Keyword extraction

Combine analytical & synthetic skills.

1h20 per week, for 3 weeks

Representing documents as graphs

Graphs have been successfully used in Information Retrieval to encompass relations between entities (e.g. PageRank [Page et al., 1999]). Graph of Word (Rousseau et Vazirgiannis, 2015) is an approach for text mining which adresses the term independence assumption through the bag-of-word representation. It takes into account word dependencies and word order through a graph based representation of a document (graph-of-word).

Unweighted directed graph

The weight of a word in the document can be set to the number of neighbors in the graph -> favor words that occur with many different other words.

Pros/Cons: Robust to varying document length, weight of a word increases only with new context of co-occurrences.

Directed graph

The weight of a word in the document can be set to the indegree of each node (numbers of incoming edges) in the graph-of-word. It represents the number of distinct contexts of occurrence. For example: information retrieval is the activity of obtaining information resources relevant to an information need from a collection of information resources, gives the following nodes and features/weights:

information (5), retrieval (1), activity (2), obtaining (2), resources (3), relevant (2), need (2), collection (2).

Keywords extraction and document summarization

Keywords are used for:

• looking up information on the Web (e. g., via a search engine bar)
• finding similar posts on a blog (e. g., tag cloud)
• for research paper indexing and retrieval (e. g., SpringerLink)

Applications are numerous:

• summarization (to get a gist of the content of a document)
• information filtering (to select specific documents of interest)
• indexing (to answer keyword-based queries)
• query expansion (using additional keywords from top results)

Graph-based Keyword Extraction selects the most cohesive sets of words in the graph as keywords. K-core decomposition retains the main core of the graph (weighted edges), and nodes based on their centrality and cohesiveness.

Evaluated against PageRank on scientific papers with keywords manually assigned by human annotators, in terms of precision, recall, F1-score, # of keywords, the algorithm is relatively simple and works well.

Note: Other graph algorithms, for example Minimum Spanning Trees or Min Vertex Cover can help visualise, summarize documents for various applications.

Quiz

Reference

F Rousseau, M Vazirgiannis. Main core retention on graph-of-words for single-document keyword extraction. European Conference on Information Retrieval. 2015.

F Rousseau, M Vazirgiannis. Graph-of-word and TW-IDF: new approach to ad hoc IRF. Proceedings of the 22nd ACM CIKM. 2013.