceed to perform the next alternative approach.

At step 808, a value for a given field may be determined based on the position of each of the plurality of text entities. The position may be in the form of (x and y) coordinates. As an example, the locations may be predetermined based on templates stored in the data repository. For example, in some scenarios, the value of a particular name may be written next to that name or under a particular name header, which may be identifiable based on (x and y) coordinates. For example, in an "Aadhar card", or a "PAN card", or a driver's license, the name of an entity, eg a user, may be obtainable at a particular location, from which the name can be identified. The location technique can be particularly useful for tables. This is because the data in the table is available in a structured format so that entities can be easily and accurately located. Therefore, the type associated with the input file may first be identified, ie whether the input file is an "Aadhar card" or a "PAN card" or a driver's license. Additionally, the location of one or more associated textual entities in the input file may be determined based on the type associated with the input file.

At step 810, a check may be made to see if it was successful in locating one or more associated text entities in the input file. If the location was successfully determined, method 800 may again proceed to step 822 (“Yes” path), at which time one or more related text entities from the input file are extracted based on location. and a representation may be generated using the extracted text entities. If it is determined that the position is not determined, at step 810 the method 800 may proceed to step 812 .

At step 812, a Regex-based approach may be attempted on the input file to extract relevant textual entities. To this end, text input may be received from a user to identify related text entities from a plurality of text entities, and search patterns corresponding to the text input may be automatically generated. In other words, a Regex may be generated. Further, a pattern associated with each of the plurality of text entities may be generated, a search pattern (Regex) corresponding to the text input may be associated with the pattern associated with the plurality of text entities, and the correspondence may include: Based on this, one or more matching patterns from patterns associated with multiple text entities may be identified. In some embodiments, the method 800 may include calculating probabilities of determining related entities from the extracted text data corresponding to Regex.

At step 814, a check may be made to determine if the Regex-based approach was successfully applied. If the Regex-based approach is successfully applied, method 800 may proceed to step 822 (“Yes” path), at which time associated text entities corresponding to one or more match patterns are multiple text Entities may be extracted, and expressions may be generated using the extracted textual entities. If the Regex-based approach is not successful, method 800 may proceed to step 816 (“No” path).

At step 816, a part-of-speech (POS) based approach may be tried. For example, a POS associated with each of a plurality of textual entities may be determined. In other words, at step 816 an attempt may be made to determine which POS the entity may be associated with, eg, whether the POS may be a noun, an adverb, an adjective, or the like. Additionally, at step 816, an attempt may be made to determine the entity type associated with each of the one or more unique text entities. For this purpose, one or more unique textual entities with no POS identification may be selected. For example, a POS may not be identified for a unique entity (such as the email ID or field name mentioned above) that has a combination of alphabets, numbers, punctuation marks, and the like. For such text entities, entity types associated with each of the one or more unique text entities may be determined. An entity type may be one of a value entity and a pattern entity. Each value entity may have an associated value and each pattern entity may have an associated pattern. Further, for each value entity, an attempt may be made to automatically identify one or more text entities that have the same associated value (eg, email ID) as the value entity's associated value.

A check may be made at step 818 to determine if the attempt was successful (ie, if the POS-based approach was successful). If the attempt is deemed successful, method 800 may proceed to step 822 (“Yes” path), at which point one or more texts having associated values that are the same as the associated value of the value entity. Entities (related text entities) may be extracted and representations may be generated using the extracted text entities.

Further, for each pattern entity, automatically generate a search pattern corresponding to the pattern associated with the pattern entity, match the search pattern corresponding to the pattern entity with the patterns associated with multiple text entities, and An attempt may be made to determine one or more matching patterns from patterns associated with a plurality of text entities based on. A check may be made at step 818 to determine if the attempt was successful (ie, if the POS-based approach was successful). If the attempt is deemed successful, method 800 may proceed to step 822 (“Yes” path), at which matching text entities corresponding to one or more matching patterns are extracted from the plurality of text entities. and a representation may be generated using the extracted text entities. In step 818, if the POS-based approach is determined to be unsuccessful, method 800 may proceed to step 820 (“No” path).

At step 820, an attempt may be made to extract relevant text entities (information) from multiple text entities using a machine learning (ML) model. It will be appreciated that the ML model may first be trained to extract the types of relevant textual entities. Therefore, based on training the ML model, ML-based classification may be applied to the input files to extract relevant textual entities. Accordingly, at step 822, relevant textual entities may be extracted based on the ML-based taxonomy, and representations may be generated using the extracted textual entities.

In some embodiments, the various steps (i.e. steps 804, 808, 812, 816, and 820) may be performed in the order described above or may be performed in any other order. Please note that

Further, note that NLP-based approaches may be used in the scenarios below.
(i) If the search key (ie text input from the user) is available and its associated text extraction (ie related entity) is expected in the input file.
(ii) if the search key (ie text input from the user) is not available, but the associated text extraction (ie related entity) is expected in the input file.
(iii) if the search key (ie text input from the user) is available but the associated text extraction (ie related entity) is not expected in the input file.

Additionally, note that a confidence score (probability score) may be calculated for each approach. For example, for a "domain-based" approach, the highest probability score (e.g., 1) may be obtained when the search key (i.e., text input from the user) exactly matches the remaining text entities in the input file. . Further, in a "location-based" approach, the probability score (p) may be based on the probability of matching the search key (w) and the identity of the location (l), ie, p=w*l. For the Regex-based approach, the probability for location identification can be either 0 or 1, and the probability score (e) for extraction can be 1 or 0. Further, there may be multiple (n) values to be extracted and one of those values may be selected. Therefore, the probability of the selected value may be "1/n" and the confidence score may be e*(1/n). If both "location-based" and "Regex-based" approaches are used, each approach may be given 50% weight. Therefore, the confidence score may be (w*l)/2+(e*(1/n))/2. For ML-based approaches, a probability score may be derived based on the confusion matrix (accuracy rate).

Additionally, the following approaches are used: a pure word matching approach, a rule-based approach (i.e. one or more of domain-based, location-based, POS-based, and Regex-based approaches), and an ML-based approach. If so, a composite probability score (P) may be calculated.
therefore,
Composite Probability Score (P)=(P1*0.3)*(P3*0.7), or Composite Probability Score (P)=(P1*0.3)*(P2*0.7)
During the ceremony,
P1 = probability score for pure word matching approach,
P2 = probability score for rule-based approach, and P3 = probability score for ML-based approach.

<Case scenario 1>
Referring now to FIG. 9, a snapshot of an exemplary input file 900 with noisy entities 902 present in the input file 900 is shown. For example, a first portion of an entity may be illegible, such as smudged or smudged, and a second portion of the entity may contain the word "United." Thus, the actual entity could be "United Airlines" or "United States" and so on. To determine this entity, one or more approaches (ie domain-based, location-based, POS-based, Regex-based, or ML-based) may be used to determine the actual entity. For example, if domain-based, location-based, POS-based, and Regex-based approaches fail to provide extraction, an ML-based approach may start, using classification-based machine learning (ML) models to Text extraction may be provided based on the data. ML models may be deterministic and/or probabilistic models.

<Case scenario 2>
Referring now to Figure 10, a snapshot of the input file "Local Order" 1000 is shown from which various textual entities may need to be extracted. As shown in Figure 10, the input file "Local Order" 1000 has a tabular structure.

For example, to extract the value for the attribute "RO Number", "RO Number" may be extracted from the domain dictionary database (lookup table) 1002 by mapping the attribute of RO date.

To extract the attribute Part Number, the key "Part Number" present in the input file 1000 may be identified or received as user input. A location-based approach may then determine the possible locations of the Part Number using the template information stored in the database. For example, a position-based approach may suggest that the Part Number may be on the “right” side of the input file 1000 . Therefore, the required Part Number '32145643' may be extracted from the right cell coordinates of the tabular structure of the input file 'Local Order' 1000 .

A Regex-based approach may be used to extract the email ID attributes present in the input file 1000 . Accordingly, a key (Regex) associated with the attributes of the email may be identified or received from a database. Therefore, Regex is "apapapa" (where "a" means one or more alphabetic characters and "p" means one or more punctuation marks) corresponding to the email ID "ajay.thakur@gmail.com". and "d" means one or more digits). Thus, a text entity with a matching pattern (Regex), namely the email ID text entity, may be identified and the value "ajay.thakur@gmail.com" extracted from the corresponding cell of the tabular structure of input file 1000. you can

A POS-based approach may be used to extract the attribute Name from the input file 1000 . For example, first the keys that exist for the attribute (Name) may be identified in the input file 1000 . The POS of the identified text entities in the input file 1000 may then be determined using, for example, Named Entity Recognizer (NER) taggers. A NER tagger specified in the database for the attribute Name may be retrieved. Therefore, the text entity corresponding to the attribute Name (“Ajay Thakur”) with the noun POS may be extracted. Therefore, an expression using the extracted text entity "Ajay Thakur" may be generated.

One or more text extraction techniques are disclosed above for extracting information from the content of input files. The techniques described above provide one or more advantages over conventional techniques. For example, these techniques provide various approaches to text extraction: domain-based, location-based, POS-based, Regex-based, and ML-based approaches. Thus, text extraction can be performed using a single approach, or multiple approaches applied in any order. In addition, these techniques allow representations to be generated using the extracted information. Furthermore, these techniques are capable of extracting information and generating representations using the extracted information, even in complex documents where tags are absent or patterns cannot be discerned in the extracted information. and

One or more computer-readable storage media may be utilized in implementing embodiments consistent with this disclosure. Computer-readable storage medium refers to any type of physical memory in which information or data readable by a processor may be stored. Thus, a computer-readable storage medium may store instructions for execution by one or more processors, including instructions for causing a processor to perform steps or stages consistent with embodiments described herein. . The term "computer-readable medium" is to be understood to include tangible items and exclude carrier waves and transients, ie, non-transitory. Examples include random access memory (RAM), read only memory (ROM), volatile memory, non-volatile memory, hard drives, CD-ROMs, DVDs, flash drives, disks, and any other known physical storage medium. is mentioned.

It is intended that the disclosure and examples be considered as exemplary only, with a true scope and spirit of the disclosed embodiments being indicated by the following claims.

Claims (16)

A method for extracting information from the content of an input file, comprising:
identifying text data from the input file, wherein the identified text data includes a plurality of text entities;
receiving text input from a user to identify related text entities from the plurality of text entities;
automatically generating a search pattern corresponding to the text input;
determining a pattern associated with each of the plurality of textual entities;
associating the search pattern corresponding to the text input with patterns associated with the plurality of text entities;
identifying one or more matching patterns from the patterns associated with the plurality of textual entities based on the correspondence;
extracting from the plurality of text entities related text entities corresponding to the one or more matching patterns. 2. The method of claim 1, comprising generating a representation using the extracted relevant text entities. 2. The method of claim 1, wherein the input file comprises at least one of an image file and a Portable Document Format (PDF) file. further comprising determining a domain-based name associated with each of the plurality of textual entities;
the domain-based name is determined based on associating each of the plurality of textual entities with a list of domain-based names stored in a data repository;
The method of claim 1. identifying a type associated with the input file;
determining locations of one or more associated textual entities in the input file based on the types associated with the input file;
2. The method of claim 1, further comprising: extracting the one or more related textual entities from the input file based on the location. 3. The position is based on (x,y) coordinates, the position being predetermined based on a template stored in the data repository corresponding to a type associated with the input file. 5. The method described in 5. identifying a part of speech (POS) associated with the plurality of textual entities;
extracting one or more textual entities that the POS identifies as nouns;
said POS is one of nouns, pronouns, verbs, adverbs, adjectives, conjunctions, prepositions, and interjections;
The method of claim 1. identifying one or more unique textual entities for which no POS is identified;
determining an entity type associated with each of the one or more unique textual entities;
the entity type is one of a value entity and a pattern entity, each value entity having an associated value and each pattern entity having an associated pattern;
8. The method of claim 7. For each value entity,
9. The method of claim 8, further comprising: automatically identifying one or more text entities that have the same associated value as the associated value of the value entity. For each pattern entity,
automatically generating search patterns corresponding to patterns associated with the pattern entity;
associating the search pattern corresponding to the pattern entity with patterns associated with the plurality of text entities;
determining one or more matching patterns from the patterns associated with the plurality of textual entities based on the correspondence;
9. The method of claim 8, further comprising: extracting matching text entities corresponding to the one or more matching patterns from the plurality of text entities. 2. The method of claim 1, comprising using a classification-based machine learning (ML) model to generate recommendations based on historical data. A system for extracting information from the content of an input file, comprising:
a processor;
a memory communicatively coupled to the processor;
The memory stores processor-executable instructions that, when executed by the processor, cause the processor to:
identifying text data from the input file, wherein the identified text data includes a plurality of text entities;
receiving text input from a user to identify related text entities from the plurality of text entities;
automatically generating a search pattern corresponding to the text input;
determining a pattern associated with each of the plurality of textual entities;
associating the search pattern corresponding to the text input with patterns associated with the plurality of text entities;
identifying one or more matching patterns from the patterns associated with the plurality of textual entities based on the correspondence; and
causing extraction of related text entities corresponding to the one or more matching patterns from the plurality of text entities;
system. 13. The system of claim 12, wherein the input files comprise at least one of image files and Portable Document Format (PDF) files. The processor-executable instructions, when executed by the processor, cause the processor to:
A domain-based extraction, wherein performing the domain-based extraction comprises:
determining a domain-based name associated with each of the plurality of textual entities, wherein the domain-based name associates each of the plurality of textual entities with domain-based names stored in a data repository; domain-based extraction, including determining, determined based on matching with the list;
location-based extraction, wherein performing the location-based extraction comprises:
identifying a type associated with the input file;
determining the location of one or more associated textual entities in the input file based on the type associated with the input file; and
extracting the one or more related textual entities from the input file based on the location, the location corresponding to a type associated with the input file stored in the data repository; extracting, predetermined based on a template;
location-based extraction, including
Part-of-Speech (POS)-based extraction, wherein performing the Part-of-Speech (POS)-based extraction comprises:
identifying a POS associated with the plurality of text entities, wherein the POS is one of nouns, pronouns, verbs, adverbs, adjectives, conjunctions, prepositions, and interjections; ,
extracting one or more textual entities that POS identifies as nouns;
identifying one or more unique textual entities for which no POS is identified;
determining an entity type associated with each of the one or more unique textual entities, wherein the entity type is one of a value entity and a pattern entity, each value entity associated with determining that each pattern entity has an associated pattern;
For each value entity,
automatically identifying one or more text entities that have an associated value that is the same as the associated value of the value entity;
For each pattern entity,
automatically generating search patterns corresponding to patterns associated with the pattern entity;
associating the search pattern corresponding to the pattern entity with patterns associated with the plurality of text entities;
determining one or more matching patterns from the patterns associated with the plurality of text entities based on the correspondence; and
part-of-speech (POS)-based extraction, comprising: extracting from the plurality of text entities matching text entities corresponding to the one or more matching patterns; and
Machine learning (ML)-based extraction, wherein performing the ML-based extraction comprises:
further causing at least one of machine learning (ML)-based extraction, including using a classification-based ML model to generate recommendations based on historical data;
13. The system of claim 12. identifying text data from an input file, wherein the identified text data includes a plurality of text entities;
receiving text input from a user to identify related text entities from the plurality of text entities;
automatically generating a search pattern corresponding to the text input;
determining a pattern associated with each of the plurality of textual entities;
associating the search pattern corresponding to the text input with patterns associated with the plurality of text entities;
identifying one or more matching patterns from the patterns associated with the plurality of textual entities based on the correspondence;
extracting from the plurality of text entities associated text entities corresponding to the one or more matching patterns; non-transitory computer-readable storage medium. The set of computer-executable instructions instructs the computer, including the one or more processors, to:
A domain-based extraction, wherein performing the domain-based extraction comprises:
determining a domain-based name associated with each of the plurality of textual entities, wherein the domain-based name associates each of the plurality of textual entities with domain-based names stored in a data repository; domain-based extraction, including determining, determined based on matching with the list;
location-based extraction, wherein performing the location-based extraction comprises:
identifying a type associated with the input file;
determining the location of one or more associated textual entities in the input file based on the type associated with the input file; and
extracting the one or more related textual entities from the input file based on the location, the location corresponding to a type associated with the input file stored in the data repository; extracting, predetermined based on a template;
location-based extraction, including
Part-of-Speech (POS)-based extraction, wherein performing the Part-of-Speech (POS)-based extraction comprises:
identifying a POS associated with the plurality of text entities, wherein the POS is one of nouns, pronouns, verbs, adverbs, adjectives, conjunctions, prepositions, and interjections; ,
extracting one or more textual entities that POS identifies as nouns;
identifying one or more unique textual entities for which no POS is identified;
determining an entity type associated with each of the one or more unique textual entities, wherein the entity type is one of a value entity and a pattern entity, each value entity associated with determining that each pattern entity has an associated pattern;
For each value entity,
automatically identifying one or more text entities that have an associated value that is the same as the associated value of the value entity;
For each pattern entity,
automatically generating search patterns corresponding to patterns associated with the pattern entity;
associating the search pattern corresponding to the pattern entity with patterns associated with the plurality of text entities;
determining one or more matching patterns from the patterns associated with the plurality of text entities based on the correspondence; and
part-of-speech (POS)-based extraction, comprising: extracting from the plurality of text entities matching text entities corresponding to the one or more matching patterns; and
Machine learning (ML)-based extraction, wherein performing the ML-based extraction comprises:
using a classification-based ML model to generate recommendations based on historical data;
16. The non-transitory computer-readable storage medium of claim 15.

Images