- August 2021
Summary: Word Sense Disambiguation (WSD) is a challenging AI problem. The solution to which is evolving with primarily supervised learning techniques. However, most of the unsupervised methods still need more reductionism while supervised techniques predominately require huge tagged data for a higher accuracy. This paper presents an unsupervised novel technique for WSD. The key contribution of this paper is firstly to introduce the application of technique of unsupervised Fuzzy Rough Set based Span for the problem of Word Sense Disambiguation. Secondly, to evaluate the efficiency of the proposed techniques for standard WSD evaluation tasks. The fuzziness comes given that the representation of the target word is a Fuzzy Set, hence, hybrid model of Fuzzy Rough Sets has been proposed while using the intuition of Span for Rough Sets computation for WSD. The proposed technique is novel in its application and elaboration for use. The results are motivating given that these are the initial applications of Span and are motivating when compared with unsupervised tasks in WSD competitions of the evaluated datasets. As in any other application the accuracy and efficiency are open problems which has been analyzed for further improvements.
Word Sense Disambiguation (WSD) (Edmonds et al, 2001) is one of the essential tasks of Natural Language Understanding. It has a variety of applications in various tasks in Artificial Intelligence such as machine translation, information retrieval, speech recognition, chat bots to mention a few. WSD has been designated as an AI-complete problem (M`arquez et al, 2000, Navigli, 2009) in which finding its solution can lead to finding solution to the General Artificial Intelligence Problem, which involve common-sense reasoning and intelligence.
The state-of-art works at present in the area of WSD are now primarily dominated by supervised learning algorithms. The major trend in recent research papers is of deep learning-based techniques for WSD both from supervised and semi-supervised ways. The supervised approaches provide for better precision, recall and accuracies over baselines. However, supervised techniques for this particular problem of Word Sense Disambiguation are not like typical classification problem. Here, each word to be disambiguated has several senses and the tagged data for each sense of the word is often not in substantial amount as required in complex learning tasks. WSD algorithms can be classified as knowledge-based systems, supervised algorithms and unsupervised algorithms. Supervised algorithms require the target word sense being annotated in training data while evaluations are performed on testing data. These methods use best of the algorithms. However, they suffer from the fact that not much tagged data is available, especially with growing multilingual use of contexts and new slangs being added to dictionary often. Unsupervised techniques do not require target class tagging however to deal with sense disambiguation, external database or a knowledge base is required. Knowledge-based systems can themselves be unsupervised or supervised in nature. However, the drawback of supervised techniques has been mentioned above, apart from those these algorithms require huge training time and fine tuning of parameters, especially in terms of deep learning techniques. And the key aim to foster unsupervised approach is this is a gradual process to solve the problems, by putting measures of improvements in unsupervised learning for normal to difficult tasks.
Rough Sets (Pawlak, 1982, Pawlak, 2012) on the other hand deal with uncertainties that occur in knowledge-based systems. This is the motivation behind this paper to consider Rough Sets for solving the problem of Word Sense Disambiguation. The key motivation of the paper is firstly, to build an unsupervised analytical and experimental model of WSD using a new concept of Rough Set based Span. However Rough Set based Span cannot be applied directly to task of WSD for reasons as explained in detail in Section 3. The key reason is the word to be disambiguated is itself a Fuzzy Set made from its gloss definition from WordNet or in other ways, as discussed in subsequent sections. The explanation for fuzziness of target word to be disambiguated is that its membership varies for all words of the Universe. Hence, modelling of problem had been performed using hybrid model of Fuzzy Rough Sets based Span (Yadav, 2021). Secondly, the motivation of paper is also to explain the applications of Span of Rough Sets to NLP task of WSD. The concept of Span (Yadav et al, 2019) was proposed to measure the covering (Spanning) capabilities of a subset of universe. Rightly choosing the subset and the universe, the Spanning or covering capabilities of a word sense in a context can be computed given the knowledge about the sense. Thirdly, the proposed model is analysed for performance of the basic model and the computational performance have been computed on several standard WSD task datasets including SENSEVAL-2 (Edmonds et al, 2001), SENSEVAL-3 (Snyder et al, 2004), SEMEVAL-2007 (Pradhan et al., 2007), SEMEVAL-2013 (Navigli et al., 2013) and SEMEVAL-2015 (Moro et al, 2015).
The paper is organised as follows. Section 2 briefly discuss main previous research work related to this paper and evaluations. The proposed techniques are discussed in Section 3. Section 4 presents the results on standard datasets for WSD and Section 5 concludes the paper with Future Works.
2. Previous Work
Rough Set theory is a key tool to deal with uncertainty present data especially in knowledge based systems. Rough Sets require an equivalence relation R. And the properties of a subset X of universe U can be analysed by the theory, which relies on two key subsets of universe, namely the lower approximation and the upper approximation. The lower approximation is the set of all elements of the universe which for sure belong to the set X while the upper approximation consists of those elements of universe which may probably belong to the set X. Boundary region, consists of those elements of universe which belong to upper approximations but not to the lower approximations. Rough Set theory based uncertainty handling have several applications and have been used often in case of uncertainty handling. Some of the application areas of previous works of Rough Sets are in financial data analysis (Tay, 2002), feature selection (Chouchoulas et al, 2001, Jensen et al, 2004, Jensen et al, 2007, Wang et al, 2007, Yadav et al, 2017), classification (Jensen et al, 2011, Yadav et al, 2020), information retrieval (Das-Gupta, 1988, Yadav et al, 2018) and text summarization (Yadav et al, 2019, Yadav et al, 2020, Yadav, 2021), sentence similarity (Yadav et al., 2019) to mention a few.
Rough Set based Span have been recently proposed for computing the Spanning capabilities of a subset of a universe. Spanning capabilities means given a subset what is the percentage of universe that this subset covers. Recent works on Rough Set based Span had been in applications in unsupervised Text Summarization, decision making using the concept of Span and the second definition of Span. Here, the concepts of Span of a Rough Set. Let X be a subset of universe U, the Span intuitively determines the Spanning or covering capabilities of the subset X. Span is formally defined as (Yadav, 2021).
The definition makes computations much straight forward, howsoever, the intuition behind these computations have been derived from definition 1, definition 2 and the past works in this area. The paper utilizes the computation of Span using this definition. This definition does not directly reflect the Spanning capabilities and the theory but its derivation (Yadav, 2021) leads to a considerable explanation to the concept.
While Word Sense Disambiguation (WSD) is the task wherein a word in a sentence or any other context is provided and its correct sense has to be automatically determined by the algorithm. Several techniques both supervised, unsupervised and knowledge based have been proposed in literature and many evaluation datasets have been constructed, some of them have been evaluated in this paper for proposed technique. One of the most famous technique in WSD is by Lesk (Lesk, 1986) which is an unsupervised knowledge-based technique. It works on intersection of gloss of context and gloss of the target word to be disambiguated. Other unsupervised algorithm for WSD task are discussed here.
WordNet based co-occurrence of target word and each word in the context was proposed (Seo et al, 2004) as an unsupervised WSD task evaluated on SENSEVAL2 apart from Korean language dataset and noted a recall measure of 0.4548. Moldovanet al (2004) provided techniques using to disambiguate WordNet glosses. Yoon et al (2007) proposed unsupervised knowledge-based WSD for Korean language using acyclic weighted digraphs. The method uses co-occurrence of word pairs and computation of similarity matrix from it, generating vector representation of words and applying graph-based processing. The results of system on English words from SENSEVAL2 dataset were reported to be 0.3070.
McInnes et al (2014) provided in their research paper the difficulties posed in medical data and the constraints of medical lexicons for automatic WSD tasks for processing medical documents. The corelation between the difficulty measures of computation and results of accuracies reached were presented in the work. The paper also presented various supervised and unsupervised techniques for medical data WSD tasks.
Augat et al. developed a NLTK computational package for WDS using Lexical chains which reaches results of 0.7856 on Brown Corpus of Semcor. Lexical chains are knowledge-based systems which creates chains of synsets and words related to a synset is added in the relevant chain. The largest chain is typically taken as the best representative of the text fragment. Various combinations have been analysed for results. Unsupervised knowledge-based technique of WSD using Lexical Chains were elaborated by Nelken et al (2007). The results reported by the authors on SENSEVAL3 task measured as F1-Scores varied to 0.7020 for models combined with HMM. Improvements in weight computations and forming of lexical chains were presented in work by Chen et al (2009). Fernandez-Amoros et al (2011) proposed WSD algorithm using conceptual relations among words and evaluation performed on SENSEVAL2 dataset. The best recall score reached by some of the experiments in paper was .4700.
Fernandez-Amoros et al (2011) proposed WSD algorithm using conceptual relations among words and evaluation performed on SENSEVAL2 dataset. The best recall score reached by some of the experiments in paper was .4700.
Unsupervised Graph based method was proposed by (Sinha and Mihalcea, 2007) and evaluated on SENSEVAL2 and SENSEVAL3 datasets. The research proposes graph centrality-based technique including indegree and PageRank for WSD task and is dependent on the computation of word similarity techniques used. A recall score of 0.5230 was noted in paper for all word problems, while higher scores were evident for all noun, and other POS tags. The algorithms have been evaluated for various kinds of similarity measure and various graph-based techniques which are indegree, betweenness, closeness and PageRank. The best system developed were combination of similarity and voting schemes and evaluations were performed on this system for test set with a recall of 0.5637 for all word WSD task. An extended graph-based technique (KĊdzia et al, 2014) has been proposed for graph-based unsupervised knowledge-based algorithm. The use of PageRank in WSD has been enhanced by authors. The algorithm had been also evaluated using Polish WordNet and reached a precision of 0.43 on language specific dataset viz. part of the KPWr corpus of Polish, having 60 sense disambiguated words and 1996 documents.
Yaun et al (2016) proposed a semi-supervised deep learning technique using LSTM for WSD and performed evaluations on SENSEVAL-2, SENSEVAL-3, SEMEVAL-2007, SEMEVAL-2013 datasets reaching an F1-score of 0.8550 from among all results. However, results specific to a particular dataset varied. (Lacobacci et al, 2016) worked on word embeddings for WSD task. Various combinations were experimentally analysed for performance. Some of the best F1 scores are 0.7770, 0.7410 and 0.7150 for SENSEVAL-2, SENSEVAL-3 and SEMEVAL-2007 respectively.
Deep Learning techniques of bi-directional LSTM was analysed by Kageback et al (2016) to obtain benchmark results in supervised learning tasks evaluated on SENSEVAL-2 and SENSEVAL-3 datasets. Bidirectional LSTM sequence labelling architecture for WSD (Raganato et al, 2017) utilized deep learning-based techniques. The F1-score reached best score of 0.7010 on all word task of datasets SENSEVAL2, SENSEVAL3, SEMEVAL2007, SEMEVAL2013 and SEMEVAL2015.
Various competitions had been conducted for WSD task. And most participants in these WDS tasks had been algorithms using supervised learning tasks. Not many of the unsupervised techniques produced competition results in the benchmarks. Hence, it becomes even more important to study, analyse and evaluate the unsupervised knowledge-based technique of WSD using Fuzzy Rough Set based Span, as proposed in this paper. The next Section describe the proposed Span technique in detail.
3. Proposed Method
Consider a word w in sentence S, where the word w has senses s1, s2,…, sk. And let the sentence (or context) S, where the word w has to be disambiguated has words v1,…,vn, which are all primarily non-stop words. The task of WSD can be explained as determining sense si such that si is the best fit to the meaning of word w in sentence S. The algorithm proposed in this paper is solely unsupervised in nature and has used WordNet (Miller et al., 1990) as knowledge base to gain information about the contexts of the senses.
The proposed technique is based on concept of Rough Set based Span. However, for using the knowledge-based system of WordNet the hybrid model of Fuzzy Rough Sets fit well. Fig 1 presents the algorithm to perform Word Sense Disambiguation using Fuzzy Rough Set based Span. The definition 3 of Fuzzy Rough Set based Span is used in the proposed algorithm for efficiently computing the Span.
Algorithm: Span based Unsupervised Word Sense Disambiguation
Input: Target_Word w , Sentence/Context S=[v1, v2,……vn]
Output: Predicted sense of w
- Evaluate the senses of the word w, say [s1, s2,….sk]
- For each sense si in [s1, s2,….sk] do
- Find the lemma and gloss corresponding to the synset_id si of sensei from WordNet API
- Compute glosses of hypernyms and hyponyms of synset si upto level b for the word synset. b chosen as 2 here, for elaboration.
- Compute bag of words from these 2*b lists
- Sense_i_expanded <- Remove stop words and non essentials from this bag of words
- Compute the SPAN of sense_i_expanded with the given context which is the sentence given in input
- Save the SPAN as Span_i[i]
- Sort the array Span[i] and select the sense that has maximum Span
- Return sense[max_Span_index]
The algorithm starts with an input target word whose sense has to be evaluated and the context in which it appears, this is ideally a sentence in which the target word is present, is provided. All senses of the target word present in the WordNet Lexicon are determined and stored in an array sense. For each sense the Span of the sense gloss is determined with the context given in input. There are various ways to compute the word sense expanded definition, simply as the definition of the synset to be considered or as a combination of glosses when hypernyms and hyponyms up to a certain depth are considered. Both these ways have been evaluated.
Once the gloss has been computed by either way, the key words from the text of list of glosses are extracted. The set X here is the set of essential words from these words for sense_i of the target word. The universe is the set of all words under consideration, those in given sentence(context) and those in this list of gloss or the expanded gloss, whichever is the case.
The proposed algorithm returns the senses which have maximum values of Span. There may be cases through less frequent wherein the maximum value of Span may not be unique. Hence, the output of system can be considered as multiclass, even though these instances occurred with lesser probability in dataset.
The proposed algorithm as given in Fig 1 relies primarily on Span computation. And as per definition of Span as given in definition 2 and definition 3, the values of Span depend on the Fuzzy Rough Set computations. The definition of Fuzzy Rough Set is highly depended on the techniques used to compute the membership of a word in context to the Fuzzy Set which makes the target word. Further, it also depends on similarity of words in the context (or sentence, as the case may be). Hence, here the technique to compute both the Span and Fuzzy Rough Sets for a data instance in standard WDS datasets are discussed in details. Two techniques have been evaluated in this paper. Reasons to look over traditional way of computing Fuzzy Rough Sets have been provided below before introducing the second approach.
The definition of Span as per definition 2 and definition 3 require computation of lower and upper approximations. This requires a deeper analysis in the definitions from point of view of application area which is Word Sense Disambiguation. The WSD standard datasets considered are for task of all-word disambiguation. While there are problems which disambiguate some words at a time only. This problem with the current application can be considered as a sub problem of the all-word problem. The dataset snippets are described in Section 3. Consider one instance of the data to illustrate the problems and techniques to handle in the baseline developed. One instance of data consists of sentence, a target word and the target sense of the word as provided by expert annotators.
Considering the target word which is required to be disambiguated as a Fuzzy Set ϕ. The context which here is the sentence is considered as the universe under consideration, as this is the context which shall resolve the sense of a word in the sentence itself. Why target word is considered as a Fuzzy Set? The reason is that the membership of words of a sentence vary for each word of the Universe, which is the context here. This makes the target word a Fuzzy Set. Hence the use of Fuzzy Rough Set based Span is well suited for the problem. Word Sense Disambiguation is the problem where one disambiguates a word given the context, which is the Universe under consideration. In these terminologies the lower and upper approximation can be explained as follows. The lower approximation (Pawlak, 1982) intuitively defines the elements of universe which for sure belong to a set X.
- Augat, M., & Ladlow, M. (2004). CS65: An NLTK Package for Lexical-Chain Based Word Sense Disambiguation. Word J. Int. Linguist. Assoc.
- Chen, J., Liu, J., Yu, W., & Wu, P. (2009, December). Combining lexical stability and improved lexical chain for unsupervised word sense disambiguation. In 2009 Second International Symposium on Knowledge Acquisition and Modeling (Vol. 1, pp. 430-433). IEEE.
- Chen, P., & Liu, S. (2008, October). Rough set-based SVM classifier for text categorization. In 2008 Fourth International Conference on Natural Computation, 2, pp. 153-157.
- Chouchoulas, A., & Shen, Q. (2001). Rough set-aided keyword reduction for text categorization. Applied Artificial Intelligence, 15(9), pp. 843-873.
- Dan Meng, Xiaohong Zhang, Keyun Qinc , Soft rough fuzzy sets and soft fuzzy rough sets, Computers and Mathematics with Applications, (https://dx.doi.org/10.1016/j.camwa.2011.10.049)
- Das-Gupta, P. (1988, May). Rough sets and information retrieval. In Proceedings of the 11th annual international ACM SIGIR conference on Research and development in information retrieval (pp. 567-581). ACM.
- Dubois, D., & Prade, H. (1991). Fuzzy Sets in approximate reasoning, Part 1: Inference with possibility distributions. Fuzzy Sets and systems, 40(1), pp. 143-202.
- Edmonds, P., & Cotton, S. (2001, July). Senseval-2: overview. In Proceedings of SENSEVAL-2 Second International Workshop on Evaluating Word Sense Disambiguation Systems (pp. 1-5).
- Eberhart, R., & Kennedy, J. (1995). A New Optimizer Using Particle Swarm Theory. In MHS’95. Proceedings of The Sixth International Symposium on Micro Machine and Human Science IEEE. (pp. 39-43).
- Fernandez-Amoros, D., & Heradio, R. (2011). Understanding the role of conceptual relations in Word Sense Disambiguation. Expert Systems with Applications, 38(8), 9506-9516.
- Foitong, S., Srinil, P., & Pinngern, O. (2008, October). Rough Sets Based Approach to Reduct Approximation: RSBARA. In 2008 Fifth International Conference on Fuzzy Systems and Knowledge Discovery 2, pp. 250-254.
- Grzymala-Busse, J. W., & Ziarko, W. (2003). Data Mining Based on Rough Sets. In Data Mining: Opportunities and Challenges. IGI Global, 2, pp. 142-173.
- Hassanien, A. E., & Ali, J. M. (2004). Rough set approach for generation of classification rules of breast cancer data. Informatica, 15(1), pp. 23-38.
- Hu, Q., Zhang, L., Chen, D., Pedrycz, W., & Yu, D. (2010). Gaussian kernel based fuzzy rough sets: model, uncertainty measures and applications. International Journal of Approximate Reasoning, 51(4), 453-471.
- Hu, Q., Yu, D., Liu, J., & Wu, C. (2008). Neighborhood rough set based heterogeneous feature subset selection. Information sciences, 178(18), pp. 3577-3594.
- Laorden, C., Santos, I., Sanz, B., Alvarez, G., & Bringas, P. G. (2012). Word sense disambiguation for spam filtering. Electronic Commerce Research and Applications, 11(3), 290-298.
- Iacobacci, I., Pilehvar, M. T., & Navigli, R. (2016, August). Embeddings for word sense disambiguation: An evaluation study. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) (pp. 897-907).
- Jensen, R., & Shen, Q. (2004). Fuzzy–rough attribute reduction with application to web categorization. Fuzzy Sets and systems, 141(3), pp. 469-485.
- Jensen, R., & Shen, Q. (2007). Fuzzy-rough sets assisted attribute selection. IEEE Transactions on fuzzy systems, 15(1), pp. 73-89.
- Jensen, R., & Cornelis, C. (2011). Fuzzy-rough nearest neighbour classification. In Transactions on rough sets XIII(pp. 56-72). Springer, Berlin, Heidelberg.
- Jensen, R., Parthaláin, N. M., & Shen, Q. (2014). Tutorial: Fuzzy-Rough data mining (using the Weka data mining suite).
- Kędzia, P., Piasecki, M., Kocoń, J., & Indyka-Piasecka, A. (2014). Distributionally extended network-based word sense disambiguation in semantic clustering of Polish texts. IERI Procedia, 10, 38-44.
- Kågebäck, M., & Salomonsson, H. (2016). Word sense disambiguation using a bidirectional lstm. arXiv preprint arXiv:1606.03568.
- Lingras, P., & Peters, G. (2012). Applying rough set concepts to clustering. In Rough Sets: Selected Methods and Applications in Management and Engineering (pp. 23-37). Springer, London.
- Marelli, M., Bentivogli, L., Baroni, M., Bernardi, R., Menini, S., & Zamparelli, R. (2014). Semeval-2014 task 1: Evaluation of compositional distributional semantic models on full sentences through semantic relatedness and textual entailment. In Proceedings of the 8th international workshop on semantic evaluation (SemEval 2014) (pp. 1-8).
- McInnes, B. T., & Stevenson, M. (2014). Determining the difficulty of word sense disambiguation. Journal of biomedical informatics, 47, 83-90.
- Mihalcea, R., & Edmonds, P. (2004). Senseval-3: Third International Workshop on the Evaluation of Systems for the Semantic Analysis of Text. In Association for Computational Linguistics. ACL (Vol. 4).
- Moldovan, D., & Novischi, A. (2004). Word sense disambiguation of WordNet glosses. Computer Speech & Language, 18(3), 301-317.
- Moro, A., & Navigli, R. (2015, June). Semeval-2015 task 13: Multilingual all-words sense disambiguation and entity linking. In Proceedings of the 9th international workshop on semantic evaluation (SemEval 2015) (pp. 288-297).
- Mihalcea, R., Chklovski, T., & Kilgarriff, A. (2004, July). The Senseval-3 English lexical sample task. In Proceedings of SENSEVAL-3, the third international workshop on the evaluation of systems for the semantic analysis of text (pp. 25-28).
- Namburu, A., kumar Samay, S., & Edara, S. R. (2017). Soft fuzzy rough set-based MR brain image segmentation. Applied Soft Computing, 54, 456-466.
- Navigli, R. (2009). Word sense disambiguation: A survey. ACM computing surveys (CSUR), 41(2), 1-69.
- Navigli, R., Jurgens, D., & Vannella, D. (2013, June). Semeval-2013 task 12: Multilingual word sense disambiguation. In Second Joint Conference on Lexical and Computational Semantics (* SEM), Volume 2: Proceedings of the Seventh International Workshop on Semantic Evaluation (SemEval 2013) (pp. 222-231).
- Nelken, R., & Shieber, S. M. (2007). Lexical chaining and word-sense-disambiguation.
- Pawlak, Z. (1982). Rough sets. International journal of computer & information sciences, 11(5), pp. 341- 356.
- Pawlak, Z. (2012). Rough sets: Theoretical aspects of reasoning about data (Vol. 9). Springer Science & Business Medi.
- Pradhan, S., Loper, E., Dligach, D., & Palmer, M. (2007, June). Semeval-2007 task-17: English lexical sample, srl and all words. In Proceedings of the fourth international workshop on semantic evaluations (SemEval-2007) (pp. 87-92).
- Raganato, A., Bovi, C. D., & Navigli, R. (2017, September). Neural sequence learning models for word sense disambiguation. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing (pp. 1156-1167).
- Seo, H. C., Chung, H., Rim, H. C., Myaeng, S. H., & Kim, S. H. (2004). Unsupervised word sense disambiguation using WordNet relatives. Computer Speech & Language, 18(3), 253-273.
- Singh, S., & Dey, L. (2005). A rough-fuzzy document grading system for customized text information retrieval. Information processing & management, 41(2), pp. 195-216.
- Sinha, R., & Mihalcea, R. (2007, September). Unsupervised graph-basedword sense disambiguation using measures of word semantic similarity. In International conference on semantic computing (ICSC 2007) (pp. 363-369). IEEE.
- Srinivasan, P., Ruiz, M. E., Kraft, D. H., & Chen, J. (2001). Vocabulary mining for information retrieval: rough sets and Fuzzy Sets. Information Processing & Management, 37(1), pp. 15-38.
- Tay, F. E., & Shen, L. (2002). Economic and financial prediction using rough sets model. European Journal of Operational Research, 141(3), pp. 641-659.
- Wang, X., Yang, J., Teng, X., Xia, W., & Jensen, R. (2007). Feature selection based on rough sets and particle swarm optimization. Pattern recognition letters, 28(4), pp. 459-471.
- Wu, Z., & Palmer, M. (1994, June). Verbs semantics and lexical selection. In Proceedings of the 32nd annual meeting on Association for Computational Linguistics (pp. 133-138). Association for Computational Linguistics.
- Yadav, N., & Chatterjee, N. (2014). Text Summarization using Rough Sets. In International Conference on Natural Language Processing and Cognitive Processing.
- Yadav, N., & Chatterjee, N. (2017, July). A novel approach for feature selection using Rough Sets. In 2017 International Conference on Computer, Communications and Electronics (Comptelix) (pp. 195-199). IEEE.
- Yadav, N., & Chatterjee, N. (2018). Fuzzy rough set based technique for user specific information retrieval: a case study on Wikipedia data. International Journal of Rough Sets and Data Analysis (IJRSDA), 5(4), 32-47.
- Yadav, N. and Chatterjee, N. (2019). Rough Set based Span and its Application to Extractive Text Summarization. Journal of Intelligent and Fuzzy Systems, IOS Press.
- Yadav, N., & Chatterjee, N. (2020). Rough Set based Aggregate Rank Measure & its Application to Supervised Multi Document Summarization. arXiv preprint arXiv:2002.03259.
- Yadav, N. (2021). Neighborhood Rough Set based Multi-document Summarization. arXiv preprint arXiv:2106.07338.
- Yadav, N. (2021). Decision Making Using Rough Set based Spanning Sets for a Decision System. arXiv:2107.12477.
- Yadav, N. (2021). Novel Span Measure, Spanning Sets and Applications. arXiv:2107.12178.
- Yadav, N. (2021). Computing Fuzzy Rough Set based Similarities with Fuzzy Inference and Its Application to Sentence Similarity Computations. arXiv:2107.01170.
- Yoon, Y., Seon, C. N., Lee, S., & Seo, J. (2006). Unsupervised word sense disambiguation for Korean through the acyclic weighted digraph using corpus and dictionary. Information processing & management, 42(3), 710-722.