TU Darmstadt / ULB / TUbiblio

Principled Approach to Natural Language Generation

Puzikov, Yevgeniy (2021):
Principled Approach to Natural Language Generation. (Publisher's Version)
Darmstadt, Technische Universität Darmstadt,
DOI: 10.26083/tuprints-00019115,
[Ph.D. Thesis]

Abstract

The research field of Natural Language Generation offers practitioners a wide range of techniques for producing texts from a variety of data types. These techniques find their way into various real- world applications and help many people to automate time-consuming tasks of text production in many areas. At the moment, the design and evaluation of text generation approaches is largely empirical. Many systems are being developed to solve one particular task and work on a single data type, which makes it hard to compare the approach to any other technique and critically evaluate its performance. Some systems employ complex machine learning algorithms to learn rich data representations and perform joint modeling of the steps involved in the process of text generation. Such approaches offer an attractive trade-off between the development costs and output quality, but often lack transparency in terms of the reasoning about the behavior of the system. The number of the proposed approaches constantly grows, but the methodology lags behind and sometimes fails to solicit a better understanding of which approaches work, and the reasons for it. In this thesis we present our view on the task of text production from a methodological point of view. We analyze the existent scientific literature, examine common text generation approaches and the established evaluation protocols. We further propose a principled view on the problem: we break it into components, examine their interaction and develop a set of recommendations which are envisioned to offer assistance during the design or analysis of a study. We further conduct a range of experiments to test this framework in several text generation tasks. First, we show that task specification analysis sometimes allows one to solve the problem at hand with very simple techniques, without resorting to the complex machinery of advanced statistical learning methods. We further demonstrate the potential of the developed framework to find discrepancies in the established evaluation protocols. We show that sometimes neither metric, nor conventional human evaluation is sufficient to draw conclusions about system performance. We demonstrate how a system can fit the data to achieve high automatic metric scores, while falling short in terms of actual output quality. Finally, we use the framework to demonstrate how one can develop effective text generation systems without sacrificing the transparency of the inner working logic, making the developed systems both accurate and reliable.

Item Type: Ph.D. Thesis
Erschienen: 2021
Creators: Puzikov, Yevgeniy
Status: Publisher's Version
Title: Principled Approach to Natural Language Generation
Language: English
Abstract:

The research field of Natural Language Generation offers practitioners a wide range of techniques for producing texts from a variety of data types. These techniques find their way into various real- world applications and help many people to automate time-consuming tasks of text production in many areas. At the moment, the design and evaluation of text generation approaches is largely empirical. Many systems are being developed to solve one particular task and work on a single data type, which makes it hard to compare the approach to any other technique and critically evaluate its performance. Some systems employ complex machine learning algorithms to learn rich data representations and perform joint modeling of the steps involved in the process of text generation. Such approaches offer an attractive trade-off between the development costs and output quality, but often lack transparency in terms of the reasoning about the behavior of the system. The number of the proposed approaches constantly grows, but the methodology lags behind and sometimes fails to solicit a better understanding of which approaches work, and the reasons for it. In this thesis we present our view on the task of text production from a methodological point of view. We analyze the existent scientific literature, examine common text generation approaches and the established evaluation protocols. We further propose a principled view on the problem: we break it into components, examine their interaction and develop a set of recommendations which are envisioned to offer assistance during the design or analysis of a study. We further conduct a range of experiments to test this framework in several text generation tasks. First, we show that task specification analysis sometimes allows one to solve the problem at hand with very simple techniques, without resorting to the complex machinery of advanced statistical learning methods. We further demonstrate the potential of the developed framework to find discrepancies in the established evaluation protocols. We show that sometimes neither metric, nor conventional human evaluation is sufficient to draw conclusions about system performance. We demonstrate how a system can fit the data to achieve high automatic metric scores, while falling short in terms of actual output quality. Finally, we use the framework to demonstrate how one can develop effective text generation systems without sacrificing the transparency of the inner working logic, making the developed systems both accurate and reliable.

Place of Publication: Darmstadt
Collation: viii, 172 Seiten
Divisions: 20 Department of Computer Science
20 Department of Computer Science > Ubiquitous Knowledge Processing
TU-Projects: Bund/BMBF|01IS17050|Software Campus 2.0
DFG|GRK1994|Gurevych_GRK_1994_Au
DFG|GU798/17-1|Deutsch-Israelische
Date Deposited: 19 Jul 2021 11:33
DOI: 10.26083/tuprints-00019115
Official URL: https://tuprints.ulb.tu-darmstadt.de/19115
URN: urn:nbn:de:tuda-tuprints-191154
Referees: Gurevych, Prof. Dr. Iryna ; Dagan, Prof. Ido ; Gardent, Prof. Dr. Claire
Refereed / Verteidigung / mdl. Prüfung: 22 June 2021
Alternative Abstract:
Alternative abstract Language

Das Forschungsgebiet der natürlichen Text-/Sprachgenerierung stellt Anwendern eine breite Palette von Techniken zur Generieren von Texten aus Vielheit von Datenarten bereit. Diese Techniken werden in konkreten Anwendungen benutzt und helfen vielen Menschen in vielen Bereichen bei der Automatisierung zeitaufwändiger Textproduktion. Im Moment ist das Enwickeln und die Auswertung von Algorithmen zur Textegenerierung weitgehend empirisch. Viele Systeme werden entwickelt, um eine bestimmte Aufgabe zu lösen und mit einer einzigen Datenart zu arbeiten, was es schwierig macht, den Ansatz mit anderen Techniken zu vergleichen und kritisch seine Leistung zu bewerten. Einige Systeme verwenden komplexe Algorithmen des maschinellen Lernens, um umfangreiche Datendarstellungen zu lernen und gemeinsame Modellierung der Schritte, die am Prozess der Textgenerierung beteiligt sind, auszuführen. Solche Ansätze bieten einen attraktiven Kompromiss von Entwicklungskosten und Ausgabequalität, erlauben aber häufig nicht, das Verhalten des Systems zu verstehen und zu analysieren. Die Zahl der vorgeschlagenen Ansätze wächst ständig, aber die Methodik hinken hinterher und manchmal tragen diese nicht zu einem besseren Verständnis bei, welche Ansätze funktionieren und warum. In dieser Arbeit präsentieren wir unseren Standpunkt zur Aufgabe der Textproduktion von einem methodischen Blickpunkt. Wir analysieren die existierende wissenschaftliche Literatur, untersuchen gemeinsame Ansätze zur Texterzeugung und etablierte Bewertungsprotokolle. Wir schlagen ferner eine grundlegende Sichtweise auf das Problem vor: Wir zerlegen es in Komponenten, untersuchen deren Wechselwirkung und entwickeln eine Reihe von Empfehlungen, die zur Unterstützung bei der Gestaltung oder Analyse einer Studie genutzt werden können. Wir führen eine Reihe von Experimenten durch, um dieses Framework in mehreren Textgenerierungsaufgaben zu testen. Zunächst zeigen wir, dass die Aufgabenspezifikationen so ist, dass es manchmal das vorliegende Problem mit sehr einfachen Techniken zu lösen erlaubt, ohne Rückgriffe auf die komplexe Maschinerie des fortgeschrittenen statistischen Lernens zu benötigen. Wir zeigen weiter das Potenzial des entwickelten Frameworks, um Abweichungen in den existierenden Bewertungsprotokollen zu finden. Wir zeigen, dass manchmal weder metrische, noch konventionelle menschliche Bewertung ausreichend ist, um Rückschlüsse auf die Systemleistung zu ziehen. Wir demonstrieren, wie ein System die Daten anpassen kann, um hohe automatische metrische Werte zu erreichen, während sie in Bezug auf die tatsächliche Ausgangsqualität stagnieren oder schlechter werden. Schließlich verwenden wir das Framework, um zu demonstrieren, wie man effektive Texter- zeugungssysteme ohne Einbußen bei der Erklärbarkeit der inneren Arbeitslogik entwickeln kann, wodurch die entwickelten Systeme genau und zuverlässig zugleich werden.

German
Export:
Suche nach Titel in: TUfind oder in Google
Send an inquiry Send an inquiry

Options (only for editors)
Show editorial Details Show editorial Details