Ravichandran and Hovy, ACL 2002: Learning Surface Text Patterns for a Question Answering System
Ravichandran, D. and Hovy, E. 2002. Learning Surface Text Patterns for a Question Answering System. Proceedings of the 40th Annual Meeting on Association for Computational Linguistics. 41--47
An online version of this paper is available .
This paper introduces a method to learn surface patterns from text and use them to answer factoid style questions.
The authors consider the problem of answering factoid questions (e.g. When was X born?). Some possible approaches for this problem include using NER, WordNet, parsers, hand-tagged corpora, and ontology lists. It turns out that a very powerful approach is to use surface text patterns, e.g. “<NAME> was born in <BIRTHDATE>”, “<NAME> (<BIRTHDATE>–”. One could create such patterns manually, but this is time-consuming, and one needs to create patterns afresh for each new attribute. Also, there's a limit on how much data can be looked at, and the method is not very accurate. The authors therefore try to learn such patterns automatically and with high precision.
- Use bootstrapping to build a large set of patterns starting with only a few examples of QA pairs.
- Rank learned patterns in order of precision
- Use patterns to answer unseen questions
The authors first select an example QA pair e.g. “Mozart 1756” and submit that to a search engine; they download the top 1000 hits and retain only sentences that contain both terms. They then pass each sentence through suffix tree constructor. This finds all substrings, of all lengths, along with their counts. For example, for the sentences:
- The great composer Mozart (1756–1791) achieved fame at a young age”
- “Mozart (1756–1791) was a genius”
- The whole world would always be indebted to the great music of Mozart (1756–1791)”.
The longest matching substring is “Mozart (1756–1791)”, which the suffix tree would extract as one of the outputs, with score 3. Then, for each phrase in suffix tree, they retain only those phrases that contain both terms. Some examples of learned patterns for the birth-year attribute include:
- born in <ANSWER> , <NAME>
- <NAME> was born on <ANSWER> ,
- <NAME> ( <ANSWER> -
- <NAME> ( <ANSWER - )
Calculating precision of each pattern
The authors query their corpus with only the question term, download the top 1000 hits, and retain only sentences that contain the question term. For each learned pattern, they check for its presence in the sentence for two cases:
- Presence with <ANSWER> matched by any word.
- Presence with <ANSWER> matched by correct A term.
Then, the precision of a pattern is Ca / Co where Ca = # of patterns with correct A term, Co = # of patterns with any word as A term They retain only patterns matching some k # of examples.
Finding answers using learned patterns
The authors determine the question type of a new question and the question term using their existing QA system. They then query the target corpus with this question term, obtain a list of sentences, and use the pattern table for that question type to search for the presence of each pattern. They select words matching the “<ANSWER>” tag as their answers. They sort answers by their generative pattern's precision scores, discard duplicates, and return the top 5 answers.
The authors considered 6 different question types: BIRTHDATE, LOCATION, INVENTOR, DISCOVERER, DEFINITION, WHY-FAMOUS. They took questions from the TREC-10 dataset, and ran two experiments: one in which they used the TREC-10 corpus and performed IR by their own QA system, and another in which they used the web as a corpus and performed IR by using AltaVista.
They find that their approach performs better on the Web than the TREC corpus. The abundance of data on Web compared to TREC probably makes it easier to find candidate answers.
The main drawbacks with the authors' approach include:
- No external knowledge added e.g. POS tags
- Good for factoid questions, not so for definition or list questions
- Cannot handle long-distance dependencies.
- Can handle only one anchor point (the Q term) in candidate A sentence. Cannot work for Qs that requires multiple words from Q to be in the A sentence, possibly apart
- Ad-hoc way of determining length of the answer
The authors present an elegant bootstrapping method to learn patterns that extract answers for factoid questions. This is essentially the same idea as extracting attributes of given entities (Our IE term project). This is an early paper in the field of Attribute Extraction using patterns, and it sets stage for lot of work done later that addresses its shortcomings.