Word sense disambiguation for Turkish

As being one of the pervasive characteristics of the natural languages, the research on word sense ambiguity aims at resolving the problem of having more than one sense. A Word Sense Disambiguation (WSD) task is defined as automatic assignment of the most appropriate meaning to a polysemous word within a given context. The problem of word sense ambiguity, which can be resolved during human communication by using previous experiences and complex cognitive processes, is also one of the active topics in computer science and Natural Language Processing (NLP) area. The identification of word meanings is required in almost all applications of the NLP area to provide them proper functioning. These applications include the areas such as Information Retrieval (IR), Machine Translation (MT), Semantic Annotation (SE), Question Answering (QA) and many others. From this aspect, WSD is an important intermediate step for all these applications that increases their performances. There is a huge need in NLP related fields and internet environment for the development of Computational Linguistics (CL) methods. As a result, several algorithms have been developed for the different fields of the NLP area. In the scope of these works, the properties such as nature of the languages, available resources and constraints, application requirements play important role to develop methods. The WSD methods are classified under three broad categories: knowledge-based, corpus-based and hybrid methods. The family of knowledge-based methods primarily relies on dictionaries, thesauri, ontologies and lexical knowledge bases. Corpus-based methods are further classified into supervised, unsupervised and semi-supervised methods (or minimally supervised). Previous efforts on WSD have mainly focused on supervised approaches that require sense annotated corpora. There are also alternative approaches of unsupervised and semi-supervised methods that try to lower the sense-annotated portion of the texts. Although sense ambiguity can be resolved in supervised systems with high accuracy, usage of semi-supervised and unsupervised methods has gained attention recently since the sense annotation scheme is labor intensive and expensive. In some of the studies, word senses are extracted from corpus itself where sense-annotated corpora are insufficient or not used. Recently, the extraction of word senses from corpus is preferred by the researchers since the pre-defined sense definitions of dictionaries may be too limited. On the other hand, the adaptation of solutions and methods to new domains may be difficult because of the dynamic nature of word senses. In the scope of hybrid methods, knowledge-based and corpus-based methods are combined. WSD can also be classified according to the scope of approach to the problem. The level of supervision is the first criterion to classify the methods. A secondary classification for generic WSD can be made by considering two variants. These options of WSD can be selected from: 10 Lexical Sample (LS) task and, 20 All-words (AW) task. The former approach disambiguates the occurrences of a small sample of target word that has been determined previously. Since the words and the set of senses are limited, supervised Machine Learning (ML) methods are usually used to handle LS tasks. Hand-labeled examples are used to train the classifier. Then unlabeled test portion of the target words can be labeled by using trained classifier. In contrast, AW approach comprises the disambiguating all the words in a running text. All the entries in a given system are required to be disambiguated. There is a similarity between AW task and Part of Speech (POS) tagging. The only difference is that the former needs much larger set of tags. This larger set of tags resulted in data sparseness problem since it is hard to find adequate training data for each word. Knowledge is the fundamental component for a WSD system which can be acquired from dictionaries or learned from a training corpus. The sources can be classified into "lexical knowledge" and "learned world knowledge" categories. The lexical knowledge category includes the knowledge sources such as "sense frequency", "concept trees", "selectional restrictions", "subject code" and the POS information. The latter category includes the usage of "Indicative words", "syntactic features", "domain specific knowledge" and "parallel corpora". It is usually observed that the unsupervised systems need lexical knowledge sources while supervised systems use world knowledge. But in practice the combinations of these sources have been used in WSD systems. ML techniques are used to automatically acquire disambiguation knowledge in the scope of corpus-based WSD methods. A typical WSD system may utilize sense-tagged corpora, online dictionaries and large scale linguistic resources as components. The set of features to be used and the learning algorithm are two of the important decisions that have to be considered for the design of a WSD system. Many NLP systems rely on linguistic knowledge acquired from hand-labeled training text data and ML methods. The supervised methods of the WSD can be classified according to the induction principle they use to acquire model or rules. These methods consist of probabilistic models (e.g., Naïve Bayes), similarity based methods (e.g., k-Nearest Neighbor algorithm), methods based on discursive properties (e.g., one sense per discourse/collocation, attribute redundancy), methods of discriminative rules (e.g., decision lists, decision trees or methods based on rule combination), linear classifiers and Kernel-based methods. Compared to the other subjects in NLP such as POS determination and syntax parsing, a WSD problem introduces extra difficulties. Since each word is associated with unique meaning, complete training set requires a huge number of examples. This language sparsity problem is dealt with by selecting features used in training algorithms. These features can be found in local or wider context. Before applying the ML algorithm, all the examples of a particular ambiguous word have to be encoded in a way the learning algorithm can handle. Collocational and Bag-of-Words (BoW) features are two important classes of features that are generally extracted from neighboring contexts in WSD tasks. Almost all of these approaches are employed by defining a window of "n" content words around the word to be disambiguated in the corpus. Collocational features encode information about the lexical inhabitants of specific positions located to the left or right of the target word. The basic elements may consist of the word, its root form and the part of speech information. BoW is the second feature set in which the text is treated as an unordered bag of words. Within this approach, similarity measures are calculated by looking at the semantic similarity between all the words in the window regardless of their positions. Considering the limited resources available for natural languages, it is especially important to select knowledge sources and the feature sets carefully to disambiguate senses. Overall results of this study can be summarized as follows: • Turkish Lexical Sample Corpus: In the scope of this study, a special corpus for Turkish has been prepared. For this task, the Turkish nouns and verbs have been determined by considering highly ambiguous ones among the dictionary of Turkish Language Association (TLA). Then samples have been collected for each ambiguous candidate word. Voters annotated the samples by using the sense definitions of the TLA dictionary. • Extracting Effective Features: Two approaches have been tested to extract effective features: 10 Collocational Features (CF), 20 Bag-of-Words. • Supervised methods have been used to extract effective features on disambiguating word senses. The effective features of neighbor words around ambiguous headword, have been determined. The contribution of these features on disambiguating word senses has been investigated. A fixed window size has been used along the experiments. In the scope of the study, collocational features which comprise the morphological analysis outputs of the word roots and suffixes have been investigated. • The contribution of using BoW features has also been investigated. The proper size for selected features has been determined. This work has been conducted for Turkish noun and verb sets. The optimal extent around headword to encode BoW features is determined. • The results of collocational and BoW features have been compared. It is shown that the collocational features are more effective than BoW features on resolving sense ambiguities. Better results are achieved by combining two feature sets. • Our research on supervised methods shows that a comprehensive and very large corpus that represents the language effectively is needed to be able to continue conducting research on WSD. There is no such a large corpus in Turkish. We focused our research on unsupervised methods since it is too labor-intensive to prepare such a corpus. • A graph-based unsupervised method which previously used in a few NLP related studies have been developed for Turkish. At the initial phase of the study, a program has been developed to implement the algorithm. Then the effect of supervised method findings is investigated to enhance the results. The Turkish lexical sample corpus has been used to generate graph and evaluate the accuracy results. • Our research show that sense ambiguities can be resolved by using unsupervised methods. We propose the gold standard evolution at the final stage and use annotated word senses of Turkish lexical sample corpus to map hub meanings. This method yielded nearly as reliable results with the supervised methods.