1. A Novel Pattern Learning Method for Open Domain Question Answering IJCNLP 2004 Yongping Du, Xuanjing Huang, Xin Li, Lide Wu

2. Abstract They develop a novel pattern based method for implementing answer extraction in QA For each type of question, the corresponding answer patterns can be learned from the Web automatically Given a new question, these answer patterns can be applied to find the answer They give a performance analysis of this approach using the TREC-11 question set

3. Introduction Three main components in QA system Many other question answering systems have used pattern based method Patterns are not learned They can handle only one question term in the candidate answering sentence

4. Introduction (Cont.) Each answer pattern is consisted of three parts + [ConstString] + : the key phrases of question : the answer, any string holding the position will be extracted as the answer [ConstString]: a sequence of words indentification -> learning answer patterns -> answer extraction -> performance analysis

5. Question Analysis A set of symbols are defined to represent questions Q_Focus: the head word of the noun phrase which is binding with interrogative Interrogative + be verb + NP… How + Adj+ be verb…

6. Question Analysis They select 182 questions from TREC, and these questions contain all the Q_Tag symbols Every element of the questions is replaced with its corresponding Q_Tag The classification of questions will be built based on the Q_Pattern and the answer type Six answer type

7. Example Question: “What is the largest city in Germany?” Question type: [LCN] What Q_BeVerb Q_Focus in Q_LCN? Doc 1: “…,Berlin is the largest city in Germany and is develop...” Answer pattern: “, Q_BeVerb Q_Focus in Q_LCN”

8. Pattern Learning 1. Constructing Query: “Q_Tag + Answer”, for example, “the largest city” + ”Germany” + “Berlin” 2. Searching: Submit query into Google, and download the top 100 documents 3. Snipped Selection: Extract the snippets containing 10 words around the answer 4. Answer Pattern Extraction: Replace the question term in each snippet by the corresponding Q_Tag, and the answer term by the tag The shortest string containing the Q_Tag and the tag is extracted as the answer pattern

9. Pattern Learning (Cont.) 5. Computing the weight of each answer pattern They learn answer patterns for each question type

10. Pattern Evaluation answer pattern “ Q_BeVerb Q_Focus” may extract candidate answer “Portland” from the snippet “Portland is the largest city in Oregon…” They provide an approach to evaluate the answer patterns 1. Query for each answer pattern of the question is formed and submitted to Google, and top 100 snippets are downloaded for evaluation The query consists of three parts: [Head] + [Tail] + [Q_Focus + Q_NameEntity] answer pattern “ Q_BeVerb Q_Focus” query: “is the largest city” + “Germany”

11. Pattern Evaluation (Cont.) 2. The confidence of each answer pattern is calculated by the formula 3. At last the score of each answer pattern is computed as the formula The major advantage over other pattern based QA systems is that more than one question term can be included in the answer pattern

12. Sample Answer patterns

13. Answer Extraction Use google as the search engine and get top 100 snippets for answer extraction 1. Identify the Q_Tag of the new question and generate its Q_Pattern “What is the most populous city in the United States?” What Q_BeVerb Q_Focus in Q_LCN? 2. Determine the question type of the question. The corresponding answer patterns of this question type are also selected 3. Replace the Q_Tag symbols of each answer pattern with the corresponding question term of the question “, Q_BeVerb Q_Focus in Q_LCN”, is the most populous city in the United States

14. Answer Extraction (Cont.) 4. Select the words matching tag as the candidate answer 5. Discard the candidate answers which do not satisfy the answer type of the question, using name entity tagger 6. Sort the remainder candidate answers by their answer pattern’s scores and their frequencies. The highest score is selected as the final answer

15. Performance Analysis TREC-9 and TREC-10 are training examples to learn answer patterns 500 questions of TREC-11 are experimented

16. Conclusion They take part in the TREC-12 this year and their result is above the median score of all runs submitted Some answer patterns they learned are too specific and are useless for answering new question “What is a shaman?” “Q_Focus was the priest, the and”