This section presents possible lines of future research.
8.2.1 Future Work in Ontology Construction
With respect to M-OntoBUILD, this technique and the tool itself requires more extensive research. One of the concerns discussed in Chapter 4 is improving the coverage of the domain ontologies, for which we can explore the use of Wikipedia’s Category Graph as an alternative for adding hierarchies of concepts. We could also consider a systematic approach for defining concept associations that is not limited to WordNet and Wikipedia’s wikilinks. For instance, we can analyse other relations defined in DBPedia [Auer et al., 2008], YAGO [Suchanek et al., 2008], ConceptNet [Havasi et al., 2007], BabelNet [Navigli and Velardi, 2010], WiSeNet [Moro and Navigli, 2012] and Probase [Wu et al., 2012]. Alternatively, we can explore other ontological properties to make the M-Ontos more robust for other purposes (e.g. conceptual properties and relation labelling), which are out of the scope of our system’s requirements.
8.2.2 Future Work in the Effects of Domain over Semantic Relatedness
The domain-sensitive semantic relatedness metric learned in Chapter 6 showed a higher cor- relation to human assessments compared to existing automatic metrics. However, “domain” can only be considered one of the many factors considered by humans while assessing seman- tic relatedness between concepts. The metric itself is affected by the set of domains that are considered for the application, for which in our case, more M-Ontos are required. These changes may also affect the coefficients of the learned metric which was defined in Section 6.7. We could also consider comparing this metric to other existing metrics over other datasets, to see if the high correlation effect prevails. However, as above, we would require domain ontologies covering the terms included in these datasets.
8.2.3 Future Work in Mixed-initiative Conversational Agents
As mentioned above, the use of semantic relatedness in open dialogue to maintain coherence is only a first step taken in this thesis. Using the Toy as a framework and improving its capabilities is not under our control, as these decisions depend on our industry partner. To
continue our research on this line, we could evaluate our Semantic Relatedness Selection Mechanism in a more sophisticated conversational agent framework. This requires more conversational domains available for the agent, in such a way that we can also measure the effect of switching to a topic in a different domain for judges in dialogue. We could also consider applying semantic relatedness as a mechanism for opportunistic topic-switching, so the agent can produce behaviours that users consider interesting, while remaining coherent. We expect that having multiple domains plugged into the Toy implies more potential for variability in constructed conversations. In particular, switching topics to make con- versations more “surprising” and thereby more engaging is a possibility, although retaining coherence becomes an interesting challenge. While the analysis of the learned metric of semantic relatedness takes into account domain switching, longer conversations and human evaluation subjects is warranted. Furthermore, we can expect some ambiguity because of the multiple senses that a word can bear and the process conducted to match QA-fragments to M-Onto concepts. Finally, having more than one domain can pose a challenge in terms of adding computational complexity to the system, which is also a further line of research.
Top-level domain concepts used in
the Domain Appropriateness user
study
A.1 Concept Terms Selected for Domain Amusement park
Concept Definition Very
related
Related Unrelated
Arson malicious burning to destroy property 0 1 6
Budget a summary of intended expenditures along with
proposals for how to meet them
1 4 2
Building a structure that has a roof and walls and stands more or less permanently in one place
1 4 2
Chairlift a ski lift on which riders (skiers or sightseers) are seated and carried up or down a mountainside; seats are hung from an endless overhead cable
3 4 2
Crime (criminal law) an act punishable by law; usually considered an evil act
0 2 5
Desegregation the action of incorporating a racial or religious group into a community
1 3 2
Enterprise a purposeful or industrious undertaking (especially one that requires effort or boldness)
4 6 3
Fair a sale of miscellany; often for charity 4 3 0
Fast food inexpensive food (hamburgers or chicken or milkshakes) prepared and served quickly
5 8 3
Funfair a commercially operated park with stalls and shows for amusement
8 1 0
Go-kart a small low motor vehicle with four wheels and an open framework; used for racing
3 5 1
Hamburger a fried cake of minced beef served on a bun 0 6 3
Hot-dog a smooth-textured sausage of minced beef or pork usually smoked; often served on a bread roll
5 1 0
Industrial revolution the transformation from an agricultural to an industrial nation
0 2 7
Jaws holding device consisting of one or both of the opposing parts of a tool that close to hold an object
1 1 4
Parking lot a lot where cars are parked 0 5 4
Penny arcade an arcade with coin-operated devices for entertainment 4 4 1
Person a human being 3 4 0
Playground yard consisting of an outdoor area for children’s play 6 2 1
Recreation an activity that diverts or amuses or stimulates 6 1 0
Concept Definition Very related
Related Unrelated
Roller coaster elevated railway in an amusement park (usually with sharp curves and steep inclines)
8 0 1
Suburb a residential district located on the outskirts of a city 2 4 1
Sweet a food rich in sugar 3 2 1
Tournament a sporting competition in which contestants play a series of games to decide the winner
0 4 3
Urban planning the branch of architecture dealing with the design and organization of urban space and activities
2 4 1
World all of the inhabitants of the earth 1 2 4