The process of
inspection, transformation and shaping up of data, keeping the discovery of
useful information and conclusions as goal is known as Data Analysis. This
supports decision making. Information retrieval is carried out by various IR
methods and data is further analyzed.
evaluation of relevance with the help of some document representations with
respect to the query is done by an IR system. There are various models for
representation documents and queries. Thus, each model has its pros and cons.
has multiple facets and approaches, encompassing diverse techniques under a
variety of names, in different business, science, and social science domains. Data
analysis is closely associated to the visualization and dissemination of data.
The term data analysis is often referred to as data modelling.
retrieval refers to the task of extracting out relevant information resources applicable
to an information requirement from a set of information resources collected. Usually,
metadata or on full-text (or other content) based indexing searches can be
Hidden Markov models have been successfully
designed and implemented, over the period of last two decades, covering a wide
variety of speech and language related recognition problems which include
speech recognition, named entity ending, optical character recognition, and
topic identification and a lot more 1. In the present work, an application of
this technology is described by us with respect to the ad hoc IR technique 2.
In every HMM implementation, the observed data is modelled by the output
produced by passing any unknown key through certain noisy channel(s). In the case
of ad hoc IR proposition(s), we represent the observed data as the query, and
an unknown key that makes up a desired relevant document. Thus, for each
document we can compute the probability that it highly probable that this was
the relevant document as imagined by the user, given the query. We then rank
the documents based on this measure.
mining is a distinct technique for data analysis which does not concentrate
upon purely descriptive purposes, rather, focuses on modelling and discovery of
knowledge for predictive purposes. Data relying excessively on aggregation and
aiming in business information comes under business intelligence. Customer data
and IT tools build the substructure on which a victorious CRM strategy is created.
Also, the quick expansion of the web and related technologies has substantially
extended the number of marketing opportunities. In addition, this has altered
the way alliance between
and their clients are balanced and supervised 3.
aims at application of statistical models for estimating or categorization, while statistical,
language-producing, and systemic techniques are applied to text analysis to
acquire and classify information from textual resources.
from the web incorporates handling the abstractness and volume of data
contained on the internet. When including aspects like as word ambiguity and a
large number of typographical errors, it is made increasingly difficult. There
exist a variety of key pitfalls comprehending IR- relevance, evaluation, and
However, this is
not the complete set of issues involving IR. Common information retrieval
problems include potential, scalability and paging update occurrences. The relational
value of an input provided by the user in the form of query, within a dataset,
is known as relevance, which is calculated using a ranking algorithm.
The larger complications
with IR that are evaluation and relevance are still significant subject matters
that require attention, amongst others.
and the respective queries form a corpus of terms where every term within that
document is indexed. 1 and 0 denote the presence and absence of some text in a
text source respectively 4,5.This is the Boolean model. Maintenance of an
inverted index of every term is necessary in order to process matching of
document and query. Nonetheless, this model holds certain limitations as
explained further. Binary decision criterion has a disadvantage that it exists lacking
any grading scale concept. Another problem includes overloading of documents.
Certain researchers have worked upon this to control the fragility of the above
said model by improvising the existing one. Certain researches have also approached
data analysis with a different search strategy of vectors. This is known as the
Vector Space model 5.
This Model denotes
documents and queries as vectors. In this model, every query and document is
expressed as vectors that exist in a |V|-dimensional space. Here V is the
collection of all distinct terms in the set of documents. Here, the documents
set is the vocabulary 5.
were first proposed by Russian Mathematician Andrei Markov. A Markov model in
probability theory is a stochastic model. This model is used to model systems
that change randomly. In this model, it is presumed that the future states depend
only on the present ones rather than the sequence of events that occurred prior
to it 1, 2, 6.
There exist four
Markov models that are used in different situations, depending on the
observational degree of every sequential state.
III. MODEL TO BE IMPLEMENTED
A hidden Markov
model (HMM) constitutes of a Markov model that is statistical. Here, the system
to be modelled is presumed to be a Markovian process with states that are hidden,
which implies that the states are unobserved. The simplest dynamic Bayesian
network can refer to as HMM 7.
the measurement of effectiveness of spontaneous information retrieval in the
standard way, we require a collection of tests consisting of three things:
? Collection of documents
? Test suite of requirements represented
as a set of queries.
? Set of conclusions, which standardly
is a binary assessment of relevance computed as either relevant or irrelevant
for every text-query pair.
following parameters were in use for the evaluation of performance of IR systems:
Precision: It is the fraction of documents relevant among the completly
retrieved document. Practically it gives accuracy of the judgement.
Recall: The fraction of the documents retrieved and relevant among all relevant
documents is referred to as recall. Practically it gives coverage of result.
Set of relevant documents retrieved
Set of all relevant documents
recognition system and IR with binary classification, precision refers to the
fraction of instances retrieved that are found to be relevant, while recall
refers to the fraction of relevant instances that are extracted and retrieved.
Both precision and recall are henceforth derived from an understanding and
degree of relevance.
A. Language used- Python
For both small
and large scale, Python helps enabling clear programs by providing constructs.
Its features include a dynamic type system and an automatic memory management.
It also has a huge and all-inclusive standard library.
standard library provides tools to users that are suited for numerous tasks.
Modules for creating GUIs, connecting them to relational databases, pseudorandom
number generators, and arithmetic decimals with arbitrary precision,
manipulation of regular expressions are included. It is also capable of performing
B. Dataset used
OHSUMED test collection is a combinational set of 348,566 references from
MEDLINE. It is the on-line database for medical information present on World
Wide Web. It has a title, MeSH indexing terms, author, and an abstract with
source as available fields in the database.
OHSUMED topics define the real requirements. Although, the judgements of
relevance does not have the same coverage as given by the pooling process of
TREC. The information requirements aren’t directly expressed by MeSH but these
terms manage the indexing terms. The standard TREC format provides the topic
statements and includes only
relevant document files are described below which simulate human judgement and
contain information for 0 or 1 for every MeSH term expressed in the filtration
of any given topic.
relevance judgments (files: qrels.ohsu.*)
searchers replicate each query. Out of these four, two are physicians who are experienced
in searching and the other two are medical librarians. A completely different
set of physicians estimate the results for relevance. This judgement is
performed on a three point scale. The pointers are: definitely, possibly, or
not relevant. Consideration for relevance is done for all documents that are
checked to be either definitely relevant or possibly relevant.
relevance judgments (files: qrels.mesh.*)
The document is considered to be relevant to a
MeSH topic if its concept is included in the list of MeSH term fields.
C. WHOOSH: Python Library
created by Matt Chaput.
? Whoosh uses only pure python hence
runs anywhere python can, and so is fast. It runs without requiring a compiler.
? Whoosh uses the Okapi BM25F ranking
function by default, but can be easily modified.
? Fairly small indexes are created by
Whoosh as compared to numerous other search libraries.
? All indexed text in Whoosh must be
permits you index free structured text for quickly searching matching documents
with respect to either simple or complex search guidelines.
predefined field types are provided by whoosh:
is used for indexing the text and storing locations for the terms. These
positions or locations further allow phrase searching.
entire value of the field is indexed into a single unit using the ID field,
rather than breaking it up into separate terms.
is neither an indexed type nor a searchable one. This is useful for displaying
the information to the user in the search results.
indexed and searchable type, this is created for comma and space separated
is capable of storing int, long, or floating point numbers in a format that is
sortable and compact
of boolean values is done by this field and this type allows users to search
for results like: true, false, 1, 0, t, f, yes, no.
objects are stored in this field in a compact and extremely sortable format.
A Format object is
made to define the type of information is recorded by a field about each term.
It also describes how it has to be stored on the disk. For example, this is how
the postings are stored by the Existence format:
While on the
other hand, this is how the Positions format would do the same:
string is passed to the field’s format object for a field by the indexing code.
An analyser is called by the format object which breaks the string into tokens.
Further, encoding of the information is done about each of them.
index performs mapping of the terms to the documents in which they appear. Also,
sometimes it is useful to store a term vector that maps all the terms that arise
in the documents to the original document sources.
inverted index of a field is:For the image above, the respective
forward index is:
D. Creating An Index
opening an existing index in a directory,
Open the index directory
import whoosh index index
creating an index in a directory,
Create a new index
Import os and os path
if os path doesn’t exist, make index directory
and create an index with schema as parameter.
schema using which the index is created is stored with the index itself. Indexes
can be kept in the same directory using the index-name keyword.
use the functions for convinience
Create index with schema and index name usage as parameters.
Open this index
use Storage object
Call storage.create with schema and index name usage as parameters
Open the storage
The relevance of
the documents using Hidden Markov Model is compared with the tf.idf approach.
Tf.idf is an approach based on numerical statistic based vector model. It reflects
necessity of a word to a document in a collection of documents. Often, it is used
in IR and data mining as a weighting factor.
The tf-idf value
is proportional to the frequency of appearance of a word given in the document.
Although, it is offset by the frequency of the word in the collection. This
helps to relate to the fact that in general some words have more frequency of
appearance than others.
implementation, the first step is to design the schema and then indexing is
performed 5. Then tf.idf values are calculated using Whoosh Library in Python.
For HMM calculation the data observed is assumed to be the query Q, and an
unknown key is assumed to be a relevant document D that is desired. The mind of
the user is a noisy channel, who is having either some precise or rough notion of
the documents he requires. This channel transforms that expressed notion into
the query text Q. Hence, we compute the probability for each document D that it
was the relevant one in the user mind, provided that Q was the query which was
expressed or produced, i.e. P (D is RjQ). We further rank the documents with
respect to this measure 6. This can be incorporated in the form of graphs. These graphical structures represent information
about a domain that is uncertain. Particularly, nodes denote random variables
with the edges denoting the probabilistic dependencies transitioning between all
the random variables 8.
is the term represents that an observer cannot realise the transition of states
and the underlying sequences by which the output is generated. But he view the output
states only 9.
P (qjD) is the
output distribution of any document D. It is set to be the sample distribution
for the words that appear in that document. For any document Dk, we can
has the maximum probability of producing Dk by repeatedly sampling the state
“General English”. It is estimated by
here is taken for all documents present in the collection. Using the parameters
estimated above, the formula for P (QjDk is R) is stated as under:
1. Hidden Markov
model is a formal substructure used for creating probabilistic models for problems
of linear sequence ‘labelling’. Just by drawing an intuitive image, a
conceptual toolkit is provided. This is very useful for building complex models.
They are at the hub of a set of miscellaneous programs. These programs include
gene finding, multiple alignments of sequence, profile searches and identification
of regulatory site.
2. HMM is a complete
probabilistic model. The overall ‘scores’ generated for sequences and the
parameters calculated are all probabilities 6, 9. Hence, Bayesian probability
theory can be incorporated for the manipulation of these numbers in more
powerful ways. This includes optimization of parameters and interpretation of
the significance of scores 5.
3. HMMs can be proved
useful for modelling of processes which contain different stages that occur in
definite orders 9.
If, for example,
you want to model the behaviour of a technical system that first boots, then
operates, then enters sleep mode, and iteratively changes between sleep and
operation later on, you might need three states (boot, operate, sleep) and can
use this process model to find out what’s going on in the system at any one
time. Similar is the case with a human biological system where the observations
can be the sequence of symptoms of a human being. Human genome project also
requires the assimilation of HMM for DNA sequencing and RNA structuring 10.
like scalability and frequencies of paging update are familiar IR issues. Ranking
algorithms are implemented with the usage of methods that elucidate
relationships amongst the given query and the accumulated documents. All the
feedback provided by the IR system has to be evaluated, which is another issue
with IR. The way the system behaves, may or may not converge with the
suppositions of the user. All the documents that are extracted from the
procedure may not be able to give relevance to a given query.
The way a user
interacts with the IR system is termed as Information needs. Retrieval of a lot
of information might be disruptive in a number of systems. On the other hand,
in another number of systems, not returning a complete set of relevant data may
handling a set of voluminous information from the internet might be extremely difficult
because of the extremely large size of documents the server manages.
A thousand of
documents can be returned by a simple retrieval query. Many of those documents
are loosely related to the original criteria of retrieval. To deal with this,
an IR system is required to have a query management that is efficient enough as
well as contains a good level of ability in order to give weight as priority to
documents that are closer for relevance to the user query.