ARTH: Algorithm For Reading Text Handily - An AI Aid for People having Word Processing Issues

Misha is a mildly mentally disabled child of age 15 years. She is declared unfit to continue further schooling and left her school in 3rd grade. Her mother tried to teach her at home, but due to her poor vocabulary level, it took a lot of time to read one page of her textbook. Thus, she gave up on her reading habits even though she loved reading. This technology aims to revive the joy of reading/learning in the lives of such people. With this technique, she can read stuff on any reading device, as the meaning of the words which are unknown to her are displayed next to them while she is reading, thereby not hindering her experience.

The algorithm starts with the reading of the raw text file. The document received is preprocessed, to make it ready for further steps. It uses algorithms for natural language processing to solve the designated task. The process can be viewed graphically by using flowchart in the Figure LABEL:fig:2 :

Lemmatization and stemming are two algorithms that deal with the process of reduction of inflectional forms and sometimes derivationally related forms of a word to a common base form. Stemming refers to a crude heuristic process that chops off the ends of words in the hope of achieving this goal correctly most of the time and often includes the removal of derivational affixes. Lemmatization usually refers to doing things correctly with the use of vocabulary and morphological analysis of words, typically aiming to remove inflectional endings only and to return the base or dictionary form of a word, which is known as the lemma.

Tokenization is the task of chopping document up into words, also called tokens, perhaps at the same time throwing away certain characters, such as punctuation. A token is an instance of a sequence of characters which becomes a useful semantic unit for processing in some particular document. Tokenization in other terms is the detection of the end of the word and start of next word. Therefore, tokenization is sometimes also referred as word segmentation.

The stopwords are the words frequently occurring in most of the texts. The general strategy for determining a stop word is to sort the terms by collection frequency (the total number of times each term appears in the document collection), and then to take the most frequent terms, often hand-filtered for their semantic content relative to the domain of the documents being indexed, as a stop list , the members of which are then discarded during indexing. In ARTH case, a generalized list of stopwords mostly consisting of articles, prepositions, etc are used.

ARTH uses two parameters for the detection of difficulty,

Syllables: Syllables have a direct relation to the difficulty of the word. It is defined as a unit of pronunciation having one vowel sound, with or without surrounding consonants, forming the whole or a part of a word. Therefore an increase in the number of syllables will cause an increase in the difficulty of the word.

Usage Frequency: The usage frequency is indirectly related to the difficulty of the word. More is the usage of a particular word, more likely it is that the user will be familiar with the word.

Syllabification is a method to know the number of syllables in a word. A syllable is a unit of organization for a sequence of speech sounds. A syllable tells us how troublesome it is for a person to enunciate a word. Greater the number of syllables, harder it is to pronounce and understand them. Thus, the number of syllables is chosen as a parameter to decide the difficulty of word. There are different ways of doing syllabification:

1. 1.

   Rule Based Approach

2. 2.

   Machine Learning Approach

Using rule-based approach, we can actually develop a fine algorithm. I researched a lot about various rules that can tell about the number of syllable in a word[2]. The accuracy of above approach is 82.5%

The calculation of the number of syllables in a word can be regarded as the classification problem where for a particular word we have to predict the number of syllables (class), which is essentially a set of discrete classes. For the experimentation purpose, I have used the most basic classifier of all the classifiers known, called Naive Bayes Classifier.

The words, which we read occasionally are more familiar and easily understood by the people. By the term usage of the word, we mean the number of times a word has occurred in the texts available online. The relationship between usage and difficulty of the word is inverse. The masses quickly understand the frequently used words, but the meaning of less commonly used words has less probability of being known to the user. For calculation of usage, a database of the pool of words with their frequency is stored. All the words read in the text are searched in this database and their corresponding frequency is stored.

Normalization is a re-scaling technique. The maximum value usage frequency feature can go up to millions, while the highest value of syllable feature is 12. There is a huge gap in the features. Since we are going to use these features for a clustering, so it is important to rescale them. As in K-means, we use Euclidean distance, and if we want that all the features to contribute equally, then we have to do normalization. There are two methods of rescaling the features:

1. 1.

   Min - Max scaling: In this score, the data is rescaled to lie within the fixed range, usually 0 to 1. The effect of outliers is diminished, when we use min-max scaling

   $$X_{norm}=\frac{X-X_{min}}{X_{max}-X_{min}}$$

2. 2.

   Z-Score:The features will be rescaled so that they’ll have the properties of a standard normal distribution with $\mu=0$ and $\sigma=1$ , where $\mu$ is the mean (average) and $\sigma$ is the standard deviation from the mean; standard scores (also called z-scores) of the samples are calculated as follows:

   $$z=\frac{X-\mu}{\sigma}$$

Z-score is usually a preferred method while dealing with the clustering algorithm and hence z-score is used in the project for normalization.

Unsupervised machine learning is the machine learning task of inferring a function to describe hidden structure from "unlabeled" data. We try to find structure in the dataset using unsupervised learning. There are many algorithms for unsupervised machine learning. In ARTH for clustering of words into different levels of difficulty clustering algorithm is used. There are a lot of clustering algorithms like K-means, DBSCAN, Agglomerative Clustering, etc. I have implemented K-means clustering.

The K in K-means clustering is based on length of the text entered. If the length is greater than 400 words we keep K as 5 else we keep as 3.

After dividing words into K clusters, we then try to identify vocabulary level of users.

The question formation algorithm is a brute force algorithm due to limited time constraint. In future, will research and make it better. Algorithm can be divided into following steps:

Step - 1 : Words Selection We want to check the vocabulary level of the person so we choose the words. For this reason, we select randomly four words from the clusters obtained in previous step.

Step 2 : Selection of sentences having those words The text is tokenized into sentences and select the sentence which have randomly chosen words.

Step 3: Creation of questions We create a template for question generation, “What is the meaning of” + word + “ in the sentence” + sentence + “?”

Correct option generator, generates answer for the question asked. We can use a dictionary based approach to find the correct meaning of the words. But there is a problem in this approach. Consider following two examples: I am going to bank for depositing money. I am going to bank for picnic. In first sentence, the word “bank” means a financial establishment. In second sentence, the word “bank” means the land sloping down to a river or lake. A single word can be used in multiple ways, having multiple meaning. This problem is known as Word Sense Disambiguation Problem in NLP. We implemented simplified Lesk algorithm for ARTH.

The task of wrong option generator is to generate wrong options so that the choices can be emerged. Steps involved in wrong option generator:

• •

  Antonym Generation: Using Wordnet we can find the antonym of the words thus generating one wrong example.

• •

  Find those synonyms of words, using Synsets of words from Wordnet, whose meaning doesn’t matches to the word. For measuring the similarity scores, we have used similarity metrics.

• •

  If similarity measures is greater than 5 then we accept it as wrong answer and add it to the list.

• •

  We randomly select words from the vocabulary and try to find similarity between them, if they are not similar, we include the word in our list.

Wu & Palmer measure (wup) : The Wu & Palmer measure (wup) calculates similarity by considering the depths of the two concepts in the UMLS, along with the depth of the LCS. The formula is

This means that 0 < score <= 1. The score can never be zero because the depth of the LCS is never zero (the depth of the root of a taxonomy is one). The score is one if the two input concepts are the same.

After the formation of questions, their correct answers and other wrong options, quiz can be generated in the UI for the user. The process of automated quiz generation is shown in the figure fig:6 :

The answers given by users are evaluated. As four questions represents one cluster. If for a particular cluster, the person answers correctly 75% of the questions, then we mark the cluster easy for the user. For those clusters, the person gives all wrong answers or only answers 25% correctly, we declare them as hard for user. If a person answers 50% correctly, then we again try to cluster that particular cluster into two groups, and ask quiz of six questions, if he answers 2 correct questions from a particular group, the group is labeled as easy for the person.