As I’ve recently had the joy of struggling with using PyLucene (after many years), I re-entered the void of documentation straight after actually managing to compile and install the thing. I ended up at the five year old blog post “PyLucene 3.0 in 60 seconds — tutorial sample code for the 3.0 API” by Joseph Turian (that conveniently lets one infer syntax and functionalities of PyLucene) many times while googling. However, the example code no longer works, as in pylucene 4.0 some things changed, in particular;
Starting with version 4.0, pylucene changed from a flat to nested namespace, mirroring the java hierarchy. ~ source
I am running PyLucene 4.10.1, so I find whatever I need in the 4.10.1 Javadocs. Below is the PyLucene 3.0 in 60 seconds blogpost example updated for PyLucene 4.0 (and beyond…?), which I figured may be of use to those that start to dabble in PyLucene. Many thanks to Joseph for the original post!
from java.io import File
from org.apache.lucene.analysis.standard import StandardAnalyzer
from org.apache.lucene.document import Document, Field
from org.apache.lucene.index import IndexWriter, IndexWriterConfig
from org.apache.lucene.store import SimpleFSDirectory
from org.apache.lucene.util import Version
if __name__ == "__main__":
indexDir = SimpleFSDirectory(File("index/"))
writerConfig = IndexWriterConfig(Version.LUCENE_4_10_1, StandardAnalyzer())
writer = IndexWriter(indexDir, writerConfig)
print "%d docs in index" % writer.numDocs()
print "Reading lines from sys.stdin..."
for n, l in enumerate(sys.stdin):
doc = Document()
doc.add(Field("text", l, Field.Store.YES, Field.Index.ANALYZED))
print "Indexed %d lines from stdin (%d docs in index)" % (n, writer.numDocs())
print "Closing index of %d docs..." % writer.numDocs()
from java.io import File
from org.apache.lucene.analysis.standard import StandardAnalyzer
from org.apache.lucene.document import Document, Field
from org.apache.lucene.search import IndexSearcher
from org.apache.lucene.index import IndexReader
from org.apache.lucene.queryparser.classic import QueryParser
from org.apache.lucene.store import SimpleFSDirectory
from org.apache.lucene.util import Version
if __name__ == "__main__":
analyzer = StandardAnalyzer(Version.LUCENE_4_10_1)
reader = IndexReader.open(SimpleFSDirectory(File("index/")))
searcher = IndexSearcher(reader)
query = QueryParser(Version.LUCENE_4_10_1, "text", analyzer).parse("Find this sentence please")
MAX = 1000
hits = searcher.search(query, MAX)
print "Found %d document(s) that matched query '%s':" % (hits.totalHits, query)
for hit in hits.scoreDocs:
print hit.score, hit.doc, hit.toString()
doc = searcher.doc(hit.doc)
As I am starting to gather testing data, I figured it’d be a good time to determine how to measure the performance of the different results of my 24 text representation algorithms. What I want to measure per algorithm: how many keywords are predicted the same as the experts, and how many aren’t. After some research and valuable advice, I came across confusion matrices, which seemed appropriate. For each algorithm I measure the amount of:
A. True Negatives (excluded by algorithm & excluded by experts)
B. False Positives (included by algorithm & excluded by experts)
C. False Negatives (excluded by algorithm & included by experts)
D. True Positives (included by algorithm & included by experts)
I found this page from the University of Regina explaining confusion matrices, and decided to implement it. My implementation in pseudocode:
For each text:
fill expertList with expertResults #list of URIs
fill algoList with algorithmResults #list of URIs
algoPOS = algoList #URIs included by algorithm
algoNEG = [item for item in NCIthesaurus if item not in algoPOS] #URIs excluded by algorithm (ontology-algoPOS)
expertPOS = expertList #all URIs included by experts
expertNEG = [item for item in NCIthesaurus if item not in expertPOS] #all URIs excluded by experts (ontology-expertPOS)
A += len(set(algoNEG).intersection(expertNEG)) #True Negatives (number of URIs that overlap in algoNEG and expertNEG)
B += len(set(algoPOS).intersection(expertNEG)) #False Positives (number of URIs that overlap of algoPOS and expertNEG)
C += len(set(algoNEG).intersection(expertPOS)) #False Negatives (number of URIs that overlap in algoNEG and expertPOS)
D += len(set(algoPOS).intersection(expertPOS)) #True Positives (number of URIs that overlap in algoPOS and expertPOS)
matrix.append([[A,B],[C,D]]) #Put numbers in a matrix
With this information I can calculate a set of standard terms: Accuracy (AC), Recall or True Positive Rate (TP), False Positive Rate (FP), True Negative Rate (TN), False Negative Rate (FN), Precision (P).
The only thing with the rates are that proportions are heavily skewed because of the size of the ontology (90.000 URIs), which means the negative cases will always be much more frequent than the positives. This means that accuracy is always around 99.9%, and so is the TN. I still have to figure out exactly what information I want to use and how (visualize?).
First results are in! My computer spent about 20 hours to retrieve and store neighboring concepts of over 10.000 concepts which my Breadth-First-Search algorithm passed through to find the shortest paths between 6 nodes. But here is the result, first the ‘before’ graph, which I showed earlier: all retrieved concepts with their parent relations. Below that is the new graph, which relates all concepts by finding their shortest paths (so far only the orange concepts – from the Gene Ontology).
What were two separate clusters in the first to the left is now one big fat cluster… Which is cool!
Less cool is the time it took… But oh well, looks like I’m going to have to prepare some examples as proof of concepts. Nowhere near realistic realtime performance so far… (however I got a big speed increase by moving my Sesame triple store from my ancient EeePC900 to my desktop computer… Goodbye supercomputer). The good news is that all neighboring nodes I processed so far are cached in a local SQLite database, so those 20 hours were not a waste! (considering my total ontology database consists out of over 800.000 concepts, and 10.000 concepts took 20 hours, is something I choose not to take into consideration however :p).
It is important to note that the meaning or interpretation of the resulting graph (and particularly the relations between concepts) is not the primary concern here: the paths (their lengths, the directions of the edges and the node’s ‘depths’) will be primarily used for the ontology-based semantic similarity measure I wrote about in this post.
The first steps towards creating graph-based text visualizations is thinking about how I would like to visually represent the structure of the information I extract from texts. I have included a preliminary example of the output at the bottom of this page. I use three different methods of extracting concepts from a piece of text:
Counting literal occurrences of concepts (from ontologies)
Finding related concepts by textual comparison of the text to the concepts’ descriptions
Finding related concepts by exploring the ontological structure (aka relating concepts within one ontology by finding paths and parents, and possibly relevant neighbours)
The primary distinction I want to make is that of relevance (aka ‘likeliness of topic’). In the case of 1 this would be the frequency of the word (more occurrences = more relevant concept), in the case of 2 this is the calculated similarity of the source text to the concepts’ descriptions (which is a number between 0 and 1). In the case of 3, this would be by ‘connections’: the more concepts the concepts I find by exploring an ontologies’ structure would link, the more relevant this found concept would be. I want to model this distinction by node’s size: the more relevant a concept is, the bigger I want to draw the concept in the graph.
The second distinction is that of literal to non-literal (1 being literal, 2 and 3 being non-literal). I want to model this distinction by style: literal concepts will be drawn as a filled circle, non-literal concepts as outlined circles.
The third distinction is that of the concepts’ source: from which ontology does a concept originate? This distinction will be modeled by color: each ontology (of the six I use) will have its distinct color. Explored concepts (from step 3) such as parents and shared parents will be colored in distinct colours as well, since they are connected in the graph to the coloured nodes, their source will implicitly be clear.
Since Gephi doesn’t fully support Graphviz’ DOT-language, and the graph library I use in Python conveniently parses graphs in DOT, I use Graphviz to directly render the results.
An issue I came across with the scaling (to represent relevance), is that I’m working with multiple measures: frequency of literal words (1), percentage of text-similarity (2), and degree-count. In an effort to roughly equalize the scaling factor, I decided to use static counters. Each node gets an initial size of 0.5 (0.4 for shared parents, and 0.3 for parents). For each occurrence of a literal word, I add 0.05. For the text-similarity, I add the percentage of 0.5 (26% similarity = 0.5 + (0.5*0.26)). For the degree, I add 0.1 for each in-link the node receives. This is an initial attempt at unifying the results. Anyway, these are just settings I’m playing around with.
These are very rough examples, because:
I use the literal representation of the input text (as I have not yet determined the most suitable keyword extraction method)
I haven’t determined a proper ‘cut-off’ for the text similarity measure; currently it includes every concept it finds with similarity ≥ 25%.
It doesn’t yet fully incorporate step 3 (it includes parents, but not yet paths between nodes)
It doesn’t scale nodes according to in-links
There is no filtering applied yet (removing obsolete classes, for example).
All this time I have been working on my ontology-powered topic identification system: combining semantic web technologies with natural language processing and text-mining techniques to extract (a) topic(s) from a text. But I never really decided what my program should output: a topic? A list of potential topics with their ‘likeliness’? That would mean the relationships between topics my SPARQL-powered ‘pathfinder’ finds would disappear in an algorithm that used those paths to calculate semantic similarity, which would be a pitty. This weekend it finally all came together:
I already played around with drawing graphs with Gephi, as a method to check the results in a way other than processing huge lists in my python interpreter. But now I realize it could be the perfect ‘final product’. What I want to create is a ‘semantically-augmented tag-cloud-graph‘. As opposed to a standard tagcloud, my augmented tag graph will be a visualization of the text, composed of both interlinked concepts and solitary concepts, concepts that can be found literally in the text and concepts that aren’t mentioned anywhere in the text. The tag graph will benefit from the linked data nature of ontologies to:
1. Show relationships between concepts
2. Show extra concepts which do not occur literally in the text: nodes that occur in a path between two nodes, and maybe ‘superClass’ and ‘subClass’ nodes.
In this way, it will be similar to a tag cloud as it conveys the content of a text, but augmented as it could convey the meaning of a text and relationships between tags. It will also be similar to the DBPedia RelFinder, but augmented as it will show links but also contain a hierarchy of more to less important concepts. I hope that in the end it could be a viable alternative of graphic text-representation.
As I see it, my semantic graph-cloud should communicate at least these three things:
1. The most likely topic of the text
2. The concepts which occur literally in the text versus concepts that do not occur in the text
3. Clusters of similar concepts
My first idea is to model these three properties by size, alpha-channel and colour respectively, the bigger node, the more likely the topic, transparant nodes are the ones that do not occur in the text, and coloured nodes to group semantically similar nodes. But that’s just my initial plan, I might have to give it some more thoughts and experiment with it.
Next, I should think about a method to ‘measure’ whether my semantic tagcloud conveys more information, or has ajy added value. In the end, I am totally unsure whether the resulting tag graphs will make sense as it depends on multiple factors such as the efficiency of my keyword extraction, the efficiency of the vector-space based string comparison, the quality of the ontologies, etc.
The good news is, most of the technical work is in a close-to-finished state. I have various methods of extracting keywords from large texts (which I will be validating soon), I have a functional method to find ‘new’ concepts by comparing a text to all the descriptions of ontology concepts, I have a functional shortest-path finder to explore how two concepts are related (and produce a graph of it). It’s just a matter of putting it all together and selecting a suitable tool to draw the graphs. I don’t think I’ll use Gephi, as I want to fully integrate the graph drawing in my script, so who knows, maybe it’s back to NetworkX, igraph, or maybe Protovis?
In the mean time I’m looking at validating the data I generated with the various keyword extraction algorithms. By using human experts and cross-validating with existing keywords from certain entries, I’ll be able to evaluate which method of keyword extraction is the most efficient for my purposes. Once that’s done, and I’ve picked my graph-drawing method, I can start showing some preliminary results!
Wordle van de 200 meest voorkomende woorden in tweets met hashtag #OccupyAmsterdam. Gemaakt van 5.239 tweets van tussen zaterdag 8 oktober 09:55 uur en 16 oktober 15:50 uur.
Handmatig gefilterd op nicknames en nietszeggende woorden. Hier is de lijst van de 1000 meest voorkomende woorden: OccupyAmsterdam-woorden.
“The tf–idf weight (term frequency–inverse document frequency) is a weight often used in information retrieval and text mining. This weight is a statistical measure used to evaluate how important a word is to a document in a collection or corpus.”[wikipedia]
It is also the weight I use to measure similarity between texts, for these two tasks of my thesis project (click for pic!):
– Step 3: measure the similarity of a cyttron-db entry to a concept-description from an ontology. This will allow me to find concepts in the text that do not appear literally.
– Step 5: to be able to relate concepts which come from different ontologies. By measuring how similar the text surrounding a concept found in the text is compared to another found concept.
As mentioned before, I am using the excellent Gensim “vector space modelling for humans” package, which takes all the complicated mathematics off my hands (like the scary and intimidating formula up top!). Perfect for me, as I’m not mathematician, nor a computational linguist, nor a statistician, but I AM a human, who wants to work with a solid and proven method of similarity measures and feature extraction for texts. Since I am what I am, I won’t attempt to explain any of the inner workings of Bag-of-word models, vector space, and TF-IDF measures, sorry, there are much better places for that. I’ll simply show how I made Gensim work for me (assuming it does).
The first step is to create a training corpus. The training corpus defines the features of the text – the words that will be considered ‘important’ when looking at a text. The training corpus needs to be from the same domain as the target application: in my case the biomedical domain.
I wrote a simple script using lxml2 to parse the individual files: extracting all plaintext from the article body, cleaning them and storing them in a new text-file (1 article per line) for later processing. The cleaning process consists out of 3 steps: tokenizing articles (aka breaking an article up in words), filtering for common stopwords, and finally stemming the remaining words. I chose to include stemming, in order to unify such words as ‘hippocampal’ and ‘hippocampus’ (stemming returns the ‘root’ of a word). As I stem both the training corpus and the strings that need to be compared, it is not a disaster if words get stemmed incorrectly: in the end I don’t need to make sense out of the stemmed words, I only need them for counting. The plaintext file my script created is 650MB (vs 8,8GB for the uncompressed XML-files)!
The cleaning of the article is pretty straightforward, using pre-cooked NLTK modules: the WordPunct tokenizer, set of English stopwords and NLTK’s implementation of the Porter stemmer. For the quality of the similarity measurement it is important to follow the exact same cleaning procedure with the strings I want to compare – I use the same function for both the corpus-preparation as that of the comparison strings:
stopset = set(stopwords.words('english'))
stemmer = nltk.PorterStemmer()
tokens = WordPunctTokenizer().tokenize(doc)
clean = [token.lower() for token in tokens if token.lower() not in stopset and len(token) > 2]
final = [stemmer.stem(word) for word in clean]
Creating a training corpus in Gensim
Gensim‘s documentation is very extensive, and I can recommend going through the tutorials if you want to get an idea of the possibilities. But I couldn’t find much documentation on how to do simple string-to-string comparisons, so I wrote down what I did (and errrm yes, it’s pretty much exactly the same as string-to-index querying you can find in the Gensim tutorials :p):
1. Create a ‘dictionary’ of the training corpus’ raw text:
The dictionary contains words:frequency mappings and will be used to convert texts to vector space at a later stage:
>>> dictionary = corpora.Dictionary(line.lower().split() for line in open('corpus.txt'))
>>> print dictionary
Dictionary(1049403 unique tokens)
2. Convert the training corpus to vector space:
for line in open('corpus.txt'):
>>> corpus = MyCorpus()
>>> corpora.MmCorpus.serialize('corpus.mm', corpus) # Save corpus to disk
>>> corpus = corpora.MmCorpus('corpus.mm') # Load corpus
>>> print corpus
MmCorpus(99432 documents, 1049403 features, 39172124 non-zero entries)
Thankfully it’s possible to store the generated corpus, dictionary and tfidf to disk: parsing all these documents takes quite a while on my computer. That’s it for the preparation of the training corpus!
Comparing two strings
Now whenever I want to compare two strings, using features gathered from the training corpus, I need to:
Clean both strings in the same way I cleaned the articles in the corpus (NLTK stopword-filter + tokenization) » cleanDoc(string)
Convert both strings to vector-space using the dictionary generated from the training corpus » dictionary.doc2bow(string)
Convert both vector-space representations of the strings to TF-IDF space, using the TF-IDF model initialized earlier » tfidf[string]
When both strings are prepared, all is left to compare them, by creating an ‘index’ (the reference string) and a ‘query’ (the other string). Order doesn’t matter.
index = similarities.MatrixSimilarity([tfidf1],num_features=len(dictionary))
sim = index[tfidf2]
print str(round(sim*100,2))+'% similar'
Resulting in, for example, the comparison of the description of “Alzheimer’s disease” and “Cognitive disease” in the Human Disease (DOID) ontology:
>>> compareDoc("""A dementia that results in progressive memory loss, impaired thinking,
disorientation, and changes in personality and mood starting in late middle age and leads
in advanced cases to a profound decline in cognitive and physical functioning and is marked
histologically by the degeneration of brain neurons especially in the cerebral cortex and
by the presence of neurofibrillary tangles and plaques containing beta-amyloid. It is
characterized by memory lapses, confusion, emotional instability and progressive loss of
mental ability.""","""A disease of mental health that affects cognitive functions including
memory processing, perception and problem solving.""")
Or another example: the Wikipedia article of “Alzheimer’s disease” compared to the ontology description of “Alzheimer’s disease”:
alzheimer in wikiTxt
>>> compareDoc(wikiTxt,"""A dementia that results in progressive memory loss, impaired thinking,
disorientation, and changes in personality and mood starting in late middle age and leads in
advanced cases to a profound decline in cognitive and physical functioning and is marked
histologically by the degeneration of brain neurons especially in the cerebral cortex and by
the presence of neurofibrillary tangles and plaques containing beta-amyloid. It is characterized
by memory lapses, confusion, emotional instability and progressive loss of mental ability.""")
Final example: the top 5 most similar ontology concepts to the Wikipedia page of “Alzheimer’s disease”:
Now the second task (of matching a string to all the descriptions from my ontologies is much the same process, with the only difference that I need to use the similarities.Similarity object when creating the index (of the descriptions): the MatrixSimilarity object resides fully in RAM, the Similarity object on disk.
I am pretty confident about these preliminary results. It all seems to work as it should, and should be much more robust than my earlier attempts at similarity measurement using difflib and some crummy homegrown keyword-extraction and comparison (which I will still use for generating synonyms, crumminess works for that).
To my right is a visualization of the output of my SPARQL-powered shortest path algorithm, finding a link between ‘intracellular and extracellular accumulation’ & ‘developmental and adult structural defect’, 2 concepts in the Mouse Pathology ontology. Click it! It shows the two ‘source’ concepts in white, and the shortest path (of 3 nodes: 4 hops) in grey. It looks like this in python:
My algorithm generates a directed subgraph of all the nodes and edges it encounters during its search for a path between two ontology concepts. I figured generating this subgraph would make it easier to get some of the variables I need for the semantic similarity measurement calculation (such as amount changes in directions in the path, node’s depth from the root, etc.). Furthermore, I can use the subgraph to more easily assign weights to the textual data surrounding the nodes, when assembling my bag-of-words model of a node’s direct context, as I’ve explained in my previous post.
What to use
There are heaps of libraries for managing graphs in Python, and loads of programs to visualize and manipulate graphs. Here’s my stuff of choice.
After looking at several python-graph libraries ( NetworkX, igraph, graph-tool, etc.) I choose to use python-graph, which in my opinion is the most straightforward, compact, simple yet powerful graph-lib for python. I use python-graph to generate a file describing the subgraph in DOT language (“plain text graph description language”). This file can then be imported in a wide array of programs to visualize and manipulate the graph.
Visualizing the subgraph containing the shortest path between two nodes would allow me to get a better picture of what my algorithm fetches, and also to double-check the results (seeing is believing ;)). To visualize graphs, there are plenty of options again. After sifting through Tulip, Graphviz and some other obscure programs I stumbled upon Gephi, a very complete, pretty and simple open-source graph visualization & manipulation program. It has extensive documentation, and several advanced features to manipulate the graph and fetch some values. Ideally though, I will manage all those ‘advanced value-fetching tasks’ in my python script. Gephi still provides a nice and quick way to double-check some of the output and get a more concrete idea about what’s happening, as things can get pretty complex, pretty fast:
1136 nodes of the DOID-ontology, subgraph produced when finding a link for DOID_4 (disease) and DOID_10652 (Alzheimer's Disease) - red nodes in the graph. Linked by the three yellow nodes.
As explained in an earlier post, I am working on a simple method of extracting ‘important words’ from a text-entry. The methods I am using at the moment are frequency distributions and word collocations. I’ve bumped into some issues regarding finetuning my methods. Read on for a short explanation of my approaches, and some issues regarding them.
Frequency Distribution: POS-tagging y/n?
Extracting keywords by frequency distribution is nothing more than counting words and sorting the list of words by occurrence. Before doing this, I filter stopwords from the text entry. The short explanation on how I’m doing this (sourcecode available at github):
» Tokenize the text (using NLTK’s WordPunctTokenizer)
» Lowercase all the words
» ‘Clean’ the list by removing common stopwords from the list (using NLTK’s English stopwords-list)
This is straightforward enough, an example of the results (from the WikiPedia page of ‘Apoptosis‘):
Earlier I was thinking about using POS-tagging (Part-Of-Speech tagging to identify word-types) in order to only extract frequently occurring nouns. I figured losing relevant adjectives (such as ‘red’ in red blood cell) could be compensated by the word collocations extraction. POS-tagging the tokenized text, and retrieving only the most frequent nouns results in:
My problem here is I’m not sure which is ‘better’ (if any of those two), or if I should maybe use a combination of both. Also, I haven’t decided yet how to handle non-alphabetic words. Initially I planned on using regular expressions to filter non-alphabetic strings, but I figured later that it wouldn’t make sense in my case. In the above example, this would omit retrieving ‘p53’: a tumor suppressor protein (P53), which is very relevant.
With earlier playing around with POS-tagging I noticed the precision was not quite high enough to perform chunk extractions (by looking for specific phrases / grammatical constructions). Extracting only nouns does seem to do quite the job, even if I still miss some and get some false positives.
Word Collocations: Stopword filtering y/n?
Collocation defines a sequence of words or terms that co-occur more often than would be expected by chance. I generate bi- and trigram word collocations, which mean ‘2-word strings’ and ‘3-word strings’. My issue here is whether or not to use stopword filtering. Here are the results of the word collocation function on the same WikiPedia page, the 1st list being the bigram collocations, the 2nd being the trigrams. Example without stopword filtering:
As you can see, lots of garbage in the first example, but still some collocations that do not appear in the cleaned version. Similar to the noun-extraction issue with the previous approach, I wonder if I should choose for one of the two, or combine them.
In other news
Finding Gensim has been a life-saver! Instead of using Difflib to compare two strings, I now use a proper text-similarity metric, namely cosine similarity measurement. I do so by creating a TF-IDF weighted corpus out of the (stopwords-cleaned) descriptions of ontology-terms I use, and calculating the cosine similarity between an input string and each entry in the corpus. Gensim makes this all a breeze to do. An example of the ouput:
alzheimer in wikiTxt
Label: Alzheimer's disease
Description: A dementia that results in progressive memory loss, impaired thinking, disorientation, and changes in personality and mood starting in late middle age and leads in advanced cases to a profound decline in cognitive and physical functioning and is marked histologically by the degeneration of brain neurons especially in the cerebral cortex and by the presence of neurofibrillary tangles and plaques containing beta-amyloid. It is characterized by memory lapses, confusion, emotional instability and progressive loss of mental ability.
Label: vascular dementia
Description: A dementia that involves impairments in cognitive function caused by problems in blood vessels that feed the brain.
Description: A cognitive disease resulting from a loss of brain function affecting memory, thinking, language, judgement and behavior.
Label: cognitive disease
Description: A disease of mental health that affects cognitive functions including memory processing, perception and problem solving.
Description: Encephalitis is a nervous system infectious disease characterized as an acute inflammation of the brain. The usual cause is a viral infection, but bacteria can also cause it. Cases can range from mild to severe. For mild cases, you could have flu-like symptoms. Serious cases can cause severe headache, sudden fever, drowsiness, vomiting, confusion and seizures.
I’m not sure if the similarity numbers it produces indicate I’m doing something wrong (there’s no high similarity), but intuitively I would say the results do make sense.
My thesis is about keyword extraction of biological notes, using semantic ‘dictionaries’ called ontologies. These ontologies are large networks, where each node stands for a concept, and each connection between nodes for relations. See the picture on the right for a visual representation of an ontology.
To identify the subject of a text, I need to see what terms that are described in an ontology appear in a text. This leaves me with multiple concepts, of which I need to find the ‘common denominator’. To do this, I have to measure the similarity (or inversely: the distance) between two concepts: if I find a bunch of very similar concepts in one text, I can be more confident of the subject.
Luckily, a lot of people have dealt with this ‘ontology-based semantic similarity measurement’. I gathered and studied a couple of papers, and provide a quick overview of my findings. See my literature list for a more complete overview.
DISCLAIMER: This is by no means intended to be an exhaustive overview. It’s short. I’m sure I’ve not read every relevant paper. Due to time constraints my priority is finding a suitable method to carry on and checking to see if the direction I’m heading is OK (it seems that way). My overview deals with global approaches only, nothing too specific. If I’ve missed anything really obvious, I’d be grateful if you could leave a comment :).
There are two main approaches in ontology-based semantic similarity measurement: edge-based (also called structural or hierarchical approach) and node-based (also called information-content approach).
Edge-based approaches take the structure of the network as a base, focussing on the connections between nodes and their implications/meanings. In edge-based semantic similarity measurement, there are three main principles (which are fortunately pretty much globally agreed-upon – at least in the papers I found):
Shortest-path length between nodes
The most direct approach: the closer two nodes are in the network, the more similar they are. Important detail: path-length is measured by counting (only!) the nodes which have a ‘is_a’ or ‘part_of’ relation. The most primitive semantic similarity measures use only path lengths. However, this shortest-path measure can be extended with:
Node ‘depth’ (aka specifity)
The deeper a node is (farther away from the root), the more specific it is. In most papers this does not revolve around individual node’s depths, but around the depth of their Least Common Subsumer (LCS). The LCS is the deepest ‘shared parent’ of two nodes. The depth is defined as the amount of nodes the LCS is separated from the root concept. The deeper the LCS is, the more similar the concepts are. Also, the granularity of an individual concept has to be considered in calculating its specifity (more granular means more ‘subdivisions’, means a more specific concept). This is usually modeled as an extra variable that influences the concept’s specifity. This means that a highly granular node will be less similar to a less granular node.
Link’s direction Ontologies are directed graphs: a connection between two nodes has a direction (chair is_a furniture, does not work the other way around). The more changes in direction the path has between two nodes, the less similar the nodes are.
Node-based measures do not take the connections in the network as a main resource, but rather the information inside and surrounding the nodes. Text-mining and textual analysis techniques can be applied here. For example by comparing both concepts’ textual data, the concept’s contexts, or by comparing the similarity of the concepts’ LCS to the individual concepts. In these cases, a node or node’s context is frequently represented as a ‘bag of words‘, disregarding any form of grammar or semantics. Cosine similarity is a measure which is often used for textually comparing two texts. A common approach to weigh the importance of words in a text is the TF-IDF (term frequency–inverse document frequency) measure.
Other techniques involve counting the amount of surrounding nodes (in a way similar to checking a node’s granularity), the depth of a node (counting the amount of hops from the root-node), etc. The way I see it, a node-based approach is a useful extension on an edge-based approach.
Both these approaches (edge & node-based) are applied in a multitude of algorithms combining edge & node-based, or dealing with either one of the two. For an overview of some common algorithms and their use of edge- and/or node-based approaches, I highly recommend .
What’s left for me is to formulate an approach: picking an edge-based similarity measurement algorithm and implementing it, and finding a node-based approach to extend the edge-based approach with.
For the edge-based algorithm, the bare essentials are already in place:
A breadth-first search algorithm to determine paths between nodes
A method of finding the common parent (LCS) of two nodes
A method of counting the depth of a node
A method of exploring the context of a node
Then, I am looking to extend the structure-based approach by including a similarity-comparison of the linguistic data ‘surrounding’ each node. By retrieving all surrounding nodes of both nodes that I’m comparing, throwing all textual data of the surrounding nodes in a ‘bag of words’, and comparing this bag of words to the second node’s bag of words. This is also called a node-based similarity measure (as opposed to the previously described ‘edge-based’ measure). I will also look into combining this text-comparison system with keyword extraction.
An extremely useful Python framework for all text-comparison tasks at hand (BOW model, cosine similarity measures, TF-IDF) I came across is Gensim. It has all the features I could want to use, plus excellent documentation.
“Gensim is a Python framework designed to automatically extract semantic topics from documents, as naturally and painlessly as possible.”
Pesquita C, Faria D, Falcão AO, Lord P, Couto FM (2009) Semantic Similarity in Biomedical Ontologies. PLoS Comput Biol 5(7): e1000443. doi:10.1371/journal.pcbi.1000443
Al-Mubaid H, Nguyen HA. “A cluster-based approach for semantic similarity in the biomedical domain”, Conf Proc IEEE Eng Med Biol Soc. 2006;1:2713-7.
A. Bramantoro, S. Krishnaswamy, and M. Indrawan, “A Semantic Distance Measure for Matching Web Services”, in Proceedings of WISE Workshops 2005. pp.217~226
E. Gabrilovich and S. Markovitch. Computing semantic relatedness using wikipedia-based explicit semantic analysis. In Proc. IJCAI, 2007
M. Andrea Rodríguez, Max J. Egenhofer, “Determining Semantic Similarity among Entity Classes from Different Ontologies,” IEEE Transactions on Knowledge and Data Engineering, pp. 442-456, March/April, 2003
Lee, W N et al. “Comparison of ontology-based semantic-similarity measures.” AMIA Annu Symp Proc 2008 (2008) : 384-388.
I. Spasic, S. Ananiadou, J. McNaught, and A. Kumar, “Text mining and ontologies in biomedicine: Making sense of raw text”, presented at Briefings in Bioinformatics, 2005, pp.239-251.
P. Resnik. “Using information content to evaluate semantic similarity in a taxonomy”, In Proceedings of the 14th international joint conference on Artificial intelligence – Volume 1 (IJCAI’95), pp. 448-453. 1995.
R. Thiagarajan, G. Manjunath, and M. Stumptner. “Computing Semantic Similarity Using Ontologies”, HP Labs, 2008