Automated summarization for English language has been an active area of research for quite a long time, but very less works have been done for multilingual and regional languages specially for the Indian languages. A summarization system consists of reduction of a text document to generate a new form which conveys the key meaning of the information contained in the text. This technique is useful due to the information overload and useful information being just a tiny fraction of it. Summarization reduces the time overload of reading the entire document and thus improves the readability and eficiency. Summarization can be understood as targeted dissemination of information for too long; didn't read (TLDR) texts. In this paper we have presented an extractive technique of text summarization employing K-means clustering on Indian Language Summarization (ILSUM) 2022 dataset and achieved satisfactory results by the team _ℎ.
Text Summarization task was very popular in past decades but requirement was diferent as
compared to now a days. At that time storage space was so scarce. Large documents needed
to be replaced with summaries for storage. There are now many inexpensive storage devices
available compared to that time, but because there is such a vast volume of data readily available,
it is necessary to transform that data into knowledge and information [
The two diferent techniques to summarization are extractive and abstractive. Abstractive is human way of summarization which requires knowledge outside of the current text too. Abstractive techniques are resource heavy and are yet to reach an accuracy comparable to a human, therefore Extractive techniques which, as the name suggests, generate summary by selecting words from the parent text, dominate the field of text summarization. Utilizing this strategy, summaries are produced by removing essential text segments from the text after analysing variables like term frequency, sentence placement, and many others to pinpoint the sentences that need to be eliminated. Most of the previous works on summarization for monolingual and native English but there is shortage of non-native English and regional language specially Indian languages. Indian Language Summarization (ILSUM) 2022 1 task given the opportunity for the researchers to cope with code-mixing and script mixing diferent multilingual Indian languages.
For the current work, we applied Extractive approach for the ILSUM 2022 task by the team
_ℎ. We tried for all the given tasks which are in English, Hindi and Gujarati
languages and achieved the satisfactory results. A very few teams submitted on all the given
shared tasks [
The paper organized as follows: Section 2 gives a short review on previous work done in the area of summarization. Section 3 discussed the dataset and Section 4 provides the detail about methodology. Section 5 presents the results and finding of our work. Finally, we concluded in Section 6 with highlighted the future scope.
Automatic text summarization is an active area from the Natural Language Processing (NLP)
research community [
various constraints , the maximum number of clusters has been set to 6 for texts containing more than 15 sentences, 5 for texts having sentences greater than 6. Text less than this size has been left as it is as it is perceived to be already summarised.
The main challenge was posed by the fact that any Indian written snippet is multilingual, in this instance, Hindi, English and Gujarati. A lot of datasets are available for english language but not for Hindi and Gujarati, in addition the script is much more complex adding to the dificulties. Each cluster thus can only be composed of words from the same language.
In this section, we presented the dataset used in this work. There are variety of datasets are
available for any single language like German, English, Frech, Spanish and many more languages
but there is no dataset for Indian regional languages. Another problem with those dataset are
they are not public and size of the datasets are also small. Therefore, through the ILSUM 2022
dataset [
In this section, we preset the detailed description of our approach used for summarization tasks.
The approach presented uses K-Means clustering to create extracts from the parent text. The number of clusters depend upon the size of the input text. In broader view though, too less number of clusters are unsuitable as they may change the meaning of the parent text entirely. Similarly large number of clusters would mean that the size of the extracted text is large which contradicts the purpose of summarization. Keeping in mind the various constraints, the maximum number of clusters has been set to 6 for texts containing more than 15 sentences, 5 for texts having sentences greater than 6. Text less than this size has been left as it is as it is perceived to be already summarised.
The main challenge was posed by the fact that any Indian written snippet is multilingual, in this instance, Hindi, English and Gujarati. A lot of datasets are available for English language but not for Hindi and Gujarati, in addition the script is much more complex adding to the dificulties. Each cluster thus can only be composed of words from the same language.
• Splitting and Tokenization • Word Vectorization • Sentence Vectorization • K-means Clustering
For tokenization and training purposes we have used NLTK library2 for splitting the words. We have designed custom tokenizer for all English, Gujarati and Hindi languages. Separate splitters have been used for English, Gujarati and Hindi words as end of the sentence delimeters are diferent in the 3 languages. We have also used sentence tokenizers based on this splitters to tokenize the sentences for further analysis.
After receiving the tokenized words from previous sentences we have trained a Word2Vec model. Each word is converted into a vector which represents its similarity index with other words in the text corpus. This forms basis for distance calculation in cluster formation based upon the trained centers.
Now that we have the word vectors we concatenate them together to form the sentences. Now each sentence is flattened into its constituent words to make it suitable for neural net input. The dimensions of the sentences are reduced using Autoencoders. A total of 5 layers including the flattening layer are used. We first reduce the dimensions of sentence and then increase them again to the original which is the basis for Autoencoder approach for lower dimensional representation.
K-Means Clustering is one of the most popular and simple to implement unsupervised machine learning algorithms. Typically, such algorithms make inferences from an unlabelled dataset as opposed to the supervised algorithms. The objective of K means is to group similar data points together on the basis of underlying patterns. K means creates K clusters in which the data points are grouped. In other words, K means identifies K centroid points, all the other data points are identified to one of the centroids on the basis of least distance. The centroids are not ifxed and they may change over as frequently as every next iteration, this is done to ensure small centroids and hence tighter clusters.
Finally K-means Clusters the cleaned text. The number of clusters is decided by the size of the input text. On an average we have 13 words per sentence so based on that if we have to take 75 words (as provided by the organizer) in final summary, so we have come up with the solution below: • Taken 6 for texts having more than 15 sentences • Taken 5 for texts having more than 6 sentences • Texts having less number of sentences were left unmodified
In this section, we present our results and findings of our approach. Table 2 and Table 3 are showing the results of four diferent ROUGE F1 Scores of validation and testing sets for Hindi, Gujarati and English languages.
Extractive text summarization techniques have been focused upon. K-means further complements the eficiency of extractive techniques as it is quick and suitable for small samples as well as large. Though the performance metric scores were not upto the mark, the method was elegant and hence tweaking the parameters to improve it may prove to be fruitful.
As future scope, the current work can be extended by utilizing the Abstractive approach for the summarization task.
We are thank full to our undergraduate students Shaury Srivastava and Pallav Garg for their contribution in prepossessing steps. A very special thanks to academics and Management of Indian Institute of Information Technology Ranchi for providing the necessary resources.