8.1 Main Findings

The general question governing this thesis has been: “How can we leverage conceptual knowledge in the language modeling framework to improve information access?” We have approached this question as a query modeling problem. That is, we have looked at methods and algorithms to improve textual queries or their representations using concept languages in the context of generative language models. This main question has lead us to formulate five main research questions listed in Section 1.4 which have been answered in the previous chapters. In this section we recall the answers.

We have started the thesis with an overview of current approaches to information retrieval, concept languages, and their combination (Chapters 2 and 3). We have zoomed in on a technique called query modeling, with which the information need of a user can be captured more thoroughly than solely using the initial query.

In Chapter 4, we have employed pseudo and explicit user feedback in the form of relevance ratings at the document level to improve the estimation of the query model. The first research question thus dealt with relevance feedback methods for query modeling. We asked:

RQ 1.
What are effective ways of using relevance feedback information for query modeling to improve retrieval performance?

We have presented two query modeling methods for relevance feedback that are based on leveraging the similarity between feedback documents and the set thereof. By providing a comprehensive analysis, evaluation, comparison, and discussion (in both theoretical and practical terms) of our novel and various other core models for query modeling using relevance feedback, we have shown that all the models we have evaluated are able to improve upon a baseline without relevance feedback in the case of explicit relevance feedback. One of our proposed models (NLLR) is particularly suited when explicit relevance assessments are available. In the case of pseudo relevance feedback, we have observed that RM-1 is the most robust model. Parsimonious relevance models, on the other hand, perform very well on large, noisy collections. We have further found that, in the case of pseudo relevance feedback, there exists a large variance in the resulting retrieval performance for different amounts of pseudo relevant documents, most notably on large, noisy collections. We have also concluded that the test collection itself is of influence on the relative performance of the models; there is no single model that outperforms the others on all test collections. As to the observations made when using explicit relevance feedback, here we found that the variance with respect to the number of feedback documents is much less pronounced. Furthermore, one of the two novel methods we introduce consequently outperforms the other models.

Inspired by relevance feedback methods, we then developed a two-step method in Chapter 5 that uses concepts to estimate a conceptual query model. Here, we moved beyond the lexical level by introducing an automatic method for generating a conceptual representation of a query and subsequently using this representation to improve end-to-end retrieval. We asked

RQ 2.
What are effective ways of using conceptual information for query modeling to improve retrieval performance?

We have introduced a novel way of using document-level annotations in the form of concepts to improve end-to-end retrieval performance. We have found that our proposed method obtained the highest performance of all evaluated models. We have concluded that, although each step in our method of applying conceptual language models is not significantly different from the other, the full model is able to significantly outperform both a standard language modeling and a pseudo relevance feedback approach.

After that, in Chapter 6, we have considered DBpedia as a concept language, in which case each Wikipedia article constitutes a concept. Here, we have turned to a different way of obtaining concepts pertinent to a user’s query based on supervised machine learning. The research question was:

RQ 3.
Can we successfully address the task of mapping search engine queries to concepts using a combination of information retrieval and machine learning techniques?

We have developed a novel way of associating concepts with queries that can effectively handle arbitrary features and answered this question in the affirmative. We have concluded that our proposed approach significantly outperforms other methods, including commonly used methods based on a lexical matching between query terms and concept labels.

In the next chapter (Chapter 7), we have moved to the open domain and brought together the ideas presented in all preceding chapters. We have taken the conceptual mapping approach from Chapter 6 to obtain DBpedia concepts. Next, we have used the natural language text associated with each concept (in the form of the accompanying Wikipedia article) to estimate a query model, similar to the conceptual language models presented in Chapter 5. The associated research question was:

RQ 4.
What are the effects on retrieval performance of applying pseudo relevance feedback methods to texts associated with concepts that are automatically mapped from ad hoc queries?

We have found that the conceptual mapping method presented in Chapter 6 transfers well to the open domain; the linked concepts seem reasonable and the estimated query models are to the point. When evaluated, we have concluded that our novel method is able to improve recall, precision, and diversity metrics on two large web collections.