Information Retrieval

Abstract

Information retrieval (IR) is the identification of documents or other units of information in a collection that are relevant to a particular information need—a set of questions to which someone would like to find an answer. This chapter presents the basic strategies of IR in length along with their analysis, particularly emphasizing the vector space and probabilistic models of IR, with worked examples in each category; gives the detailed coverage to construction and maintenance of index, and its parallel processing. The fuzzy logic-based retrieval, concept-based retrieval techniques, their algorithms, and worked examples are presented; and Automatic Query Expansion has been dealt with at length. Application of Bayesian networks, and inferences using these have been demonstrated for IR. The newly emerged semantic web for futuristic IR and its applications have been introduced; and the design aspects of distributed IR suited for currently distributed information resources are treated in depth. The chapter ends with the summary and a set of practice exercises.

Exercises

• 1. Show how the vector space model can be modeled using an inference network.

• 2. Consider a documents collection made of 100 documents. Given a query q, the set of documents relevant to the users is $D^*= \{d_4 , d_{15} , d_{34}, d_{56}, d_{98}\}$. An IR system retrieves the following documents $D = \{d_4, d_{15}, d_{35}, d_{56}, d_{66}, d_{88} , d_{95}\}$

  • a. Compute the number of True-Negatives, True-Positives, False-Negatives, False-Positives.

  • b. Compute Precision, Recall, and F-measure.

• 3. Consider an IR scenario in the following: It has been found in some hospital, results of blood tests taken on a specific day are unreliable for diabetic patients due to equipment malfunction. The hospital uses an IR system to identify these patients. Suppose the collection of patients’ records contains 10, 000 documents, 500 of that are relevant to the query. The system returns 350 documents, 225 of that are relevant to the query. Answer the following for this scenario:

  • a. Calculate the precision and recall for this system.

  • b. Based on your results from above, explain how well would you say about the working of hospital’s IR system.

  • c. Knowing about the precision-recall trade-off, what is likely to happen if an IR system is tuned to aim for 100% precision?

  • d. Knowing about the precision-recall trade-off, what is likely to happen if an IR system is tuned to aim for 100% recall?

  • e. For the trade-off given scenario, which measure do you think is more important, precision or recall? Why?

• 4. You are looking for information on “Economic growth in India” in a large document collection, during the period of last 3 years. You decide to search using the terms: India, banks, growth, economy, business, agriculture, using an IR system, which recommends three possible documents given below with term frequencies.

  Term	Economy	India	Growth	Banks	Business	Agriculture
  Document-1	15	10	3	4	2	9
  Document-2	0	0	9	8	7	8
  Document-3	4	2	4	4	6	10

  There is no additional information about the documents. Make use of each of the following models to find out the relevancy of the documents to the query.

  • a. Boolean model

  • b. Vector space model

  • c. tf.idf model

• 5. Take any three small documents of size, approximately 100 words.

  • a. Build a matrix of an inverted index for these documents, in the format shown in Fig. 18.5.

  • b. Weight terms by their presence/absence (binary), and also and by $tf\times idf$ (with estimated IDFs).

  • c. Compute the memory requirements for this inverted index. Make necessary assumptions for character size, pointer size, etc.

  • d. Construct a suitable query, and calculate document–query similarity, for the following scenarios:

    • i Cosine (with normalization)

    • ii Inner product (i.e., cosine without normalization)

    • iii Does the normalization has any effect? Justify.

• 6. Consider that we submit the queries to search engines for searching the needed information on WWW.

  • a. Does the search process use a stop-word list?

  • b. Can you search “The”, “The a”, “An a”, etc.? Justify.

  • c. Is it a practice to search the above terms?

  • d. Does the search process use stemming?

  • e. Are there different results for two queries “Human body”, “Humanly body”. Justify your answer.

  • f. Does it normalize words to lower case?

• 7. “Having the knowledge of the sense of a query term may help a document retrieval system, especially for short queries.” Why it is not true for longer queries?

• 8. Comment on the validity of following statements for Boolean model:

  • a. “Stemming does not lower the precision of a Boolean retrieval system.”

  • b. “Stemming does not lower recall of a Boolean retrieval system.”

• 9. Answer the following in brief:

  • a. Why is the idf of a term always finite?

  • b. What is the idf of a term that occurs in every document?

  • c. What is the idf of a term that appears in one document only?

  • d. What is the idf of a term that appears in no document?

• 10. Answer the following in brief:

  • a. Name three criteria for evaluating a search engine.

  • b. What is an easy way to maximize the recall of a search engine?

  • c. What is an easy way to maximize the precision of a search engine?

• 11. What is the difference between clustering and classification? How can they be used in a complete IR system?

• 12. Discuss the merits and demerits of following, suggest as to which one will provide better response time?

  • a. Document-distributed architecture.

  • b. Term-distributed architecture.