The basis of a search engine is a collection of documents converted to a machine-readable representation, and classified into broad topics. Imagine doing search on a small set of 1M documents. You might have 100K documents in computer science topic, 100K in the animals topic, another 100K in a politics topic, etc. When a user query comes in, say How do I learn Python?, you ideally want to classify the query into the computer science topic, so that you only have to search through 100K documents rather than the whole 1M document collection.
In this practical, we will use Wikipedia to create a small collection of documents on certain topics. We will convert those documents into vectors. And finally, we will test whether a given user query is mapped to the correct topic in the collection.
We will first retrieve a number of topics from the Wikipedia category tree. We will use the Wikimedia API for this:
python get_categories.py
The code will write out all Wikipedia categories for which at least 200 documents exist, creating the file wiki_categories.txt. (It takes a minute or two, so be patient! In case you have difficulties running the code, you will find an already processed list in the data folder: data/preprocessed_wiki_categories.txt.)
Open your wiki_categories.txt file and scroll through it. Select 10 categories from it. Make sure you don't select categories which look like Wikipedia-internal labels (e.g. Wikipedia books', Pages with script errors', etc). Copy and paste those category names into a file named tmp_cat.txt in the top directory of the repo. For instance, you might end up with the following:
Fast-food franchises
Fauna by country
Fauna of Brazil
Fauna of California
Fauna of the Amazon
Fauna of the California chaparral and woodlands
Fauna of the Southeastern United States
Featherstone Rovers players
Featherweight boxers
Featherweight mixed martial artists
We will now retrieve Wikipedia pages for our categories. We first need a list of page titles for each category. To do so, we will use the API again and run the following:
python get_category_pages.py
This will create a file titles.txt for each category in your data/ directory, containing 100 Wikipedia titles for each of the categories in your tmp_cat.txt file. We now need to retrieve the text of those documents. To speed things up and in order not to use too much hard disk space, we will just retrieve the intro text of each document rather than the whole page. Run:
python get_page_content.py
This will create a file linear.txt for each category in your data/ directory, containing the introductory text of each Wikipedia page listed in titles.txt files. The calls to the Wikipedia API may take a little while, so you can start reading the rest of this page while you wait.
Before carrying on, check the files that have been created in the data processing step. This will give you an idea of the documents you have in your collection.
Let's now convert our raw texts into sets of features. We will do this using a vocabulary of character ngrams. The repository contains scripts to compute ngram frequencies for each category. You can run it doing:
python ngrams.py [ngram-size]
For instance,
python ngrams.py 6
will create a file in each subfolder of data/categories, showing a sorted list of ngrams of size 6.
We probably want several ngram sizes. So re-run the script for size 4-6, for instance.
Which of our features should we use to generate document representations? A common way to do this in Information Retrieval is the TF-IDF measure.
To compute the td-idf of each ngram in each category, run:
python output_top_tfidf_ngrams.py
This will create a tf_idfs.txt file in each category subfolder. In addition, the script captures the 100 ngrams with highest tf-idf for each category and uses them to generate a set of features, saved in the file data/vocab.txt. In essence, what we are doing here is simple feature selection, based on the highest-ranked ngrams across categories.
Now that we have our features, we can build vector representations for each category. If we have ended up with 1000 features in our feature selection step, we will build 1000-dimensional vectors to represent each category, with each component of a vector set to the tf-idf of the relevant dimensions for that category. As an example, if actor is the 16th dimension in our 1000-dimensional space, and category russian_language_films has a tf-idf of 1.147e-05 for that ngram, then dimension 16 will be set to that number in the russian_language_film vector.
You can create your category vectors with
python mk_cat_vectors.py
Our final goal for now is to see how we can map incoming search queries to each category. The point of this is that, as our search engine grows (to potentially millions, or billions of pages), it would be very inefficient to try and find the answer to a movie-related query in our "gecko and other lizards" category. We want to narrow things down.
So let's test this. The file query_file.txt contains some sample queries. You should change this file and write down queries corresponding to the categories you selected, one query per line. We want to see whether the system finds the right category for each query. Run the following:
python classify.py query_file.txt
and you will return a ranked list of categories for each query. You of course want to most relevant categories at the top.
Play with the classify.py script, and with adding queries to the file. What works, what doesn't? Why?