diplomsko delo
Abstract
Nestrukturirani dokumenti zajemajo informacije v oblikah in postavitvah, ki se lahko od enega primerka do drugega razlikujejo, kar lahko oteži in podraži nalogo pridobivanja informacij. Kot rešitev se je v zadnjih letih za razumevanje dokumentov na področju dokumentne inteligence pričela uporaba nevronskih jezikovnih modelov, usposobljenih na učnih množicah dokumentov. V diplomskem delu za pridobivanje informacij iz skeniranih trgovinskih računov uporabljamo prehodno učeni nevronski jezikovni model, zgrajen iz transformatorjev. Model je natančno učen z uporabo učne množice SROIE za izluščitev štirih kategorij, tj. imen in naslovov trgovin, datumov in skupnih cen. Za pridobivanje informacij smo uporabili prepoznavo imenskih entitet. Za primerjavo izvajamo poskuse s spreminjanem hiperparametrov modela. S spremembo nevronskega jezikovnega modela smo pri poskusih dosegli največjo
natančnost klasifikacije: 96,7 %.
Keywords
dokumentna inteligenca;obdelava naravnih jezikov;prepoznava imenskih entitet;jezikovni modeli;transformatorji;diplomske naloge;
Data
Language: |
Slovenian |
Year of publishing: |
2021 |
Typology: |
2.11 - Undergraduate Thesis |
Organization: |
UM FERI - Faculty of Electrical Engineering and Computer Science |
Publisher: |
[U. Knupleš] |
UDC: |
004.652.8(043.2) |
COBISS: |
95975171
|
Views: |
274 |
Downloads: |
20 |
Average score: |
0 (0 votes) |
Metadata: |
|
Other data
Secondary language: |
English |
Secondary title: |
Named entity recognition on unstructured documents using neural language models |
Secondary abstract: |
Layouts and formats of information, in unstructured documents, can differ from one another and can make the extraction of information difficult and costly. Therefore, in recent years, the field of document intelligence began with the usage of neural language models trained on datasets of documents for document understanding. In the thesis, we adopt a pre-trained neural language model based on transformers, for information extraction out of scanned store invoices. The model is fine-tuned, using the SROIE dataset, based on four categories to extract store names and addresses, dates and total prices. For information extraction we used named entity recognition to classify tokens into the four prementioned categories. We conducted experiments using altered hyperparameters of the model for comparison. With the usage of the fine-tuned, altered neural language model, we achieved a maximum classification accuracy score of 96.7 %. |
Secondary keywords: |
Document intelligence;natural language processing;named entity recognition;langauge models;transformers; |
Type (COBISS): |
Bachelor thesis/paper |
Thesis comment: |
Univ. v Mariboru, Fak. za elektrotehniko, računalništvo in informatiko, Računalništvo in informacijske tehnologije |
Pages: |
IX, 35 str. |
ID: |
13344285 |