Text Search-Enhanced with Types and Entities - Text Mining: Classification, Clustering, and Applications - page 254

Features Coarse Fine

Question trigrams 91.2 77.6

All question q grams 87.2 71.8

All question unigrams 88.4 78.2

Question bigrams 91.6 79.4

+informer q-grams 94.0 82.4

+informer hypernyms 94.2 88.0

Question unigrams + all informer 93.4 88.0

Only informer 92.2 85.0

Question bigrams + hypernyms 91.6 79.4

FIGURE 10.11 : Percent accuracy with linear SVMs, “perfect” informer
spans and various feature encodings. The 'Coarse' column is for the 6 top-
level UIUC classes and the 'fine' column is for the 50 second-level classes.

10.3 Scoring Potential Answer Snippets
In Section 10.2 we established that atypes can be inferred from a natural
language question with high accuracy. The atype extraction step is an
important part of question preprocessing, because it lets us partition question
tokens into

• Tokens that express the user's information need as a type to be

instantiated, but which need not literally appear in a correct response
document or snippet, and

• Tokens that the user expects to literally match correct response

documents or snippets—we call these selector tokens.

For example, the question “What is the distance between Paris and Rome?”
gets partitioned into

• Atype NUMBER:distance (UIUC system) or distance#n#3 (WordNet

system)

• Selectors Paris and Rome thatcanbeusedtoshortlistdocumentsand

snippets that qualify to be scored

In this section we set up a machine learning framework to assign scores to
snippets that potentially answer the question.
In traditional Information Retrieval, the extent of match between the query
q and a candidate document d is often measured as the cosine of the angle
between q and d represented as vectors in the Vector Space Model (33). Each
word in the lexicon is represented by an axis in the vector space. Words