Filetype PDF | Posted on 30 Jan 2023 | 3 years ago
Authentication
digital resources
245x Filetype PDF File size 0.22 MB Source: www.temjournal.com
TEM Journal 5(2) 236–240
Pattern Recognition and Natural Language
Processing: State of the Art
1,2 2 2 2
Mirjana Kocaleva , Done Stojanov , Igor Stojanovik , Zoran Zdravev
1E-learning Center – University “Goce Delcev”, Krste Misirkov bb, Shtip, R.Macedonia
2
Faculty of Computer Science – University “Goce Delcev”, Krste Misirkov bb, Shtip, R.Macedonia
Abstract –Development of information technologies is linguistics, statistics and computational intelligence.
growing steadily. With the latest software technologies AI is an interdisciplinary branch of computer science
development and application of the methods of that has connections to other sciences such as
artificial intelligence and machine learning intelligence neuroscience, philosophy, linguistics and
embededs in computers, the expectations are that in psychology. Despite its application in industry,
near future computers will be able to solve problems nowadays-predictive methods in AI are also
themselves like people do. Artificial intelligence commonly used in social sciences, such as
emulates human behavior on computers. Rather than economics.
executing instructions one by one, as theyare
programmed, machine learning employs prior There are several areas of specialization of
experience/data that is used in the process of system’s
training. In this state of the art paper, common artificial intelligence, such as:
methods in AI, such as machine learning, pattern
recognition and the natural language processing (NLP) • games playing, i.e. computers are
are discussed. Also are given standard architecture of programmed to oppose gamers
NLP processing system and the level thatisneeded for • Expert systems: computers are programmed
understanding NLP. Lastly the statistical NLP to make decisions about situations in real-life
processing and multi-word expressions are described. (Mycin is a typical expert AI system that was
developed in the 1970’s and it has been used
Keywords –artificial intelligence, machine learning, for bacteria identification and
pattern recognition, natural language processing.
recommendation of medications and drugs
based on known symptoms).
1. Introduction • Natural language: computers are
programmed to process sentences from
Depending of the scope of application, there are spoken languages, analysing the
many definitions for the artificial intelligence. morphology, lexicography and even the
According to [2], [4] artificial intelligence maps semantics of a whole sentence.
human behaviour on computers. Regardless whether • Neural networks: combination of artificial
the human behaviour is emulated or not, the goal of neurons designed upon the neuron of a
AI is to create intelligence. Without any doubt, the human being, primarily used for recognition
yet to come challenge in AI is to emulate completely purpose.
or near-perfectly general intelligence. For different • Robotics: computers are programmed to
purposes, AI combines different methods from receive surrounding signals and to generate
intelligent reactions upon them.
DOI: 10.18421/TEM52-18 Machine learning is a branch of AI that analyses
https://dx.doi.org/10.18421/TEM52-18 systems, which are able to learn from training data,
rather than following a strict order of execution of
Corresponding author: Mirjana Kocaleva - E-learning already programmed instructions. Machine learning
Center – University “Goce Delcev”, Krste Misirkov bb,
Shtip, R. Macedonia tends to construct self-adjustable artificial systems,
Email: mirjana.kocaleva@ugd.edu.mk which can grow up alone and change through time by
© 2016 Mirjana Kocaleva et al, published by gathering experience from new training data sets [9].
UIKTEN. This work is licensed under the Creative Alike data mining, the machine learning also
Commons Attribution-NonCommercial-NoDerivs 3.0 processes raw data in order to find patterns, but
License. instead of extracting data which will undergo human
The article is published with Open Access at analysis, it uses data to improve system’s own
www.temjournal.com
236 TEM Journal – Volume 5 / Number 2 / 2016.
TEM Journal 5(2) 236–240
capabilities. Detected patterns are used for system’s knowledge discovery. Unlike machine learning
self-adjustment [4], [5], [6]. which is focused to maximize the rate of recognition
There are two different kinds of learning. Learning [15], [17], [18], [19], pattern recognition models
without supervision and learning with adequate patterns and regularities found in data.
supervision. Learning without any kind of For our research, pattern recognition is important
supervision requires an ability to identify patterns in as a field in machine learning. Supervised learning
streams of inputs, whereas learning with supervision employs training data set, which is used to identify
involves classification and numerical regressions. patterns that match or resemble already annotated
The category an object belongs to is determined with regularities. Unlike supervised learning,
classification and regression deals with obtaining a unsupervised learning does not rely on training data
set of already recorded input/output samples, thereby and it can be applied to detect unfamiliar regularities
from respective inputs we are discovering functions in data.
enabling the generation of suitable outputs. By analyzing training samples, supervised learning
Computational learning theory is a branch of methods always produce an inferred
modern computer science that deals with a function.Applying these functions, an output for any
mathematical analysis of the performance of machine valid input object can be easily predicted. There are
learning algorithms.Essentially, machine learning can two kinds of inferred functions: if the output is
be defined as a self-learning data method for discrete the function is called a classifierand if the
improving computer’s actions or behaviour. The output is continuous the function is called regression
training data set strictly depends of the domain of the function.Data clustering (k-nearest neighbour’s
problem under consideration. algorithm, support vector machine, naive Bayes
classifier) and ANNs-artificial neural networks are
common approaches to supervised learning.
According [2], unsupervised learning does not rely
on previously labelled data and it attempts to detect
built-in patterns in data. It can be used to calculate
the correct output for any new data instance. Data
clustering (k-means clustering and hierarchical
clustering), hidden Markov models and blind signal
separation using features extraction techniques for
dimensionality reduction are common methods in
unsupervised learning.
Figure 1.Retrieved from In pattern recognition, there are problems where
http://whiteswami.wordpress.com/machine-learning/ distinct representations can be obtained for the same
pattern, and depending on the type of classifier
Robot can learn to walk based on reading from (statistical or structural), one type of representation is
sensors of force correction of the output for a specific preferred versus the others.By taking into
input. Pattern recognition or the process of teaching a consideration the statistical variance of all inputs,
program or system to be able to recognize patterns algorithms for pattern recognition aim to perform
[5], [6] is another way of thinking about machine "most probably" matching.
learning.
3. Natural language processing
2. Pattern recognition Natural language processing (NLP) or
Patterns are a form of language.Pattern recognition computational linguistics is a major field of
is studied in many fields, including psychology, computer-related research. By discovering language
psychiatry, ethnology, cognitive sciences, and patterns, children learn how to make a difference
computer science andtraffic flow.Pattern recognition between singular and plural, match templates in
is a field in machine learning, but may also refer to nouns, verbs and adjectives and how to form
pattern recognition (psychology), identification of different types of sentences, such as declarative or an
faces, objects, words, melodies, etc. [8], [10].Since imperative sentence. If we can define,and moreover,
the scopes of machine learning, knowledge describe patterns from natural language then we can
discovery, pattern recognition and data mining highly teach the computer about the way we speak and
overlap, they are hard to separate. Most often, understand sentences form the spoken language.
machine learning refers to methods based on According [9], [13], [14] much of the research in this
supervised learning, while unsupervised learning is field relies on methods from cognitive science and
primarily explored by data mining and KDD -
TEM Journal – Volume 5 / Number 2 / 2016. 237
TEM Journal 5(2) 236–240
linguistics. Natural language processing techniques We use six levels about understanding standard
train computers to understand what a human speaks. architecture of NLP system with aim to find out the
Natural language processing gives machines the meaning of Natural Language Processing. Not every
ability to read and understand the languages that level is used by every system for NLP. Those levels
humans speak. NLP research aims to answer the are [1], [3]:
question of how people are able to comprehend the
meaning of a spoken/written sentence and how Phonetic level
people understand what happened, when and where
that happened or what is an assumption, belief or Way in which words are produced, transmittedand
fact. understoodin language is also known as phonetics
[1]. This level is of great importance for
understanding spoken language, but it is not
important for the written text [3].
Morphological level
A morphemein a language is the smallest unit of a
word to carry meaning [3].According to Jurafsky
[13] there are stems and affixes as types of
morphemes.The stem is the primary part of a word
and it gives the meaning (for example happy),
Figure 2.Retrieved otherwise the part of the word, which adds further
fromhttp://tex.stackexchange.com/questions/184099/logo- significance of the word, is known as affix.Affixes
for-natural-language-processing are usually suffixes (for example in happ-ily) and
The common elements of any standard architecture prefixes (for example in un-happy).
of system for NLP processing are:
Speech recognition: Turning of a spoken word into Syntactic level
array of words. Spoken words are composed of a
series of parameters related to the sense of hearing. The study of sentences is known as syntax. Syntax
Language understanding: The goal of this element analysis includes the action of dividing a sentence
is to generate a meaning for spoken words, and that into components of which it is created. The main
meaning will be used by the next element (dialogue interests of the people are regulations and conditions,
management). which are defined by grammar and designed to keep
Dialogue management: Main task of this element a sentence along.The position of the word in the
is to coordinate and hold together all parts of the sentence can be established if the single word is an
system and users, andconnecting with other systems. object or subject in that sentence.NLP systems store
Communication with external system: such as a representation of every sentences andthey store the
expert system, system for databases, or other fact that a word is a verb and what kind of verb it is
computer application. [1], [3], [11].
Response generation: Setting out a message that
system should deliver. Semantic level
Speech output: Use of different techniques to
produce the message from the system. [11] Semantics concerned with the manner syntactic
structures are constructed.Syntactic and semantic
levels are inseparable and complement each other.
Syntax analysis deals with the structure of sentences,
while semantic analysis is searching for the meaning
inthose sentences [1], [3]. According to semantics,
most words have multiple meanings.Hence, we can
identify the appropriate word by looking at the rest of
the sentence or dependent on context [3].
Discourse level
Figure 3.Standard architecture of NLP system The discourse level examines the meaning of the
sentence in dependence of the other sentence in the
text or paragraph in the same document. The
238 TEM Journal – Volume 5 / Number 2 / 2016.
TEM Journal 5(2) 236–240
structure of this level is predictable and because of a) Can be degraded into several lexemes, and
thatreason, is use by NLP to understand the role of b) Can show lexical, syntactic, semantic,
each information in a document[1], [3]. pragmatic and/or statistical idiomaticity [8].
Pragmatic level When one or more parts of an MWE are not part of
the conventional lexicon then we can say that lexical
This level deals with the analysis of sentences and idiomaticityoccurs. For example, in the word Wi - Fi,
how they are used in different situations.In addition, the components (Wi and Fi) do nothave a meaning
also how their meaning changes depending on the when they are separated.
situation [3].
All the levels described here are inseparable and
complement each other. The aim of NLP systems is
to inject these definitions into a computer, and then
using them to create a structured unambiguous
sentence with well-defined meaning [3], [7], [12],
[16].
3.1. Learning about Natural Language Processing
As a discipline of informatics, NLP deals with the Figure 4.A classification for MWEs [8]
relationship among computers and natural
languages.Understanding natural languages is an AI Syntactic idiomaticity happens when the syntax of
problem because the identification of natural the MWE can not be extracted from that of its
languages requires understanding of sciences such as building blocks [8], [16]. For example, we can
computer science, statistics, science of language and analyse a syntactically idiomatic "on the whole”.
many others. Today’s modern algorithms for NLP This word is adverbial but consists of a preposition
are based mainly on statistical machine learning. (on) and an adjective (whole).
Today’s modern algorithms for NLP are mainly When the understanding of MWE is not coming
basedon statistical machine learning. Machine from its building blocks, we are talking about
learning aims to learn computer automatically from a semantic idiomaticity [8]. We can also use semantic
large number of different corpus, in a different way idiomaticity for allegorical language [8].
from the methods of processing the natural When some MWE with a fixed location, position
language.A corpus is a set of documents that are or situation are related, then we talk about pragmatic
annotated by hand with the correct values and it idiomaticity [8]. For example good night and good
serves for learning purpose. luck. The former is a greeting associated specifically
with nights and the latter is having a good wishes for
3.2. Statistical NLP the people who work in mines.
Statistic idiomaticity occurs when some
Statistical processing of natural language uses combination of words (term under) is repeated often
many different methods to solve problems. Some of in texts, unlike the words that make up the term
those methods use probability, some use statistic, and under. Anotherexample is impeccable credentials,
others use mathematics and so on. Problems may which occur much more frequently thanspotless
arise with words that have multiple meaning and with credentials.Part of statistical idiomaticity are the
sentences that are too long. Usually long sentences binomials, for example true and false, where the
can be interpreted in several different ways. Methods opposite adjective does not keep the sense of the
for clarifying sentences usually use corpus and term under [8].
Markov models. Statistical NLP is composed of
many techniques (such as modelling based on 4. Conclusion
probability, algebra, and theory of information) for
automated processing of natural or spoken language Study about natural language processing is
[7]. progressively changing from semantics to narrative
comprehension. In addition, people opinionis that the
3.3. Multi-word expressions (MWE) processing of a natural language represents a big AI
difficulty. This is without any doubt a top challenge
Multi-word expressions (MWEs) have paid that yet to be solved- make computers smart and
attention from the NLP community. They are lexical intelligent as people are. The future of NLP is
elements and they:
TEM Journal – Volume 5 / Number 2 / 2016. 239
no reviews yet
Please Login to review.
