CUSAT TEAM@DPIL-FIRE2016: Detecting Paraphrase in Indian Languages-Malayalam Manju K Sumam Mary Idicula Research Scholar, Head of Department, Department of Computer science and Department of Computer science and Engineering, Engineering, Cochin University of Science and Cochin University of Science and Technology,India. Technology,India. manju@mec.ac.in sumam@cusat.ac.in ABSTRACT 2. RELATED WORKS This paper describes the work done as part of the shared task Paraphrase identification has a lot of significance in differ- on Detecting Paraphrases in Indian Languages(DPIL) in Fo- ent areas of Natural language Processing. Paraphrase iden- rum for Information Retrieval and Evaluation(FIRE 2016). tification techniques are mainly classified into statistical and Paraphrase identification is the task of deciding whether two semantic methods. In statistical methods, the similarity be- given text fragments have the same meaning. Our detection tween sentences is measured only on the basis of statisti- system is for Malayalam language and makes use of the co- cal information in the sentences whereas semantic method sine similarity measure, an existing state of the art method makes use of word meanings. Work which shows the compar- for determining the similarity between sentences. The ex- ison of statistical and semantic similarity measures[1], which periments were done on the standard data set and the re- was tested on the same data set stated that the performance sults showed that the system was able to give performance of both measures are comparable. One of the most com- comparable to methods employing more sophisticated pro- monly used corpora for paraphrase detection is the MSRP cedures. corpus[3], which contains 5,801 English sentence pairs from news articles manually labelled with 67% paraphrases and 33% non-paraphrases. Since there are no annotated corpora CCS Concepts or automated semantic interpretation systems available for •Information Processing → Similarity Measures; •Natural Indian languages till date, the initiative made as part of Language Processing → Paraphrase Identification; •Text the open shared task competition is highly appreciable and Mining → Text Summarization; is of great help to the research community. The automatic plagiarism detection framework for Malayalam documents[5] uses Jaccard similarity for determining the relation between Keywords sentences. Paraphrase; Cosine similarity; text tagging The proposed method implements Paraphrase Identifica- tion for Malayalam Language using similarity measures[4]. 1. INTRODUCTION Paraphrases are alternate ways to convey the same infor- 3. TASK DESCRIPTION mation. In natural languages, we can express a single event The task is focused on sentence level paraphrase identi- in different ways which conveys the same information. Para- fication for Indian languages-Tamil, Malayalam, Hindi and phrase identification, the ability to determine whether two Punjabi. The proposed method considers only Malayalam formally distinct strings are similar or not, have applica- language. Malayalam is one among the 22 scheduled lan- tion in various NLP tasks like Information retrieval, Ques- guages of India. It is the official language in the state of tion Answering, Plagiarism detection, Text Mining and Au- Kerala and in the Union territories of Lakshadweep and tomatic summarization. Paraphrase identification basically Puduchery. Malayalam belongs to the Dravidian language uses a simple lexical matching comparison of sentences. family and is spoken by approximately 33 million people.The In order to select a sentence pair as paraphrase, they task provided is divided into two sub tasks where sub task should describe the same event and should contain same 1 is to classify the given pair of sentences to paraphrase or information about the event. However there are instances non paraphrase and in sub task 2 the sentences are classi- when the concept behind the sentences are difficult to iden- fied on a 3 point scale, to completely equivalent(P), roughly tify, even for humans this is a difficult task. equivalent(SP) or not equivalent(NP). The rest of the paper is organized as follows: Section 2 dis- cusses related work in the area of Paraphrase detection. Sec- tion 3 presents the Task Description. Section 4 tells about 4. DATA SET the data set provided by the DPIL task[2] organizers. Sec- The shared task challenge provided data for four lan- tion 5 explains the methodology used and Section 6 gives the guages Tamil, Malayalam, Hindi and Punjabi. We were Result and evaluation. Section 7 presents the conclusion and provided with 2500 sentence pairs for sub task 1 and 3500 the future improvements that can be made. sentence pairs for sub task 2 as training data and 900 sen- tence pairs for sub task 1 and 1500 sentence pairs for sub task 2 as test data. The data set available was in XML D1 ∗ D2 format taken from prominent Newspapers. Sim(D1 , D2 ) = p p (3) D12 ∗ D22 Similarity score will be a value between 0 and 1. 5. SYSTEM DESCRIPTION It was decided to set a threshold for determining the classes Data was given in XML format and that file was processed Paraphrase, Semi Paraphrase and Non Paraphrase. Through to extract each pair of sentences for paraphrase detection. experiment using the training data given for task1 and task2 Cosine similarity measure was used for paraphrase identi- a threshold of 0.4 was set for Paraphrase, 0.3 for SemiPara- fication and the concerned two sentences in each pair was phrase and any value less than that as NonParaphrase. considered as two documents D1 and D2 . D1 and D2 con- tain only one sentence each. The overall architecture of the system is shown in Fig 1. D1 and D2 are subjected to tok- 6. RESULTS AND EVALUATION enization and stop word removal. A look up table was used The proposed system was experimented with the data set for stop word removal. Due to the agglutinative nature of provided by the open shared task. Fig 2 shows the similarity the language, the same word can appear with different in- score obtained for the 3 classes of sentence pairs. flections in the sentences. To eliminate these inflections, stemming was performed. Even though literature related to stemming in Malayalam language is available, there is no full fledged tool which can be used in the work. We have custom tailored the Silpa Stemmer[6] by Swathanthra Malayalam Computing group for our purpose. The stemmer removes longest matching suffix from each word with proper replacement to get the base word. Figure 2: Similarity Score Obtained The accuracy and F-score for this methodology of para- phrase identification is tabulated in Table 1 for subtask 1 Figure 1: System Architecture and subtask 2 The words in the resulting sentences after preprocessing are the bag of words(vocabulary) for the vector represen- Table 1: Results tation of the sentences. The sentence vector is formulated Language SubTask1 SubTask2 using bag-of-words model to extract frequency information Accuracy F1 Score Accuracy F1 Score of words in the sentence. The size of the vector will be the Malayalam 0.80444 0.76 0.50857 0.46576 size of the vocabulary set and the value at each vector in- dex i represents the count of word i in the sentence. This is the Term Frequency(TF) Vector.For determining the im- 7. CONCLUSION portance of each word with respect to the two documents its Inverse Document Frequency (IDF) is also calculated ac- This paper discussed on how cosine similarity can be used cording to equation(1). for Paraphrase identification. The morphological richness and agglutinative nature of the language demands for stem- ming of the sentence pairs before paraphrase scoring. The N accuracy of the preprocessing phase has got a significant Idft = log (1) Nt role in the paraphrase identification system. Performance of where N is the total sentences in a document D, here it is 2 the system can be improved by considering semantic simi- and Nt is the number of sentences in which the term t occurs. larity using word net in addition to statistical measures. An The sentence vector is computed according to equation(2). ensemble of different similarity scores may improve the ac- curacy of the system. The vague demarcation between semi Si = T ft,i ∗ Idft (2) paraphrase and non paraphrase is a challenge in this type of work. where T ft,i is the frequency of term t in Sentence Si and Idft gives the information, how important is the term t. Using equation(3) the similarity between documents are computed 8. REFERENCES where D1 contains the first sentence and D2 contains the [1] S. S. Abraham and S. M. Idicula. Comparison of second sentence in the pair. statistical and semantic similarity techniques for paraphrase identification. In 2012 International Conference on Data Science & Engineering (ICDSE). [2] M. Anand Kumar, S. Shivkaran, B. Kavirajan, and K. P. Soman. DPIL@FIRE2016: Overview of shared task on detecting paraphrases in indian languages. In Working notes of FIRE 2016 - Forum for Information Retrieval Evaluation, Kolkata, India, December 7-10, 2016, CEUR Workshop Proceedings. CEUR-WS.org, 2016. [3] W. B. Dolan and C. Brockett. Automatically constructing a corpus of sentential paraphrases. In Proc. of IWP, 2005. [4] S. Fernando and M. Stevenson. A semantic similarity approach to paraphrase detection. In Proceedings of the 11th Annual Research Colloquium of the UK Special Interest Group for Computational Linguistics, pages 45–52. Citeseer, 2008. [5] L. Sindhu, B. B. Thomas, and S. M. Idicula. Automated plagiarism detection system for malayalam text documents. International Journal of Computer Applications, 106(15), 2014. [6] S. Thottungal. Silpastemmer: http://libindic.org/stemmer.