Zopf, Markus (2019)
Towards Context-free Information Importance Estimation.
Technische Universität Darmstadt
Dissertation, Erstveröffentlichung
Kurzbeschreibung (Abstract)
The amount of information contained in heterogeneous text documents such as news articles, blogs, social media posts, scientific articles, discussion forums, and microblogging platforms is already huge and is going to increase further. It is not possible for humans to cope with this flood of information, so that important information can neither be found nor be utilized. This situation is unfortunate since information is the key driver in many areas of society in the present Information Age. Hence, developing automatic means that can assist people to handle the information overload is crucial. Developing methods for automatic estimation of information importance is an essential step towards this goal.
The guiding hypothesis of this work is that prior methods for automatic information importance estimation are inherently limited because they are based on merely correlated signals that are, however, not causally linked with information importance. To resolve this issue, we lay in this work the foundations for a fundamentally new approach for importance estimation. The key idea of context-free information importance estimation is to equip machine learning models with world knowledge so that they can estimate information importance based on causal reasons.
In the first part of this work, we lay the theoretical foundations for context-free information importance estimation. First, we discuss how the abstract concept of information importance can be formally defined. So far, a formal definition of this concept is missing in the research community. We close this gap by discussing two information importance definitions, which equate the importance of information with its impact on the behavior and the impact on the course of life of the information recipients, respectively. Second, we discuss how information importance estimation abilities can be assessed. Usually, this is done by performing automatic summarization of text documents. However, we find that this approach is not ideal. Instead, we propose to consider ranking, regression, and preference prediction tasks as alternatives in future work. Third, we deduce context-free information importance estimation as a logical consequence of the previously introduced importance definitions. We find that reliable importance estimation, in particular for heterogeneous text documents, is only possible with context-free methods.
In the second part, we develop the first machine learning models based on the idea of context-free information importance estimation. To this end, we first tackle the lack of suited datasets that are required to train and test machine learning models. In particular, large and heterogeneous datasets to investigate automatic summarization of multiple source documents are missing, because their construction is complicated and costly. To solve this problem, we present a simple and cost-efficient corpus construction approach and demonstrate its applicability by creating new multi-document summarization datasets. Second, we develop a new machine learning approach for context-free information importance estimation, implement a concrete realization, and demonstrate its advantages over contextual importance estimators. Third, we develop a new method to evaluate automatic summarization methods. Previous works are based on expensive reference summaries and unreliable semantic comparisons of text documents. On the contrary, our approach uses cheap pairwise preference annotations and only much simpler sentence-level similarity estimation.
This work lays the foundations for context-free information importance estimation. We hope that future research will explore if this fundamentally new type of information importance estimation can eventually lead to human-level information importance estimation abilities.
Typ des Eintrags: | Dissertation | ||||
---|---|---|---|---|---|
Erschienen: | 2019 | ||||
Autor(en): | Zopf, Markus | ||||
Art des Eintrags: | Erstveröffentlichung | ||||
Titel: | Towards Context-free Information Importance Estimation | ||||
Sprache: | Englisch | ||||
Referenten: | Fürnkranz, Prof. Johannes ; Dagan, Prof. Ido | ||||
Publikationsjahr: | 7 August 2019 | ||||
Ort: | Darmstadt | ||||
Datum der mündlichen Prüfung: | 29 Januar 2019 | ||||
URL / URN: | https://tuprints.ulb.tu-darmstadt.de/8976 | ||||
Kurzbeschreibung (Abstract): | The amount of information contained in heterogeneous text documents such as news articles, blogs, social media posts, scientific articles, discussion forums, and microblogging platforms is already huge and is going to increase further. It is not possible for humans to cope with this flood of information, so that important information can neither be found nor be utilized. This situation is unfortunate since information is the key driver in many areas of society in the present Information Age. Hence, developing automatic means that can assist people to handle the information overload is crucial. Developing methods for automatic estimation of information importance is an essential step towards this goal. The guiding hypothesis of this work is that prior methods for automatic information importance estimation are inherently limited because they are based on merely correlated signals that are, however, not causally linked with information importance. To resolve this issue, we lay in this work the foundations for a fundamentally new approach for importance estimation. The key idea of context-free information importance estimation is to equip machine learning models with world knowledge so that they can estimate information importance based on causal reasons. In the first part of this work, we lay the theoretical foundations for context-free information importance estimation. First, we discuss how the abstract concept of information importance can be formally defined. So far, a formal definition of this concept is missing in the research community. We close this gap by discussing two information importance definitions, which equate the importance of information with its impact on the behavior and the impact on the course of life of the information recipients, respectively. Second, we discuss how information importance estimation abilities can be assessed. Usually, this is done by performing automatic summarization of text documents. However, we find that this approach is not ideal. Instead, we propose to consider ranking, regression, and preference prediction tasks as alternatives in future work. Third, we deduce context-free information importance estimation as a logical consequence of the previously introduced importance definitions. We find that reliable importance estimation, in particular for heterogeneous text documents, is only possible with context-free methods. In the second part, we develop the first machine learning models based on the idea of context-free information importance estimation. To this end, we first tackle the lack of suited datasets that are required to train and test machine learning models. In particular, large and heterogeneous datasets to investigate automatic summarization of multiple source documents are missing, because their construction is complicated and costly. To solve this problem, we present a simple and cost-efficient corpus construction approach and demonstrate its applicability by creating new multi-document summarization datasets. Second, we develop a new machine learning approach for context-free information importance estimation, implement a concrete realization, and demonstrate its advantages over contextual importance estimators. Third, we develop a new method to evaluate automatic summarization methods. Previous works are based on expensive reference summaries and unreliable semantic comparisons of text documents. On the contrary, our approach uses cheap pairwise preference annotations and only much simpler sentence-level similarity estimation. This work lays the foundations for context-free information importance estimation. We hope that future research will explore if this fundamentally new type of information importance estimation can eventually lead to human-level information importance estimation abilities. |
||||
Alternatives oder übersetztes Abstract: |
|
||||
URN: | urn:nbn:de:tuda-tuprints-89762 | ||||
Sachgruppe der Dewey Dezimalklassifikatin (DDC): | 000 Allgemeines, Informatik, Informationswissenschaft > 004 Informatik | ||||
Fachbereich(e)/-gebiet(e): | 20 Fachbereich Informatik 20 Fachbereich Informatik > Knowledge Engineering |
||||
Hinterlegungsdatum: | 20 Okt 2019 19:55 | ||||
Letzte Änderung: | 20 Okt 2019 19:55 | ||||
PPN: | |||||
Referenten: | Fürnkranz, Prof. Johannes ; Dagan, Prof. Ido | ||||
Datum der mündlichen Prüfung / Verteidigung / mdl. Prüfung: | 29 Januar 2019 | ||||
Export: | |||||
Suche nach Titel in: | TUfind oder in Google |
Frage zum Eintrag |
Optionen (nur für Redakteure)
Redaktionelle Details anzeigen |