Information Retrieval Models

Ranks passages against queries using various models and techniques.

Structure

The following describes the purpose of each package and the contained modules.

Root

• start.py - Initialises the program, taking parameters from the user for the dataset and choice of model.
• zipfs.py - Generates a distribution plot and reports Zipfs law parameters for the dataset. Note that this runs separately from the main program.

Retrieval

• data/ - Data structures for storing and managing data.
• models/ - Information retrieval models for ranking passages against queries.
• util/ - Helper functions used for processing and managing data.
• DatasetParser.py - Takes a dataset as input and parses it using a specified model.

Data

• Dataset.py - Extracts and retrieves data from the dataset.
• InvertedIndex.py - Indexes passages from the dataset.

Models

• Model.py - Base class which all IR models inherit.
• BM25.py - BM25 probabilistic retrieval model for estimating the relevance of a passage.
• VectorSpace.py - Vector space algebraic model for representing passages as vectors.
• QueryLikelihood.py - Query likelihood language model for calculating the likelihood of a document being relevant to a given query.

Util

• FileManager.py - Reads and writes to a given file.
• TextProcessor.py - Performs text preprocessing on a collection or passage.

How to Run

The program can be initialised by running start.py, which accepts parameters in the format of:
start.py <dataset> <model> [-s <smoothing>]
To generate a distribution plot and report parameters for Zipfs law, run zipfs.py, which accepts a parameter in the form of:
zipfs.py <dataset>

Parameters

Dataset

• The <dataset> parameter is required and is the path of the dataset to be parsed.
• Expects a TSV file in the format , where qid is the query ID, pid is the ID of the passage retrieved, query is the query text, and passage is the passage text,
• Each column must be tab separated.

Model

• The <model> parameter is required and is the name of the model to be used for ranking passages.
• Expects either 'bm25' for the BM25 model, 'vs' for the Vector Space model, or 'lm' for the query likelihood model.
• Any other input will be deemed invalid, and an exception will be raised.

Smoothing

• The -s <smoothing> parameter is required only when using the Query Likelihood model, and is the name of the smoothing technique which will be applied.
• Expects either 'laplace' for Laplace smoothing, 'lidstone' for Lidstone smoothing, or 'dirichlet' for Dirichlet smoothing.
• This parameter can only ever be used if the Query Likelihood model was selected for the <model> parameter, and an exception will be raised if any other model is used.

Examples

BM25 Model

start.py candidate_passages_top1000.tsv bm25

Vector Space Model

start.py candidate_passages_top1000.tsv vs

Query Likehood Model

start.py candidate_passages_top1000.tsv lm -s laplace
start.py candidate_passages_top1000.tsv lm -s lidstone
start.py candidate_passages_top1000.tsv lm -s dirichlet

Zipfs Law Report

zipfs.py candidate_passages_top1000.tsv

Dependencies

• numpy
• matplotlib
• nltk
• num2words
• tabulate
• punkt (nltk module)
• stopwords (nltk module)
  NLTK modules are downloaded automatically at runtime