DifferentApproaches of Marathi Language Word Sense Disambiguation

This paper presents the development issues of a Manipuri word sense disambiguation system. The paper describes the general and linguistic issues as well as suggests a primary plan. Around 650 Manipuri sentences are collected as the database. Conventional positional and context based features are suggested to capture the sense of the words, which have ambiguous and multiple senses. By applying the suggested knowledge acquisition technique, finally the sense of a word can be predicted.

Procedia Computer Science, 2015

Semantic relations play an important role in resolving the ambiguity of a polysemous word. This paper investigates the role of hypernym, hyponym, holonym and meronym relations in Hindi Word Sense Disambiguation. In this work, we have considered five different cases: all relations, hypernym and hyponym, hypernym, hyponym and holonym. No semantic relations are used in the baseline. We obtained an overall improvement of 12.09% in precision over the baseline for the case when all relations are considered. The maximum improvement for single semantic relation was obtained using hyponym which resulted in an overall improvement of 9.86% in precision.

The problem of WSD started sooner attempts were made for machine translation. WSD is hands in gloves with ambiguity. As we have seen both structural and lexical ambiguity cause problem in the interpretation of meaning. This problem multiplies when we go from target langue to source language. There are interlingual as well as intralingual ambiguities. The ambiguity between source language and target langue is known as transfer ambiguity. Before bothering about the transfer ambiguity, we have resolve ambiguity exhibited by the source langue itself in terms of polysemy and homonymy. As we have seen it is the context which comes to our rescue when we try to interpret the meaning of a lexical item or word. Almost all the attempts to resolve homonymy and polysemy by means of WSD make use of context directly or indirectly. We have seen that there are a number of approaches for WSD. Fundamentally they can be grouped into three types: supervised approaches of WSD, unsupervised approaches of WSD and knowledge-based approaches of WSD. We have taken up only two categories for resolving homonymy and polysemy. They are nouns and verbs. As for as nouns are concerned we have taken up homonymous lexical items and for verbs we have taken up polysemous lexical items. For WSD of nouns we have followed cluster approach. Here we capture contexts by means of the process for clustering. The content words which appear along with the ambiguous word have been grouped into two or more clusters depending on the meaning they convey. In one cluster approach we annotate the senses denoted by the ambiguous words that appear in the selected or training corpus. By using the training corpus we train the system so that the system can distinguish the senses of polysemous items when new texts are given to it. This approach has limitation as we have to manually annotate the senses of the words in the training corpus. This system by default is a sample or toy system rather than a generic system. In order to overcome the limitation on the earlier supervised clustering approach we have taken up another clustering approach in which the system itself will group the contexts of the ambiguous lexical items based on similarity measures, i.e. making clusters based on the frequency of contextual lexical items. The number of clusters indicates the number of senses. But the problem is that the clusters are made without being assigned to senses. There could be a possibility that more number of cluster can be made than actually required by us. But this can be used as a generic system and not as a sample system. The success rates achieved by us for both the supervised and unsupervised approaches are definitely encouraging. For resolving ambiguity due to polysemy of verbs, we make use of similarity measures. We make use of neural network approach to meet the problem of polysemy in verbs. Here also the success rates are encouraging. This also can be done as a supervised or unsupervised approach. The use of clustering for Word Sense Disambiguation is relatively a new idea and it has been implemented successfully here for disambiguating homonymous words in a given Malayalam text. Apart from the use of the clustering algorithm, this approach makes use of collocations and case-markers as knowledge sources. The approach can be viewed, on the one hand- as using statistical techniques for the clustering process and; on the other hand- as using information extracted automatically from the corpus, in the form of collocations and case-markers, for disambiguation. The complete algorithm has been implemented and tested with Malayalam corpus. The results obtained are quite encouraging. The highest accuracy of the algorithm is around 90% for two words, out of the eight ambiguous words we have tested. The average accuracy of the algorithm is around 84% (Table 4.2), which fare well with the other knowledge-based approaches to WSD (Lesk, 1986; Walker 1987; Black 1988; Dagan and Itai 1994). It is important to note that we believe the accuracy can be improved to near 95% with further research and improvements such as 1) using a bigger corpus, 2) considering the case-markers that are attached to the ambiguous word (apart from the case-markers attached to the collocation words) for disambiguation. The detailed suggestions are given in the section titled “Future Scope”. The contribution of this research work is twofold. The first is demonstration of the substantial reduction in the labour of human-annotators compared to the earlier approach by Yarowsky (1995). This is achieved mainly because of the clustering algorithm used here instead of the bootstrapping approach of Yarowsky (1995). The clustering process identifies collocations automatically. Thus eliminates the need for giving seed words manually, which is a laborious process. The second important contribution is the use argument structure for WSD of verbs. This approach additionally needs a corpus, POS tagger a morphological analyser and lists of ambiguous and stop words; these are the minimal requirements for any statistical WSD approach. However, the approach does not demand expensive resources such as a dictionary, a thesaurus or a sense-tagged corpus. 9.1. Future Scope As this work is the first to use clustering techniques disambiguation, it points to a good number of possible improvements. The scope of further improvement of this work is given below. Currently, the algorithm uses collocation words. The preliminary analysis in this regard is promising and encouraging. The second major improvement in the approach is to use the “one sense per discourse” property suggested by Yarowsky (1993). This can be used both in the training and testing phase to tag occurrences that could not be tagged by the current algorithm. This property could improve the learning of the algorithm when used in the training phase. This can also increase the performance when used in the testing phase. The performance increment, if fed back to the system can be used for training the algorithm further. This will greatly improve the learning of the algorithm and ultimately results in improved performance. The similarity measures can be improved upon if we have information on argument structure of verbs. The case-markers, which express the functional relationship between the nouns and verbs are bound by the semantic rules of the language and hence require some form of semantic analysis. Consequently, we require verb frame categorisation along with some additional semantic information (to distinguish between the various roles played by a case-marker) to effectively overcome this limitation due to multi-functional marking. Now we have information on word vectors. There open source soft-wares which can give information on word vectors. Such system has embedding system with very useful semantic information extracted form corpus. It can tell us what the contexts are of words where their positions in the given text are; it can also give information on collocation and co-occurrence. By making use of word vectors we can build we effective WSD system.

International journal of engineering research and technology, 2018

Word Sense Disambiguation (WSD) is one of the pressing problems in Natural Language Processing (NLP), which determines the right sense of an ambiguous word in the specific context. WSD is considered as a harder problem as it depends on a set of classes, which vary depending on the context. There have been a number of researches performed in WSD in many languages, from dictionary-based methods to supervised learning methods and unsupervised learning approaches. In this paper, an attempt is made for the disambiguation of Malayalam words, which has a rich set of ambiguous words having different meanings. A semi supervised machine learning techniques combined with statistical approach mainly Maximum Entropy is experimented for Malayalam, which shows promising result for a set of corpus trained.

Regular, 2020

Telugu (తెలుగు) is one of the Dravidian languages which are morphologically rich. As within the other languages, it too consists of ambiguous words/phrases which have one-of-a-kind meanings in special contexts. Such words are referred as polysemous words i.e. words having a couple of experiences. A Knowledge based approach is proposed for disambiguating Telugu polysemous phrases using the computational linguistics tool, IndoWordNet. The task of WSD (Word sense disambiguation) requires finding out the similarity among the target phrase and the nearby phrase. In this approach, the similarity is calculated either by means of locating out the range of similar phrases (intersection) between the glosses (definition) of the target and nearby words or by way of finding out the exact occurrence of the nearby phrase's sense in the hierarchy (hypernyms/hyponyms) of the target phrase's senses. The above parameters are changed by using the intersection use of not simplest the glosses but...