This library contains simple functionality to tackle the problem of segmenting documents into coherent parts. Imagine you don't have a good paragraph annotation in your documents, as it is often the case for scraped pdfs or html documents. For NLP tasks you want to split them at points where the topic changes. Good results have been achieved using topic representations, but they involve a further step of topic modeling which is quite domain dependent. This approach uses only word embeddings which are assumed to be less domain specific. See [https://arxiv.org/pdf/1503.05543.pdf] for an overview and an approach very similar to the one presented here.
The algorithm uses word embeddings to find a segmentation where the splits are chosen such that the segments are coherent. This coherence can be described as accumulated weighted cosine similarity of the words of a segment to the mean vector of that segment. More formally segments are chosen as to maximize the quantity |v|, where v is a segment vector and |.| denotes the l2-norm. The accumulated weighted cosine similarity turns up by a simple transformation: |v| = 1/|v| <v, v> = <v, v/|v|> = \sum_i <w_i, v/|v|> = \sum_i |w_i| <w_i/|w_i|, v/|v|>, where v = \sum_i w_i is the definition of the segment vector from word vectors w_i. The expansion gives a good intuition of what we try to achieve. As we usually compare word embeddings with cosine similarity, the last scalar product <w_i/|w_i|, v/|v|> is just the cosine similarity of a word w_i to the segment vector v. The weighting with the length of w_i suppresses frequent noise words, that typically have a shorter length.
This leads to the interpretation that coherence corresponds to segment vector length, in the sense that two segment vectors of same length contain the same amount of information. This interpretation is of course only capturing information that we are given as input by means of the word embeddings, but it serves as an abstraction.
To optimize for segment vector length |v|, we look for a sequence of split positions such that the sum of l2-norms of the segment vectors formed by summing the words between the splits is maximal. Given this objective without constraints, the optimal solution is to split the document between every two subsequent words (triangle inequality). We have to impose some limit on the granularity of the segmentation to get useful results. This is done by a penalty for every split made, that counts against the vector norms, i.e. is subtracted from the sum of vector norms.
Let Seg := {(0 = t_0 < t_i < ... < t_n = L) | s_i natural number} where L is a documents length. A segment [a, b[ comprises the words at positions a, a+1, ..., b-1. Let l(j, k) := |\sum_i=j^{k-1} w_i| denote the vector of segment [i, j[. We optimize the function f mapping elements of Seg to the real numbers with f: (t_0, ..., t_n) \mapsto \sum_{i=0}^{n-1} (l(t_{i-1}, t_i) + l(t_i, t_{i+1}) - penalty).
There are two variants, a greedy that is fast and a dynamic programming approach that computes the optimal segmentation. Both depend on a penalty hyperparameter, that defined the granularity of the split.
Split the text iteratively at the position where the gain is highest until this gain would be below a given penalty threshold. The gain is the sum of norms of the left and right segments minus the norm of the segment that is to be split.
Iteratively construct a data structure storing the results of optimally splitting a prefix of the document. This results in a matrix storing a score for making a segment from position i to j, given a optimal segmentation up to i.
The greedy implementation does not need the penalty parameter, but can also be
run by limiting the number of segments. This is leveraged by the get_penalty
function to approximately determine a penalty parameter for a desired average
segment length computed over a set of documents.
To measure the accuracy of an algorithm against a given reference segmentation
P_k is a commonly used metric described e.g. in above paper.
The function get_segments simply applies a segmentation determined by one of
the algorithms to e.g. the sentences of a text used when generating the
segmentation.
The algorithms are fed a matrix docmat containing vectors representing the
content of a text. These vectors are supposed to have cosine similarity as a
natural similarity measure and length roughly corresponding to the content
length of a text particle. Particles could be words in which case word2vec
embeddings are a good choice as vectors. The width of docmat is the embedding
dimension and the height the number of particles.
If you want to split text into paragraphs, you most likely already have a good idea of what potential sentence borders are. It makes sense not to give the word vectors as input but sentence vectors formed by e.g. the sum of word vectors, as it is usual practice.
Clone the repo, then install it from within the root folder with
pip install -e .In the Jupyter notebook HowTo.ipynb you find code that demonstrates the use of the module. It downloads a corpus to trains word2vec vectors on and an example text for segmentation. You achieve better results if you compute word vectors on a larger corpus.