American Journal of Computer Science and Engineering Survey Open Access

Keywords

NLTK; Brown Corpus; WordNet; Sentence similarity; Word order vector; Word similarity; Automatic Text Summarizer

Introduction

With fastest growing internet facility and technological people currently trust internet more than a bank locker or private safe. People are using internet to keep data secret and safe from the attack of unauthorized personnel. As we know World Wide Web in 21st century makes revolution in technology field or to be precise in the Information Technology, and the more revolutions are waiting to come. To keep this revolution moving on the Data Mining approach becomes a buzzword in these decades. In a short-term data mining refers to approach the practice of examining large pre-existing databases in order to generate new information [1]. A small field of data mining is text mining. Text mining, also referred to as text data mining, roughly equivalent to text analytics, is the process of deriving high-quality information from text [2]. Text mining redirects to one of the important applications which is text summarization [3-5].

Text summarization is one of the challenging and important problems of text mining. It provides several benefits to users. A vast number of real-life problems can also come to a fruitful conclusion with text summarization. In this chapter we will give some idea about text summarizing and its automotive mode. We will also cover the classification of text summarization. Then we will discuss about the objectives of our paper. After that, we will cover steps of summarizing process along with brief discussion about extraction based summarizing technique. At the end of this chapter we will try to illustrate why we choose this approach for Bengali Language [6-9].

The extractive summarization process evolves with several fascinating algorithms. From those algorithms we worked with the statistical approach and graph-based sentence similarity approach [10,11].

Statistical Approach

In statistical approach we first initialize the min_cut and max_cut values. Words that have probabilistic appearance less than min_ cut and more than max_cut will be ignored. The list of stop words then will also be considered as non-countable. After tokenizing every word except the stop words, we calculated frequency of each word. In this situation all the words are mapped using python’s hash map structure that stores the words as an object. Then we did calculate the maximum among frequency of words. Then to figure out the statistical score of words we will use the formula given below-

Here, Ai is the number of appearances of a certain word obviously without stop words and M is the maximum number among frequency of words.

Doing this calculation for each of the sentences we got the statistical score of each word. After this calculation we have calculated the total score of a sentence by placing the score of all the words and for the stop words it the score is null.

The equation of the score o f a sentence can be shown as –

By implementation of this equation once we have the score of all the sentences, we can give output the sentences with most scored K sentences at our ease. As we are implanting this algorithm in Bengali, we need to make a different library of Bengali stop words.

Pre-Processing

Text pre-processing simply illustrates organize texts for further usage. Pre-Processing is vital pre-requisite while working with noisy text content like social media, news content crawler and all that. Pre-processing involves splitting content into valid tokens, words and symbols, stemming, moving out stop words and parts of speech tagging. Here we have used the nltk tokenizer with a slight modification although built for English works smoothly for Bengali as well [12].

Computation of Word Similarity

Working on the word similarity we have to first illustrate the WordNet lexicon analysis which is a Bengali lexical knowledge base that will obviously help for NLP and next we will illustrate the word similarity procedure [13].

Wordnet lexicon formation

To calculate word similarity between any two Bengali words a lexicon should be developed. In this we have used the lexicon developed by (Manjira, Sinha, Tithankar, Anupam). In that lexicon they constructed a hierarchical conceptual graph. The storage and organization of their database facilitate computational processing of the information and efficient searching to retrieve the details associated with any words [14]. Therefore, it is really a fantastic resource and tool for NLP analysing in Bengali. It supports query like DETAILS(X) where X can be any type of node in hierarchy and SIMILARITY (W1, W2) that returns the similarity between any two words based on the word’s organization i.e. either they are synonym, similar, stop word etc. The formation of the lexicon is pretty simple. The lexicon contains 30 different sections as 30 different roots. Words are also categorized, sub-categorized using parts of speech. Corresponding each of the category there are two clusters. The first one contains the synonym of a word and the other one contains related word [14]. Figure 1 below will illustrate the lexicon basic formation.

Word similarity computation

To calculate word similarity based on the lexicon we will measure the distance between two words. Let us given two words “Universe” and “Universe”. As both of them are in same cluster i.e. within the synonym cluster then the similarity is obviously same. But if we consider another two words “Universe” and “Astronomer”. Both of them are in different cluster. They are not synonyms to each other. Instead they are related word to each other and the distance between those two words are path length which is 4. As we are calculating similarity then we must have to choose the words with more similarity i.e. lesser path distance. However, this method may be less accurate if it is applied to larger and for more general semantic nets such as WordNet. For example, let us have the distance between words “Man” and “Animal” is lesser than the distance between words “Man” and “Doctor”. Then obviously similarity returns wrong result. To resolve this issue, we may create categorized word lexicon. If we create a category called “Human” then put all related words under this category and another category called “Animal” then put all the related words under this category we hope task will be a bit easier and the distance will be a larger such as adding the real distance with a constant value [15,16]. Therefore, the depth of word in the hierarchy should be taken into account. In summary, similarity between words is determined not only by path lengths but also by depth. We propose that the similarity S (w1, w2) between words w1 and w2 is a function of path length and depth as follows:

where l is the shortest path length between W1 and W2, h is the depth of subsume in the hierarchical semantic nets. We assume that (3) can be rewritten using two independent functions as:

f1 and f2 are transfer functions of path length and depth, respectively. We call these information sources of path length and depth, attributes [17-19].

Transfer function and its characteristics

Transfer functions are those functions that calculate similarities between words or sentences based on several parameters. The values of an attribute in equation (4) may cover a large range up to infinity, while the interval of similarity should be finite with extremes of exactly the same with a value of 1 and no similarity as 0, then the interval similarity is (0, 1). Direct use of information sources as a metric of similarity is inappropriate due to it infinite property. Therefore, it is intuitive that the transfer function from information sources to semantic similarity is a nonlinear function. Taking path length as an example, when the path length decreases to zero, the similarity would monotonically increase toward the limit 1, while the path length increases infinitely, the similarity should monotonically decrease to 0. So, to meet these constraints the transfer function must be a nonlinear function. The nonlinearity of the transfer function is taken into account in the derivation of the formula for semantic similarity between two words [20,21].

Path Length and its Contribution

Considering the path length between any two words w1 and w2 can be determined by one of the following three cases:

1. w1 and w2 are in the same synset.

2. w1 and w2 are not in the same synset, but the synset for w1 and w2 contains one or more common words.

3. w1 and w2 are neither in the same synset nor do their synsets contain any common words.

Case 1 simply illustrate that both of the words are of same meaning and hence the path length is 0. For case 2 partially they share the same feature and hence path length is 1. For case 3, we will compute the actual path length. Based on the previous consideration we set f 1(l) (4) to be a monotonically decreasing function of l:

where, α is a constant. The selection of the function in exponential form ensures that f 1(l) satisfies the constraint and the value of f 1(l) within the range of 0 and 1.

Effects of Depth Scaling

Words at upper layers of hierarchical semantic nets have more general concepts and less semantic similarity between words than words at lower layers. This behavior must be taken into account in calculating S (w1, w2). We therefore need to scale down S (w1, w2) for subsuming words at lower layers. As a result, f2(h)should be monotonically increasing with respect to h. So, f 2 will be:

where, is a smoothing factor. As then the depth of a word in the semantic nets is not considered. In summary, the formula for a word similarity measure as:

Where

are parameters scaling the contribution of shortest path length and depth, respectively. The optimal values of and ' are dependent on the knowledge base used and can be determined using a set of word pairs with human similarity rating. With several calculations it has been seen that the proposed measure of is 0:2 and 0:45 [22].

Semantic Similarity between Sentences

A sentence consists of some words. It would be easier to calculate things if we represent a sentence as a vector of words. Given two sentence T1 and T2, a joint word set is formed:

The joint word set T contains all the distinct word set. Since inflectional morphology may cause a word to appear in a sentence with a different form that conveys a specific meaning for a specific context, we consider them as different words. For example, boy and boys considered as two different words and all will be included in the word set [23-25].

Here in T the algorithm removes duplicate words skipping the stop words. The vector derived from the joint word set is called the lexical semantic vector, denoted by S. Each entry of the semantic vector corresponds to a word in the joint word set, so the dimension equals the number of words in the joint word set. The value of an entry of the lexical semantic vector, Si (i=1, 2,….n) is determined by the semantic similarity of the corresponding word to a word in the sentence. Take T1 as an example:

Case 1: If Wi appears in the sentence, Si is set 1

Case 2: If Wi is not contained in T1, a semantic similarity score is computed between Wi and each word in the sentence T1. Thus, the most similar word in T1 to Wi is that with the highest similarity score γ If γ exceeds a present threshold, then S I = γ otherwise, S I = 0. The reason for the introduction of the threshold is twofold. First, since we use the word Similarity of distinct words (different words), the maximum similarity scores may be very low, indicating that the words are highly dissimilar. In this case, we would not want to introduce such noise to the semantic vector. Second, classical word matching methods (1) can be unified into the proposed method by simply setting the threshold equal to one. Unlike classical methods, we also keep all function words. This is because function words carry syntactic information that cannot be ignored if a text is very short (e.g. sentence length). Although function words are retained in the joint word set, they contribute less to the meaning of a sentence than other words. Furthermore, different words contribute toward the meaning of a sentence to differing degrees. Thus, a scheme is needed to weight each word. We weight the significance of a word using its information content. It has been shown that words that occur with a higher frequency (in a corpus) contain less information than those that occur with lower frequencies [26]. The value of an entry of a semantic vector is:

where, Wi is a word of joint set, is its associated word in sentence. The use of I (Wi) and allows the concerned two words to contribute to the similarity based on their individual information contents. Here wi and can be calculated as follows:

Where, n = number of times word w in corpus N = number of words in the corpus, So, I ∈ [0,1]. The semantic similarity between two sentences is defined as the cosine coefficient between the two vectors:

It is worth mentioning that the method does not currently conduct word sense disambiguation for poly-serous words. This is based on the following consideration: First we wanted to our model to be as simple as possible and not too demanding in terms of computing resources. The integration of word sense disambiguation would scale up the model complexity. Second existing sentences similarity methods do not have included word sense [27].

Word Order Similarity between Sentences

The word order similarity attempts to correct for the fact that sentences with the same words can have radically different meanings. This is done by computing the word order vector for each sentence and computing a normalized similarity measure between them. The word order vector, like the semantic vector is based on the joint word set. If the word occurs in the sentence, its position in the joint word set is recorded. If not, the similarity to the most similar word in the sentence is recorded if it crosses a threshold η else it is 0. In equations,

Where, P1= word position vector for sentence 1

P2 = word position vector for sentence 2

That is, word order similarity is determined by the normalized difference of word order. The following analysis will demonstrate that Sp is an efficient metric for indicating word order similarity. To simplify the analysis, we will consider only a single word order difference, as in sentences T1 and T2. Given two sentences, T1 and T2, where both sentences contain exactly the same words and the only difference is that a pair of words in T1 appears in the reverse order in T2. The word order vectors are:

aj and aj+k are the entries for the considered word pair in T1, bj and bj+k are the corresponding entries for the word pair in T2, and k is the number of words from wj to wj+k. From the above assumptions, we have:

Then,

We can also derive the same formula for a sentence pair with only one different word at the kth entry. For the more general case with a more significant difference in word order or a larger number of different words, the analytical form of the proposed metric becomes more complicated (which we do not intend to present in this paper). The above analysis shows that the sp is a suitable indication of word order information. sp equals 1 if there is no word order difference. sp is greater than or equal to 0 if word order difference is present. Since sp is a function of k, it can reflect the word order difference and the compactness of a word pair. The following features of the proposed word order metric can also be observed:

1. sp can reflect the words shared by two sentences.

2. sp can reflect the order of a pair of the same words in two sentences. It only indicates the word order, while it is invariant regardless of the location of the word pair in an individual sentence.

3. sp is sensitive to the distance between the two words of the word pair. Its value decreases as the distance increase.

4. For the same number of different words or the same number of word pairs in a different order, sp is proportional to the sentence length (number of words); its value increases as the sentence length increases. This coincides with intuitive knowledge, that is, two sentences would share more of the same words for a certain number of different words or different word order if the sentence length is longer. Therefore, the proposed metric is a good one [28-31].

Sentence Similarity in Total

Similarity between sentences based on semantic nets is basically the lexical similarity of sentences. The relationship between words can be illustrated as word order similarity. This word order similarity illustrates which word will be there in the summarized sentence and which words will be there before which words. The information that is collected in semantic and syntactic way provides information about the meaning of the sentences. The similarity between two sentences is modelled as a linear combination of their semantic similarity and word order similarity. In equations –

Here the value of δ ≤ 1 decides the relative contributions of semantic and word order information to the overall similarity computation. As syntax plays a sub-ordinate role for semantic processing of text [32], should be a value greater than 0.5, i.e. δ ∈ (0.5, 1].

As a short note we would like to mention that, implementing this method requires the use of two databases. The first one is WordNet and the second one is Brown Corpus. WordNet is lexical database of words that returns similarity issue between two words. It’s a more like tree hierarchical formation. Starting from the several roots it will move towards leaf in order to find out whether words are in same corpus or not. The algorithm then takes decisions based on this return value [33,34]. The Brown corpus is also a database of lexical words but it computes the appearance of words within the corpus using relative frequency:

Here, n = number of times word w occurs in corpus N = number of words in the corpus

Result

We gave tested in 5 inputs of a Bengali text paragraph for text summarizing. We have analysis the statistical approach, Sentence Similarity approach, a result by a random Bengali person. It is shown below in Table 1.

Table 1: Text summarizing result.

One of the text summarizing result

Input: A few sentences about greatness of humanity in Bengali, words counted 70.

Output: Statistical approach A few sentences about greatness of humanity in Bengali words counted 23.

Output: Sentence Similarity A few sentences about greatness of humanity in Bengali words counted 6.

Output: A random User A few sentences about greatness of humanity in Bengali words counted 15.

Analysis:

Comparing both of the algorithm it is clear that both output illustrates greatness of human being but we look in text words sentence similarity gives the shortest summarization result but on the other hand statistical approach gives longest summarizing result If we take a look at the random user’s output it also points to the same direction. Analysing just about the output we can say that both algorithm’s accuracy is about 80 (Figure 2).

Discussion

While processing the randomized inputs we have seen that most of the Statistical Approach Summarizer returns the lengthy lines which contain more information and also in a very short time. Also, as there is no use of complex algorithms and techniques and hence processing time is pretty lower for Statistical Approach. Whenever we read about the output of the statistical approach, we have seen that it somehow misses some important gist. Whereas sentence similarity approach mostly does not miss the gist’s. The sentence similarity takes much more time with respect to the statistical approach but it returns the less sized line mostly. And also, it gives a very realistic result. One thing that we have noticed is with increasing size of text the run time is increasing almost exponential way for sentence similarity approach whereas the statistical Approach do not have such characteristics. The complete timing issue can be shown as the following graph. Text size and run time analysis is shown in Figure 2.

Above analysis show the text size – run time analysis bar chart. The slightly blue shades on the graph denote run time of Statistical Approach. And the red bar denotes the run time for Sentence Similarity approach.

Accuracy

From the analysis for each of the inputs we are able to see that both of our approaches at least capable to show the gist’s of the main content. Reading the output, we can at least gather an idea that what it is talking about. In statistical analysis we missed some of the gist’s for Sample Input 5. Whereas sentence similarity managed to show that gist’s. From this point of view, sentence similarity is more accurate. Although time complexity of sentence similarity is a big deal whereas statistical analysis takes much less time. Apart from that we think that somehow all of our idea managed to see the daylight.

Conclusion

The sentence similarity-based summarizer is able to give a satisfactory result for a wide range of documents. The statistical approach is able to process a large amount data at a very short time. The sentence similarity approach provides more realistic result and contains more gist’s than statistical approach. Lack of Bangla WordNet and Corpus decrease the accuracy of Sentence similarity approach. Caching of the calculation may also be a great choice to reduce the run time of Sentence similarity.

References

1. Verloren (2002) Wikipedia Data Mining Article.
2. Hearst M (2003) What is text mining? Wikipedia Text Mining Article.
3. Allen J (1995) Natural Language Understanding. Redwood City, Calif: Benjamin/Cummings 2nd Edition.
4. Atkinson J, Mellish C, Aitken S (2004) Combining Information Extraction with Genetic Algorithms for Text Mining. IEEE Intell Syst 19: 22-30.
5. Leech, Geoffrey, Smith N (2005) Extending the possibilities of corpus-based research on English in the twentieth century: A prequel to LOB and F-LOB. Int Comput Arch Mod Mediev English 29: 83-98.
6. Juan-Manuel Torres-Moreno (2014) Automatic Text Summarization (Cognitive Science and Knowledge Management). ISTE Ltd; Hoboken, NJ: John Wiley & Sons, Inc,1st Edition, 2014.
7. Mani I, Maybury TM (1999) Advances in Automatic Text Summarization. The MIT Press, 55 Hayward St. Cambridge, MA, United States, ISBN:978-0-262-13359-3.
8. Allahyari M, Pouriyeh S, Assefi M, Safaei S, Trippe ED et al. (2017) Text Summarization Techniques: A Brief Survey. Int J Adv Comput Sci Appl 8: 397-405.
9. Kumar YJ, Goh OS, Basiron H, Choon NH, Suppiah PC (2016) A Review on Automatic Text Summarization Approaches. J Comput Sci 12: 178-190.
10. AL-Khassawneh YA, Salim N, Jarrah M (2017) Improving Triangle-Graph Based Text Summarization using Hybrid Similarity Function. Indian J Sci Technol 10: 1-15.
11. Erkan G, Radev DR (2004) LexRank: Graph-Based Lexical Centrality as Salience in Text Summarization. J Artif Intell Res 22: 457-479.
12. Budanitsky A, Hirst G (2001) Semantic Distance in WordNet: An Experimental, Application-Oriented Evaluation of Five Measures. Proceedings at Workshop WordNet and Other Lexical Resources, Second Meeting of the North American Chapter of   the Association for Computational Linguistics 27: 1-6.
13. Rudrapal D, Das A, Bhattacharya B (2015) Measuring Semantic Similarity for Bengali Tweets Using WordNet. Proceedings of Recent Advances in Natural Language Processing: 537–544.
14. Sinha M, Jana A, Dasgupta T, Basu A (2012). A semantic lexicon and similarity measure in Bangla. Proceedings of the 3rd Workshop on Cognitive Aspects of the Lexicon (CogALex-III): 171–182.
15. Foltz PW, Kintsch W, Landauer TK (1998) The Measurement of Textual Coherence with Latent Semantic Analysis. Discourse Processes 25: 285-307.
16. Jurafsky D, Martin JH (2008) Speech and Language Processing: An Introduction to Natural Language Processing, Computational Linguistics, and Speech Recognition. Prentice Hall: 1-16.
17. Ko Y, Park J, Seo J (2004) Improving Text Categorization Using the Importance of Sentences. Inf Process Manag 40: 65-79.
18. Kozima H (1994) Computing Lexical Cohesion as a Tool for Text Analysis. Univ Electro Commun: 1-39.
19. Landauer TK, Laham D, Rehder B, Schreiner ME (1997) How Well Can Passage Meaning Be Derived without Using Word Order? A Comparison of Latent Semantic Analysis and Humans. Proceedings of 19th Annual Meeting of the Cognitive Science Society: 412- 417.
20. Landauer TK, Foltz PW, Laham D (2008) Introduction to Latent Semantic Analysis. Discourse Processes 25: 259- 284.
21. Landauer TK, Laham D, Foltz PW (1999) Learning Human-Like Knowledge by Singular Value Decomposition: A Progress Report. Advances in Neural Information Processing Systems 10, Jordan MI, Kearns MJ, Solla SA, eds., Cambridge, Mass : MIT Press 1998 :45-51.
22. Li Y, Bandar Z, McLean D (2003) An Approach for Measuring Semantic Similarity Using Multiple Information Sources. IEEE Trans Knowl Data Eng 15: 871-882.
23. Liu Y, Zong C (2004) Example-Based Chinese-English MT. Proc 2004, IEEE Int Conf Syst Man Cybern 7: 6093-6096.
24. McClelland JL, Kawamoto AH (1986) Mechanisms of Sentence Processing: Assigning Roles to Constituents of Sentences. Parallel Distributed Process 2, Rumelhart DE, McClelland JL and the PDP Research Group, MIT Press 2: 272-325.
25. McHale ML (1998) A Comparison of WordNet and Roget’s Taxonomy for Measuring Semantic Similarity. Proceedings of COLING/ACL Workshop Usage of WordNet in Natural Language Processing Systems, 1998.
26. Meadow CT, Boyce BR, Kraft DH (2000) Text Information Retrieval Systems, second edition, Academic Press, 2000.
27. Michie D (2001) Return of the Imitation Game. Electro Trans Artif Intell 5-B2: 203-221.
28. Burgess C, Livesay K, Lund K (1998) Explorations in Context Space: Words, sentences, discourse. Discourse Processes 25: 211-257.
29. Charles WG (2000) Contextual Correlates of Meaning. Appl Psycholinguist 21: 505-524.
30. Chiang JH, Yu HC (2005) Literature Extraction of Protein Functions Using Sentence Pattern Mining. IEEE Trans Knowl Data Eng 17: 1088-1098.
31. Coelho TAS, Calado PP, Souza LV, Ribeiro-Neto B, Muntz R (2004) Image Retrieval Using Multiple Evidence Ranking. IEEE Trans Knowl Data Eng 16: 408-417
32. Wiemer-Hastings P (2000) Adding Syntactic Information to LSA. Proceedings of 22nd Annual Conference of the Cognitive Science Society 22: 989-993.
33. Hatzivassiloglou V, Klavans JL, E. Eskin E (1999) Detecting Text Similarity over Short Passages: Exploring Linguistic Feature Combinations via Machine Learning. Proceedings of the Joint SIGDAT Conference on Empirical Methods in and National Language Processing and Very Large Corpora, 1999: 203-212.
34. Hatzivassiloglou V, Shaw J (1999) Ordering among premodifiers. Proceedings of the 37th Annual meeting of the Association for Computational Linguistics :135-143.