Search

In this homework, you’ll implement a basic search engine by defining your own Python classes. A search engine is an algorithm that takes a query and retrieves the most relevant documents for that query. In order to identify the most relevant documents, our search engine will use term frequency–inverse document frequency (tf–idf), an information statistic for determining the relevance of a term to each document from a corpus consisting of many documents.

The tf–idf statistic is a product of two values: term frequency and inverse document frequency. Term frequency computes the number of times that a term appears in a document (such as a single Wikipedia page). If we were to use only the term frequency in determining the relevance of a term to each document, then our search result might not be helpful since most documents contain many common words such as “the” or “a”. In order to downweight these common terms, the document frequency computes the number of times that a term appears across the corpus of all documents.

Outside Sources¶

Update the following Markdown cell to include your name and list your outside sources. Submitted work should be consistent with the curriculum and your sources.

Name: YOUR_NAME_HERE

1. Enter your outside sources as a list here, or remove this line if you did not consult any outside sources at all.

Task: Document¶

Write and test a Document class in the code cell below that can be used to represent the text in a web page and includes methods to compute term frequency. (But not document frequency since that would require access to all the documents in the corpus.)

The Document class should include:

1. An initializer that takes a str path and initializes the document data based on the text in the specified file. Assume that the file exists, but that it could be empty. In order to implement term_frequency later, we’ll need to precompute and save the term frequency for each term in the document in the initializer as a field by constructing a dictionary that maps each str term to its float term frequency. Term frequency is defined as the count of the given term divided by the total number of words in the text.

Consider the term frequencies for this document containing 4 total words: “the cutest cutest dog”.

• “the” appears 1 time out of 4 total words, so its term frequency is 0.25.
• “cutest” appears 2 times out of 4 total words, so its term frequency is 0.5.
• “dog” appears 1 time out of 4 total words, so its term frequency is 0.25.

When constructing this dictionary, call the clean function to convert the input token to lowercase and ignore non-letter characters so that “corgi”, “CoRgi”, and “corgi!!” are all considered the same string “corgi” to the search algorithm.

1. A method term_frequency that takes a given str term and returns its term frequency by looking it up in the precomputed dictionary. Remember to normalize the term before looking it up to find the corresponding match. If the term does not occur, return 0.

2. A method get_path that returns the str path of the file that this document represents.

3. A method get_words that returns a set of the unique, cleaned words in this document.

4. A method __repr__ that returns a string representation of this document in the format Document('{path}') (with literal single quotes in the output) where {path} is the path to the document from the initializer. The __repr__ method is called when Jupyter Notebook needs to display a Document as output, so we should be able to copy the string contents into a new code cell and immediately run it to create a new Document.

For each of the 4 methods (excluding the initializer) in the Document class, write a testing function that contains at least 3 pytest-style assertions based on your own testing corpus and written as expressions like 1 / 5 that show your thinking. As always, your test cases should expand the domain and range. Documentation strings are optional for testing functions.

• We’ve provided some example corpuses in the doggos directory and the small_wiki directory. For this task, create your own testing corpus by creating a New Folder and adding your own text files to it.
• Be sure to exhaustively test your Document class before moving on: bugs in Document will make implementing the following SearchEngine task much more difficult.
• Line numbers reported in error messages will be off by one due to the first line of code required to run %%ipytest. Look below the indicated line. Line comments (lines that begin #) may also result in different line numbering.

Task: SearchEngine¶

Write and test a SearchEngine class in the code cell below that represents a corpus of Document objects and includes methods to compute the tf–idf statistic between a given query and every document in the corpus. The SearchEngine begins by constructing an inverted index that associates each term in the corpus to the list of Document objects that contain the term.

To iterate over all the files in a directory, call os.listdir to list all the file names and join the directory to the filename with os.path.join. The example below will print only the .txt files in the doggos directory.

The SearchEngine class should include:

1. An initializer that takes a str path to a directory such as "small_wiki" and a str file extension and constructs an inverted index from the files in the specified directory matching the given extension. By default, the extension should be ".txt". Assume the string represents a valid directory, and that the directory contains only valid files. Do not recreate any behavior that is already done in the Document class—call the get_words() method! Create at most one Document per file.

2. A method _calculate_idf that takes a str term and returns the inverse document frequency of that term. If the term is not in the corpus, return 0. Inverse document frequency is defined by calling math.log on the total number of documents in the corpus divided by the number of documents containing the given term.

3. A method __repr__ that returns a string representation of this search engine in the format SearchEngine('{path}') (with literal single quotes in the output) where {path} is the path to the directory specified in the initializer.

4. A method search that takes a str query consisting of one or more terms and returns a list of relevant document paths that match at least one of the cleaned terms sorted by descending tf–idf statistic: the product of the term frequency and inverse document frequency. If there are no matching documents, return an empty list.

For each of the 3 methods (excluding the initializer) in the SearchEngine class, write a testing function that contains at least 3 pytest-style assertions based on your own testing corpus and written as expressions like 1 / 5 that show your thinking. Test cases for SearchEngine.__repr__ may use the given corpuses doggos and small_wiki in addition to your corpus. Documentation strings are optional for testing functions.

• The type checker can be too picky about the sorted key parameter: a type error here can be safely ignored if the code works as intended.

We recommend the following iterative software development approach to implement the search method.

1. Write code to handle queries that contain only a single term by collecting all the documents that contain the given term, computing the tf–idf statistic for each document, and returning the list of document paths sorted by descending tf–idf statistic.

2. Write tests to ensure that your program works on single-term queries.

3. Write code to handle queries that contain more than one term by returning all the documents that match any of the terms in the query sorted by descending tf–idf statistic. The tf–idf statistic for a document that matches more than one term is defined as the sum of its constituent single-term tf–idf statistics.

4. Write tests to ensure that your program works on multi-term queries.

Here’s a walkthrough of the search function from beginning to end. Say we have a corpus in a directory called "doggos" containing 3 documents with the following contents:

• doggos/doc1.txt with the text Dogs are the greatest pets.
• doggos/doc2.txt with the text Cats seem pretty okay
• doggos/doc3.txt with the text I love dogs!

The initializer should construct the following inverted index.

{"dogs":     [doc1, doc3],
 "are":      [doc1],
 "the":      [doc1],
 "greatest": [doc1],
 "pets":     [doc1],
 "cats":     [doc2],
 "seem":     [doc2],
 "pretty":   [doc2],
 "okay":     [doc2],
 "i":        [doc3],
 "love":     [doc3]}

Searching this corpus for the multi-term query "love dogs" should return a list ["doggos/doc3.txt", "doggos/doc1.txt"] by:

1. Finding all documents that match at least one query term. The word "love" is found in doc3 while the word "dogs" is found in doc1 and doc3.

2. Computing the tf–idf statistic for each matching document. For each matching document, the tf–idf statistic for a multi-word query "love dogs" is the sum of the tf–idf statistics for "love" and "dogs" individually.

   1. For doc1, the sum of 0 + 0.081 = 0.081. The tf–idf statistic for "love" is 0 because the term is not in doc1.

   2. For doc3, the sum of 0.366 + 0.135 = 0.501.

3. Returning the matching document paths sorted by descending tf–idf statistic.

After completing your SearchEngine, run the following cell to search our small Wikipedia corpus for the query “data”. Try some other search queries too!