Inspecting sub tasks of aspect based sentiment analysis in Turkish language–opinion target expression, aspect category and sentiment polarity detection

Sentiment analysis is one of the major trend research topics in natural language analysis lately. Social media and Web 2.0 has become so popular among people to share their opinions within whole society. This popularity has given rise to the need of sentiment analysis for especially commercial organizations. bBecause commercial organizations companies always wish to measure their reputation in their customers' eyes. Sentiment analysis is the key solution for measuring reputation and generating productive reports over customers' voice. Researches show that this kind of analysis needs natural language processing tools and experimental data. In this research we try to handle sentiment analysis in aspect level which is called aspect based sentiment analysis in the literature. This level of sentiment analysis is the most detailed one. Sentiment analysis problem is handled at different levels in scientific researches. These levels can be listed as document level, sentence level and aspect level sentiment analysis. We define aspect based sentiment analysis with structures called sentiment tuples. This definition comes from Aspect Based Sentiment Analysis (ABSA) shared task of Semantic Evaluation (Semeval) workshop. A sentiment tuple consists of Aspect Category, Opinion Target Expression (OTE) and Sentiment polarity sections. These sections show the opinion over an aspect in the given text. We use Semeval ABSA Turkish Restaurant dataset as the training and test set in our research. Turkish is our target language and we have tried to show Turkish natural language analysis' contribution over aspect based sentiment analysis. Turkish is an agglutinative and free constituent order language. These features make Turkish language different from other languages that most NLP researches have been conducted on, such as English, Spanish, French, German, etc. In Turkish language, due to its agglutinativeness many forms of same words may be seen in texts. This variety leads to problem for sentiment analysis. The great majority of sentiment analysis systems are data driven systems that usually use machine learning techniques. Machine learning algorithms may be applied on textual data using words as features. But different surface forms of same stems are represented as different features and causes feature sparsity. On the other hand, inter word relations are important for sentiment analysis systems. Free word order property complicates to find relations between words in sentence. To overcome this difficulty, syntactic analysis or dependency parsing becomes the key idea. Using dependency parsing, we can extract the words related to opinion target expression and take them into account during sentiment polarity classification. In order to detect opinion target expression in sentiment tuples, we use sequence labelling techniques. We considered the problem as a sematic role labeling problem. In semantic role labelling, finding whole parts of action (actor, action, theme) is aimed by systems, however only theme (object) of the expression is concerned in our problem. Conditional Random Fields (CRF) is the most common algorithm for semantic role labelling problems. OTEPOS (Part of Speech) tagging and named entity recognition are important examples of semantic role labelling. OTEs are similar to named entities. Because these terms exists in texts to represent an aspect of the examined domain. s Terms are strongly dependent to working domain. For this reason, most of the studies try to solve the problem with using lexicon solutions. Morphological richness of Turkish language becomes an important obstacle to generate a specific lexicon. Also previous research shows that lexicon approaches has lack of success. In sequence labelling method, we used features such as, surface forms, stems, POS tag, morphological features, word vectors (word embeddings) and dependency relations between words. Our experiments were designed to prove contribution of word embeddings and dependency relations over OTE detection. Another noticeable contribution comes from labelling types. Our experiments consist of 3 kind of labeling; 0-1, IOB standards and raw aspect category names. In 0-1 labelling type, the word is labelled as "1" if it is in OTE and "0" vice versa. IOB standard comes from named entity recognition and POS tagging tasks. "B" label means current word is beginning of an OTE, "I" means current word is in OTE and "O" label means not in OTE. In this form of labelling "I" and "B" labels followed with aspect category names. For instance "B-RESTAURANT#GENERAL". Latest type which is used in experiments is raw aspect category names. For example: "RESTAURANT#GENERAL". Most successful scenario for labelling has become IOB standards. Word vectors and dependency relations outperformed the other setups with a score of %53.12. We deduced that word vectors have noticeable contribution in sequence labelling because training set is lack of word features. Using word vectors is extending the feature set that is used in CRF. Using dependency relations provides better scores compared to using order of words in sentence.. In order to detect aspect category of sentiment tuple, we tried linear classification algorithm and sequence labelling. In the first scenario, we used a two layered approach. First layer consists of one-vs-all binary classifiers for the detection of each different aspect entity (E) and aspect attribute (A). This first layer is used to obtain the possibilities of an instance to belong to the corresponding classes (entity or attribute). The obtained probabilities will be further used in the second layer for the main classifier. In instance representation, we use unigram and bigram features occurring in the training set. Other experiment uses sequence labelling approach. In OTE detection our IOB standard and raw aspect category labelling types are capable to generate aspect category types of OTEs. Therefore, we could use this aspect category types to fill sentiment tuple's aspect category slot. Sequence labelling results are more successful than linear classification system. (%66.66) Third sub task aims to find OTE and aspect category at the same time. We showed sequence labelling success (%46.67) is quite well while detecting aspect category. This inference led us to use same technique in third sub task. We had chance to compare our system with just baseline results (%28.15). Because only shared task committee's baseline results are reported in Turkish language. On the other hand, our system performance was similar to other submissions on other languages. Latest sub task aims to fill sentiment polarity slots in sentiment tuples. Shared task provides different test set from previous sub tasks. In the test set, all the sentiment tuples are generated and aspect category and OTE slots are filled in gold standards. Our sentiment polarity detection method consists of a linear classification system. The classification system uses logistic regression algorithm and textual features as well as aspect category and OTE information. Our textual feature extraction method comes from bag-of-words. As preprocessing step, our words are represented as stems which come from morphological analysis. In addition to that, during testing phase, only words related to the OTEs are included to classification. This mentioned system performs %764.12 accuracy rate while evaluation. This result is % 2 higher than the best submitted constraint system in ABSA. in at the same level with best submitted constraint system in the shared task. The mentioned Best best performed performing constraint system uses deep learning techniques. Main inference becomes that the linear classification algorithms can perform as good as deep learning systems, if they use just the provided limited training sets. For extracting OTEs' related words, we applied three different approaches which employ sequence of words, dependency parsing relations and both of them together. The best performing one was employing both sequence of words and dependency parsing relations. Hereby, we can say that using both of sequence of words and dependency parsing relations is more successful than others while finding out the sentiment class of the current OTE. Our research's goal is to show the contributions of several NLP tools and Linguistic resources over aspect based sentiment analysis in Turkish Language. According to our experiments, this contribution over all subtasks of aspect based sentiment analysis can be seen clearly. Especially, dependency parsing and word vectors have powerful impacts on sentiment analysis. Another inference can be that morphological features have noticeable information. As a conclusion, Turkish is challenging language for aspect based sentiment analysis because it is agglutinative and free constituent order language. For solving problems that are created by Turkish linguistic challenges, additional NLP information and linguistic resources should be used.