| Up | Next | Prev | PrevTail | Tail |
In previous chapters we have seen various ways of updating the estimate of a query model, for example through the use of feedback information (Chapter 4) or conceptual document annotations (Chapter 5). In essence, these approaches are a form of data fusion, where information from multiple sources is combined to influence a document ranking. Such fusion methods exist in a number of related tasks. For example, in web retrieval it is common to take into account anchor texts or some function of the web graph [45]. In multimedia environments, different modalities (text, video, speech, etc.) need to be combined. In cross-lingual IR, where the queries and documents are stated in different languages, evidence from multiple languages is merged to obtain a final ranking. In our query modeling case, we have combined evidence from either top-ranked or relevant documents and the initial query. In Chapter 5 we have added to this concepts in the form of document annotations. In Chapter 6 we have linked domain-restricted queries to DBpedia and the question arose “Can we apply the semantic analysis based on Linked Open Data (LOD) to the open domain?” Furthermore, can we apply these linked concepts for retrieval, using the ideas presented in Chapter 4 and Chapter 5?
Looking from a different angle, there have been several developments in web search over the last 20 years [19, 300]. Initially, web pages were ranked solely based on term frequency (TF) and inverse document frequency (IDF) of the terms a user entered in her query. Later, this was enriched with “off-page” information, such as information from the web graph, anchortext, and related hyperlinks and from user behavior such as clicks or dwell time [45, 152]. Most recently, as users are visiting the search engines for more diverse reasons [300], the major web search engines are also moving towards semantically informed responses, aiming to interpret a user’s intent and answer the information need behind the query [19]—whether the search engines “follow” changing user behavior or whether users adapt to new functionalities offered by search engines does not really matter for this discussion [143]. Aiming to answer information needs instead of queries involves rather low-level enhancements such as spelling correction, but also more fine-grained user interface enhancements such as query suggestions [9]. A prime example is the Yahoo! query formulation tool called searchassist that we have mentioned as an example in the previous chapter. In [217] we have shown that blending in conceptual information in the query suggestion process can improve such suggestions, especially for rare, infrequent queries. Moving more towards determining the meaning of queries (or, indeed, the information needs behind them), current enhancements include determining the task the user aims to solve [46, 270] or determining the type of information that is being sought (through so-called verticals—which are typically defined as “domain-specific subcollections”) [11, 93]. Another way of attempting to answer the information need behind the query is through semantic analysis, for example by (semi-automatic) expansion of the query using synonyms [113]. Other approaches aim to infer the semantics behind the queries that are submitted [42].
Even other approaches try to understand the “things” that are being sought. For example, using the approach presented by Gabrilovich and Markovitch [107], we can obtain a mapping of free text to concepts (in the form of Wikipedia articles); the same ideas are applied in a more general sense by Turney and Pantel [322]. Medelyan et al. [205] present a comprehensive overview of approaches making use of Wikipedia to extract and make use of the concepts, relations, facts, and descriptions found in Wikipedia.
One of the current goals of the semantic web (in particular the LOD cloud or “web of data”) is to expose, share, and connect data [32, 37]. For this, it uses URIs to identify concepts and provides means by which to describe the concepts themselves as well as any possible relationships with other concepts. One of the current goals of major search engines is very similar: to move beyond a web of pages towards gathering and exposing web-derived knowledge and a “web of things” instead [19]. Indeed, in this chapter we explore what happens when we apply the semantic analysis method from Chapter 6, that links queries to a semantic “backbone,” (in the form of concepts in a concept language). We do so in order to “understand” open domain queries and to estimate query models based on this conceptual information.
In particular, we take the best performing machine learning method from the previous chapter and map queries from the open domain to DBpedia concepts. Then, we apply the most robust relevance feedback method, relevance model 1 (RM-1), from Chapter 4 to the Wikipedia articles associated with the found DBpedia concepts to estimate a query model. The guiding intuition is that, similar to our conceptual query models, concepts are best described by the language use associated with them. In other words, once our algorithm has determined which concepts are meant by a query, we employ the language use associated with those concepts to update the query model. We compare the performance of this approach to pseudo relevance feedback on the collection (in the same way as presented in Chapter 4) and to pseudo relevance feedback on Wikipedia (similar to the way we obtain conceptual query models in Chapter 5).
The research questions we address in this chapter are as follows.
The main contribution presented in this chapter is to provide an indication to what extent the LOD-based semantic analysis presented in the previous chapter can be applied for query modeling in the open domain. In this chapter, we therefore make use of the TREC Terabyte 2004–2006 (TREC-TB) and TREC Web 2009, Category A (TREC-WEB-09) test collections as introduced in Section 3.3. Recall that TREC Terabyte uses the .GOV2 document collection, a large crawl of the .gov domain. TREC Web 2009 uses the ClueWeb09 document collection, a realistically sized web collection. In the experiments in this chapter we use the largest subset, Category A. The topics associated with the TREC Web 2009 test collection are taken from a search engine’s log and representative of queries submitted to a web search engine.
We continue this chapter in Section 7.1 by introducing our method for obtaining DBpedia concepts from ad hoc queries. In Section 7.2 we detail how we estimate the query models as well as the experimental setup used. We discuss results in Section 7.3 and end with a concluding section.
| Up | Next | Prev | PrevTail | Front |