Skip to content

Python library and tools for working with term-document matrices

License

Notifications You must be signed in to change notification settings

jtauber/termdoc

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

78 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

termdoc

Python library and tools for working with term-document matrices

This project is just beginning, but a useful data structure, HTDM (hierarchical term-document matrix), is already available and TF-IDF calculations can be performed.

I am also interested in standardizing the exchange format for HTDMs.

Installation

pip install termdoc

Python 3.10 or higher is required.

HTDM

The core data structure is a term-document matrix supporting hierarchical documents. Documents are labelled with a delimited string such as "1.7.5" or "Plato.Republic.5". This hierarchy could represent divisions of a work, grouping of multiple works, or some combination. Counts are aggregated at each level of the hierarchy (including at the top level to get totals across all documents).

HTDMs can be loaded with load:

>>> import termdoc
>>> c = termdoc.HTDM()
>>> c.load("test_data/test1.tsv")

where the file looks something like:

1.1	foo	7
1.1	bar	4
1.2	foo	2
1.3	bar	1
2.1	baz	5
2.1	foo	1

with a period-separated hierarchical address / document ID, term, and count all separated with tabs.

Both the period and tab are just defaults and can be override by passing address_sep and/or field_sep to load.

The HTDM can then give counts at any level of the document hierarchy:

>>> c.get_counts()["foo"]
10
>>> c.get_counts("1")["foo"]
9
>>> c.get_counts("1.2")["foo"]
2

Note the separator used in the address / document ID defaults to a period (regardless of what was used in load) but can be override by passing address_sep to the HTDM constructor.

HTDMs can also be built up programmatically.

Here is an example with a single-level of documents (a traditional TDM):

>>> import termdoc
>>> c = termdoc.HTDM()
>>> c.increment_count("1", "foo", 3)  # document 1 contains the term "foo" 3 times
>>> c.increment_count("1", "bar", 2)
>>> c.increment_count("2", "foo", 2)
>>> c.increment_count("2", "bar", 1)
>>> c.get_counts()["foo"]
5
>>> c.get_counts()["bar"]
3

And here is an example with a two-level hierarchy:

>>> import termdoc
>>> c = termdoc.HTDM()
>>> c.increment_count("1.1", "foo", 3)
>>> c.increment_count("1.2", "foo", 3)
>>> c.increment_count("1.1", "bar", 2)
>>> c.increment_count("1.2", "bar", 2)
>>> c.increment_count("2.1", "foo", 2)
>>> c.increment_count("2.2", "foo", 2)
>>> c.increment_count("2.1", "bar", 1)
>>> c.increment_count("2.2", "bar")
>>> c.get_counts()["foo"]
10
>>> c.get_counts()["bar"]
6
>>> c.get_counts("2")["foo"]
4

Note that if the count is 1 you can omit it.

Entire lists of tokens can be added for a particular address in one go using add(address, term_list):

>>> import termdoc
>>> c = termdoc.HTDM()
>>> c.add("1.1", ["foo", "bar", "bar", "baz"])
>>> c.add("1.2", ["foo", "foo"])
>>> c.get_counts()["bar"]
2
>>> c.get_counts()["foo"]
3
>>> c.get_counts("1.2")["foo"]
2

You can prune a HTDM to just n levels with the method prune(n).

You can iterate over the document-term counts at the leaves of the HTDM with the method leaf_entries() (this returns a generator yielding (document_address, term, count) tuples). This is effectively a traditional TDM (the document IDs will still reflect the hierarchy but the aggregate counts aren't present).

You can optionally pass a prefix to leaf_entries() in each case only that subtree will be returns (with the prefix removed from the document IDs).

You can graft one HTDM under another by using the graft(prefix, subtree) method, specifying as prefix the document address you want to add the subtree under. This is useful if you have an HTDM for, say, a single work by an author, with chapters as documents and you want to incorporate that into a higher-level HTDM of multiple works by the author, or a collection of works by different authors.

Alternatively you can provide a prefix to load to load the file under a particular point in the tree. This effectively prepends the given prefix (plus address separator) to the document IDs.

The third (count) field in a loaded file can be omitted if the count is 1 and a document ID + term may be repeated with the counts accumulating.

You can deep copy an HTDM with copy(). You can also pass a prefix to copy() if you want to clone just a subtree.

You can save out an HTDM with save() which takes a filename and optional field_sep (defaulting to tab) and prefix (if you just want to save out a subtree).

Calculations

You can get a term frequency with tf(term) or tf(term, address).

>>> c = termdoc.HTDM()
>>> c.increment_count("1", "foo")
>>> c.increment_count("1", "bar", 3)
>>> c.increment_count("2", "foo", 3)
>>> c.increment_count("2", "bar")
>>> c.tf("foo")
0.5
>>> c.tf("foo", "2")
0.75

You can also get document frequency with df(term).

>>> c = termdoc.HTDM()
>>> c.add("1.1", ["foo", "bar"])
>>> c.add("1.2", ["bar"])
>>> c.add("2.1", ["foo"])
>>> c.add("2.2", ["foo", "bar", "baz"])
>>> c.df("foo")
0.75
>>> c.df("bar")
0.75
>>> c.df("baz")
0.25

By default this treats the leaves of the tree at the documents but you can instead specify an explicit number of levels to go down. For example this following will only tree the 1 and 2 as the documents (not 1.1, 1.2, 2.1, 2.2):

>>> c.df("foo", level=1)
1.0
>>> c.df("bar", level=1)
1.0
>>> c.df("baz", level=1)
0.5

Furthermore you can scope the calculate to a subtree, in this case just the documents 1.1 and 1.2 under 1:

>>> c.df("foo", "1")
0.5
>>> c.df("bar", "1")
1.0
>>> c.df("baz", "1")
0.0

This scoping can be combined with the level limit.

The TF-IDF of a term and document can be calculated with tf_idf(term, address):

>>> c = termdoc.HTDM()
>>> c.add("1", ["this", "is", "a", "a", "sample"])
>>> c.add("2", ["this", "is", "another", "another", "example", "example", "example"])
>>> c.tf_idf("example", "1")
0.0
>>> c.tf_idf("example", "2")
0.12901285528456335

You can also scope the TF-IDF calculation to a subtree.

>>> c = termdoc.HTDM()
>>> c.add("1.1", ["foo", "bar"])
>>> c.add("1.2", ["bar"])
>>> c.add("2.1", ["foo"])
>>> c.add("2.2", ["foo", "bar", "baz"])

>>> c.tf_idf("foo", "1.1", "1")
0.1505149978319906

Note that the address must start with the subtree prefix if the latter is given.

Duplicates Policy

You can optionally pass in a duplicates setting to the constructor indicating the policy you want to follow if a term-document count is updated more than once.

>>> c = termdoc.HTDM()
>>> c.increment_count("", "foo", 3)
>>> c.increment_count("", "foo", 2)
>>> c.get_counts()["foo"]
5

is the same as

>>> c = termdoc.HTDM(duplicates=termdoc.Duplicates.ALLOW)
>>> c.increment_count("", "foo", 3)
>>> c.increment_count("", "foo", 2)
>>> c.get_counts()["foo"]
5

But you can tell an HTDM to ignore attempts to update an existing count:

>>> c = termdoc.HTDM(duplicates=termdoc.Duplicates.IGNORE)
>>> c.increment_count("", "foo", 3)
>>> c.increment_count("", "foo", 2)
>>> c.get_counts()["foo"]
3

or to raise an exception:

>>> c = termdoc.HTDM(duplicates=termdoc.Duplicates.ERROR)
>>> c.increment_count("", "foo", 3)
>>> c.increment_count("", "foo", 2)  # this will raise a ValueError
Traceback (most recent call last):
ValueError: 'foo' already in ''

Note that duplicates are only checked in the leaves of the document tree.

About

Python library and tools for working with term-document matrices

Resources

License

Stars

Watchers

Forks

Sponsor this project