Cluster analysis related to computational linguistics seldom concerned with Pragmatics level. Features of corpus on Pragmatics level related to specific situations, including backgrounds, titles and habits. To improve the accuracy of clustering for conversations collected from international students in Tsinghua University, it required contextual features. Here, we collected four-hundred conversations as a corpus and built it to Vector Space Model. With the Oxford-Duden Dictionary and other methods we modified the model and concluded into three groups. We testified our hypothesis through self-organizing map neural network. The result suggested that the modified model had a better outcome.
Conversational Corpus Contextual Features VSM SOM1. Introduction
Natural Language Processing (NLP) involves different levels, including Morphology, Syntax, Semantics and Pragmatics [1]. Different levels are applied to various methods and applications, such as Segmentation based on Morphology and Syntax. With advanced statistics models and algorithms these applications are able to perform a high accuracy. However, a few clustering algorithms for conversational corpus are not aware with levels in Semantics or Pragmatics, which results in ambiguity to categorize specific situations [2]. As an illustration, the traditional method of Vector Space Model (VSM) has some limitations. The method pays its entire attention to the frequency without concern about Semantics and Pragmatics, resulted in a “false positive” or “false negative” match. As a consequence, the algorithm based on VSM above lacks enough ability to represent the corpus and cluster them into right groups.
Some researchers have already applied Semantics into their studies. To overcome the limitations of VSM, there is a combination between VSM and some lexical databases such as WordNet [3]. Same work related to Chinese corpus was conducted based on HowNet [4]. The key is to modify the default VSM with Semantics information. Instead of the entire attention to word frequency, these studies combined similar words into a same concept. As a result, the corpus was built into a conceptual tree rather than VSM before. With these efforts the model had a quite low dimensionality and resulted in a good performance related to auto-summarization or categorization.
This paper concentrates more on Pragmatics, particularly on Contextual Features. The experiment is conducted on Conversational Corpus with eight different situations in Table 1, including “Hospital”, “Restaurant”, “Renting House”, “Inside Class”, “After Class”, “Airport”, “Baber’s” and “Bank”. Contextual features are the very representative ones in each situation, such as conversation backgrounds and titles. Features in different situations are distinct to others, thus they play a very important role in clustering. We use two methods to extract contextual features from corpus. The experiment applies these contextual features to modify the default VSM that is calculated based on word frequency and testify this hypothesis by Self-organizing map (SOM) neural network [5]. The experiment result suggests that the combination with contextual features in Pragmatics level bring a better outcome for clustering.
2. Materials and Methods2.1. Conversational Corpus
The experiment was conducted on the conversational corpus collected from international students in Tsinghua University. The experiment was in eight categories and each of them represented a specific situation, including “Hospital”, “Restaurant”, “Renting House”, “Inside Class”, “After Class”, “Airport”, “Barber’s” and “Bank”. Each category consisted of fifty different conversations recorded of daily life from these international students. We got rid of titles in each conversation and reorganized them with word segmentation. As a result, the corpus had nearly 10,000 words and about 5000 word tokens after the stop list.
2.2. The Oxford-Duden Dictionary
The experiment used the Oxford-Duden dictionary to map the same category discussed above in order to extract specific contextual features on Pragmatics level. The dictionary was organized into several categories and illustrated with pictorial items within particular situation. This organization helped us to extract features like backgrounds easily. For example, there were several keywords related to “Hospital” illustrated on the dictionary, such as “Drag”, “Blood” and “Alcohol”, which were very common in the hospital and of great possibility to be referred in conversations. Besides, we were able to access to lots of features related to habits through the dictionary. It was significant to consider these features as integrity rather than separated words. With the Oxford-Du- den dictionary, we selected several keywords and maintained a list for modification of VSM (Figure 1).
Categories in Oxford-Duden dictionary
2.3. Procedures
The first step was to build VSM from collected corpus [6]. We used four hundred (50*8) conversations as input to construct VSM. The values in VSM were calculated by the frequency of each word. Without a clear-cut boundary for features in VSM, we also paid attention to low-frequency words, for some of them may contain important contextual information. Therefore, we kept all features in the model. The VSM in this state was called the “Default” group. General definition of VSM was illustrated as below:
(1)
and are two vectors while the Equation (1) is to make a comparison between them.
We selected eight keywords for each category which were often referred in each specific situation. For example, when we talked about words like “Food”, “Waiter” and “Beverage”, there was a great possibility that the conversation occurred in a restaurant. These words were very common and distinct because of their identities like backgrounds or titles. It was usual for us to hear words like beverage in a restaurant as well as various foods. Besides, the conversation in a restaurant was always occurred between customers and waiters. Therefore, the experiment maintained a self-selected keyword list of sixty-four words (8*8) which were representative for each situation (Details in Table 1). The default VSM was then modified by self-selected keyword list and was called the “Defined” group. For each word in VSM occurred in the list, we weighted it for twice than its original value.
Besides, we used the Oxford-Duden dictionary for the keywords selection. The experiment chose ten words for each specific situation from different categories in the dictionary. For example, words like “Teacher”, “Student” and “Book” were selected from the dictionary. In the dictionary, these words above were all in a same category. Within this specific situation, keywords like above were considered as a key feature to represent the situation, because all of these keywords shared a same background or common relationship between people or even their habits. As a consequence, the experiment maintained a dictionary-selected keyword list of eighty (10*8) words which were representative for each situation (Details in Table 1). The default VSM was then
ReferencesJurafsky and Martin (2000) Speech and Language Processing: An Introduction to Natural Language Processing, Computational Linguistics, and Speech Recognition. Prentice Hall.Lewis, D.D. and Hayes, P.J. (1994) ACM Trans-actions on Information Systems: Special Issue on Text Categorization, Vol. 12. ACM Press.Ji, H., Luo, Z.S., Wang, M. and Gao, X.Y. (2002) Summarizing Based on Concept Counting and Hierarchy Analysis. The Natural Language Processing and Knowledge Engineering (NLPKE) Mini Symposium of the 2002 IEEE International Conference on Systems, Man and Cybernetics (SMC2002).Liao, S.S. and Jiang, M.H. (2005) An Improved Method of Feature Selection Based on Concept Attributes in Text Classification. Advances in Natural Computation, Lecture Notes in Computer Science, 3610, 1140-1149.
http://dx.doi.org/10.1007/11539087_152Kohonen, T. (1987) Self-Organization and Associative Memory. 2nd Edition, Springer-Verlag, Berlin.Salton, G., Singhal, A., Buckley, C., et al. (1994) Automatic Text Decomposition Using Text Segments and Text Themes. Text Retrieval Conference, Washington DC.
(1)