Chapter 2 Background
2.2 Approaches to Question Answering
2.2.5 Web-based Approaches
The vast amount of information available on the World Wide Web makes it an attractive external resource for question answering (Clarke, Cormack, and Lynam, 2001; Clarke et al., 2001; Breck et al., 2001; Dumais et al., 2002; Lin, 2002). Web data is viewed as an enormous collection of unstructured, flat texts with tremendous amount of data re- dundancy. Its immense size qualitatively changes the nature of question answering task as compared to the same task on close corpora, such as news paper texts, encyclope- dias, etc. Due to the data redundancy on the Web, any piece of information might be stated in a variety of ways in different documents. Let’s borrow an example from (Lin, 2002): considering the question ”Who killed Lincoln?”, there are two possible answer sentences:
1. John Wilkes Booth killed Lincoln.
2. John Wilkes Booth is perhaps America’s most infamous assassin. He is best known for firing the bullet that ended Abraham Lincoln’s life.
Obviously, the answer could be much more easily extracted from the sentence(1) than the passage(2). QA turns to be easier if answer sentences are stated as the simple reformula- tions of questions. In this case, simple techniques, such as keyword-based retrieval and surface pattern-based matching, can perform well although they are not good enough on small corpora. The larger the text collection is, the greater the possibility of having sim- ple statement is. Therefore, data redundancy can be used as a surrogate for sophisticated natural language techniques.
Moreover, with the increase of data size, the quality and credibility of individual documents are decreased. Some documents are poorly written, or contain incorrect in-
formation. As a result, answers extracted from a single document might not be trustable enough to be globally correct. Data redundancy alleviates this issue since multiple oc- currences of an answer in different documents lead to the higher credibility level.
Considering the above benefits, more and more QA systems incorporate Web resource. The Web resource is mainly used in the two modules regarding the whole system architectures: Answer Extraction Module (Wu et al., 2003; Wu et al., 2004; Wu et al., 2005; Kaisser and Becker, 2004) and Answer Validation Module (Harabagiu et al., 2005; Xu et al., 2002).
As the most representative work of using the Web resource in Answer Extraction Module, the QA system of Fudan University (Wu et al., 2003; Wu et al., 2004; Wu et al., 2005) is solely built on the Web data. They retrieve relevant snippets from the Web using Google and find answers from the snippets. To construct queries for Google search, they parse questions using LinkParser and then extract four constituents from the parsed questions: subject, predicate, object and adverbial modifier. Next, queries are formulated according to the constituents of the questions. For example, they splitted the question ”What book did Rachel Carson write in 1962?” into the following constituents:
1. Rachel Carson - subject 2. wrote - predicate
3. in 1962 - adverbial modifier
The following queries from tight to loose are formulated from the above con- stituents.
1. ”Rachel Carson wrote” ”in 1962” 2. ”Rachel Carson wrote” ”in 1962” 3. ”Rachel Carson” wrote in 1962 4. Rachel Carson wrote in 1962
From relevant snippets, they extract answers simply using a surface pattern match- ing method. Different from other methods, they abstract question phrases using different
classes:
• Q Quotation: the quotation parts in a question;
• Q Focus: the key words representing the object or event which a question asks about;
• Q NamedEntity: the name entities in a question; • Q Verb: the main verb of a question;
• Q BNP: the noun phrases of a question.
The different classes of question phrases are assigned different weights in surface pat- terns. Surface pattern matching scores are calculated by incorporating the weights.
(Kaisser and Becker, 2004) manually construct 157 surface patterns and make strict and fussy surface pattern matching on the Web data. They further use a Google Fallback Mechanism to backup the surface pattern matching methods. The Google Fall- back Mechanism exploits n-gram information matching on Google snippets. Final scor- ing function is designed as the linear interpolation of the strict surface pattern matching, the fussy surface pattern matching and the Google fallback matching. The experiments on TREC 2004 data show that the fallback mechanism performs better than the surface pattern matching methods.
Besides directly finding answers from the Web, researches also use the Web data in Answer Validation Module. It may effectively overcome the locally correct answer problem. (Harabagiu et al., 2005) explore ”Web-boosting” features based on a web strat- egy that utilizes general linguistic patterns to construct a series of search engine queries. Once an answer candidate from TREC collection also occurs in web documents, an ad- ditional feature capturing web redundancy information will be fired to boost the answer candidate. As a result, the feature leads to another ranking of the answer candidates pro- duced by the original Answer Extraction Module. The evaluation results on TREC 2005 data show that ”web-boosting” provide an added value of 69/331 to final factoid score.
(Xu et al., 2002) apply the Web data to supplement TREC corpus. For efficiency consideration, they look for answers in Top 100 Google hits for a Web search and confine whole web pages to short summaries in order to further reduce processing cost. Two measures are proposed to incorporate the Web information:
1. The confidence of the answer A found from the Web is a function of the question type T and the answer frequency F in Top 100 Google summaries. Specifically,
p(correct|Q, A) = p(correct|T, F )
= p(correct|T ) × 0.5 + P (correct|F ) × 0.5
where, T = question type, F = frequency of A in Google summaries
2. The confidence of the answer A is a function of its frequency F in Top 100 Google summaries and a Boolean variable INT REC, which is true if and only if A is also returned from the TREC corpus. Specifically,
p(correct|Q, A) = p(correct|F, INT REC)
As a result, they also confirm the positive findings reported in the earlier stud- ies (Dumais et al., 2002). The experiments on TREC 2002 data show that answer fre- quency in Top 100 Google summaries is a strong predictor of answer correctness.