• Keine Ergebnisse gefunden

Chapter 3. Structured Summarization with Concept Maps

already seems to be warranted. Yet, we want to emphasize that CM-MDS differs from concept map mining in several aspects, making it an even more relevant task.

First, by formulating the task as a variant of MDS, we add a particular focus on the summarization aspect. In exploratory search scenarios, document collections are typically huge, requiring effective techniques to reduce the content to a small overview. Existing work for concept map mining mostly focused on the extraction subtasks, paying only little attention to summarization. For instance, we are not aware of any previous work that explicitly included a size restriction in their task formulation. Second, the concept map construction subtask has been largely ignored, with only Zubrinic et al. (2015) proposing a first heuristic algorithm. No work has explicitly formulated an optimization problem with an objective function and constraints as described in the previous section.

In addition to this new focus, we want to emphasize that also the other subtasks, namely extraction, grouping and labeling, are far from being solved. With the already explored techniques reviewed in Section 2.3.1, most of the challenges discussed in the last section cannot be handled well enough for the approaches to be broadly applicable in practice.

3.4.2 Text Summarization

Given Definition 5, several obvious parallels to the task of MDS (see Section 2.3.2) are vis-ible: Both tasks work with multi-document inputs, both have a restricted size budget and both require producing the best possible summary within that budget. The main difference is that the summaries have different forms, one being a text, while the other is a concept map. For the latter, applying known MDS techniques alone is thus not sufficient, as the content also has to be structured into concepts and relations. While we are not aware of any work from the summarization community that produces summary concept maps, other types of structured summaries have been explored. Most notably, Haghighi and Vander-wende (2009), Christensen et al. (2014) and Tauchmann et al. (2018) all propose forms of hierarchically organized textual summaries covering different parts of a collection, giving a user, similar as with our concept map–based approach, more navigation capabilities.

Both MDS and CM-MDS require some form of importance estimation. As the existing work on concept map mining in this regard is limited (see Section 2.3.1.5), the application of techniques from MDS is an interesting direction to explore. In contrast to most extrac-tive MDS work, the importance of concepts and relations, rather than sentences, has to be estimated. MDS approaches working with bigrams (Gillick and Favre, 2009, Boudin et al., 2015) or entities (Li, 2015, Li et al., 2016a) are closest to that requirement.

With regard to selection, two main differences are important. First, the notion of redun-dancy, a central topic in MDS, is less relevant for CM-MDS. As we work with smaller units, concepts and relations, and since redundant mentions of them are already grouped together in previous steps, no redundancy should exist among the elements considered during

sub-3.4. Relations to Existing Tasks

graph selection. Explicitly modeling redundancy reduction as an objective is therefore not necessary. The MDS work of Gillick and Favre (2009) goes in a similar direction by making selections based on bigrams instead of sentences. However, the set of bigrams can still con-tain different bigrams with the same meaning, such that redundancy remains a challenge.

Second, the selection problems are different optimization problems. The MDS selec-tion problem is difficult due to the mismatch of cost (length budget used by a sentence) and utility (contributed important content) per element, resulting in the NP-hard knapsack problem. For CM-MDS, the cost is always one, as only a cardinality restriction is defined, such that one can optimize by simply selecting by utility. However, the additional connect-edness requirement introduces a new restriction, making an optimal selection still difficult.

Due to the different constraints, sentence selection techniques discussed in Section 2.3.1.5 cannot be directly applied. More similar is the work on abstractive summarization by Li et al. (2016a) and Liu et al. (2015), who summarize graphs obtained by semantic parsing methods, resulting in a similar subgraph selection problem.

As challenges of the importance estimation, we mentioned the dependence on world knowledge and user-specific information. While these ideas apply to MDS as well, most of the existing work ignores them and deals only with generic summarization, i.e. using only information from the source documents. Notable exceptions are Louis (2014) and Peyrard (2018), who aim to integrate background knowledge, and P.V.S. and Meyer (2017), who propose an interactive summarization system that tailors a summary to a specific user.

3.4.3 Information Extraction

While summarization, as discussed above, has similarities with the importance estimation and concept map construction subtasks of CM-MDS, the remaining subtasks, in particu-lar mention extraction and grouping, are essentially information extraction tasks. As we already pointed out in Section 2.3.3, due to the open vocabulary property of concept map labels,open information extraction is closest to the mention extraction subtasks.

In both tasks, we identify relations between two arguments in text and use labels taken from that text to describe that proposition. Conciseness and assertedness are desired in both cases. One difference is that the notion of an OIE argument is typically broader than that of a concept, such that not all propositions extracted by OIE are relations between two concepts.19 And second, several OIE systems, in particular more recent ones, extract n-ary tuples. A concept map, however, can by definition only represent binn-ary relationships between pairs of concepts and can therefore not use all OIE extractions directly.

To the best of our knowledge, despite the high similarity between OIE and concept and relation mention extraction, no work has yet explored the application of OIE to the latter. Similarly, a large body of work on coreference resolution exists, surveyed for instance

19For more details and examples, see step 2 of the corpus creation described in Section 4.3.2.

Chapter 3. Structured Summarization with Concept Maps

by Zheng et al. (2011), but only rudimentary techniques have been used for concept and relation grouping. In Chapter 5 and Chapter 6, we therefore compare existing information extraction methods against techniques that were specifically designed for concept maps.

3.4.4 Knowledge Graphs

Concept maps have some similarity with more formal knowledge representations, of which many have been developed and used in the past. One example are knowledge bases such as Freebase (Bollacker et al., 2009) or WikiData (Tanon et al., 2016), which are databases of facts about real-world entities like politicians, celebrities, countries, organizations or places. Much work has been done in the NLP community to populate these databases au-tomatically from text, for example as part of the automatic knowledge base construction workshops (Pujara et al., 2016). Projects like Cyc (Lenat et al., 1986), which are mainly driven by the knowledge representation and reasoning community, have a different fo-cus as they aim to formally capture common-sense knowledge that is rarely represented in text. Within the semantic web community, the construction of domain ontologies has been studied extensively (Breitman et al., 2007). Rather than facts about real-world entities, ontologies model general concepts, their properties and relationships. Methods to automat-ically extract ontologies from text are studied as ontology learning (Maedche, 2002). More recently, the term knowledge graph became popular, encompassing both knowledge bases and ontologies, but it still lacks a clear definition (Ehrlinger and Wöß, 2016).

As pointed out in Section 2.1.3, a main difference to concept maps is that knowledge graphs are more formal. Whereas concept maps follow the open label paradigm and are meant to be interpretable by humans, knowledge graphs are more strictly typed — accord-ing to a predefined schema — as their purpose is to be machine-readable. Ontology learnaccord-ing and knowledge base construction methods are therefore not directly applicable. However, the work by Zouaq et al. (2011) showed that the similarities between the tasks can be lever-aged by using concept map mining as a preprocessing step for ontology learning.

Another difference is that the ultimate goal of ontology learning and knowledge base construction is to create high-quality knowledge graphs. Therefore, the focus is on making high-confidence extractions or on finding concepts and relations still missing in the graph.

Whether all the information present in a specific text is captured by extraction methods is less relevant, as the techniques could also just be applied to additional documents to grow the knowledge graph. In contrast, for CM-MDS, fully covering a given text is much more important, as only then an accurate summary can be produced.

The recently proposed open knowledge graphs (Wities et al., 2017) are a knowledge rep-resentation that is less formal than the reprep-resentations discussed above and follows, similar to concept maps and OIE, an open label paradigm for its elements. Open knowledge graphs depict entities and propositions describing their relationships, both being the result of an