Computer Science ETDs

Author

Piyo Song

Publication Date

12-1-2013

Abstract

Natural language processing tasks typically start with the basic unit of words, and then from words and their meanings a big picture is constructed about what the meanings of documents or other larger constructs are in terms of the topics discussed. Social media is very challenging for natural language processing because it challenges the notion of a word. Social media users regularly use words that are not in even the most comprehensive lexicons. These new words can be unknown named entities that have suddenly risen in prominence because of a current event, or they might be neologisms newly created to emphasize meaning or evade keyword filtering. Chinese social media is particularly challenging. The Chinese language poses challenges for natural language processing based on the unit of a word even for formal uses of the Chinese language, social media only makes word segmentation in Chinese even more difficult. Thus, even knowing what the boundaries of words are in a social media corpus is a difficult proposition. For these reasons, in this document I propose the Pointillism approach to natural language processing. In the pointillism approach, language is viewed as a time series, or sequence of points that represent the grams' usage over time. Time is an important aspect of the Pointillism approach. Detailed timing information, such as timestamps of when posts were posted, contain correlations based on human patterns and current events. This timing information provides the necessary context to build words and phrases out of trigrams and then group those words and phrases into topical clusters. Rather than words that have individual meanings, the basic unit of the pointillism approach is trigrams of characters. These grams take on meaning in aggregate when they appear together in a way that is correlated over time. I anticipate that the pointillism approach can perform well in a variety of natural language processing tasks for many different languages, but in this document my focus is on trend analysis for Chinese microblogging. Microblog posts have a timestamp of when posts were posted, that is accurate to the minute or second (though, in this dissertation, I bin posts by the hour). To show that trigrams supplemented with frequency information do collect scattered information into meaningful pieces, I first use the pointillism approach to extract phrases. I conducted experiments on 4-character idioms, a set of 500 phrases that are longer than 3 characters taken from the Chinese-language version of Wiktionary, and also on Weibo's hot keywords. My results show that when words and topics do have a meme-like trend, they can be reconstructed from only trigrams. For example, for 4-character idioms that appear at least 99 times in one day in my data, the unconstrained precision (that is, precision that allows for deviation from a lexicon when the result is just as correct as the lexicon version of the word or phrase) is 0.93. For longer words and phrases collected from Wiktionary, including neologisms, the unconstrained precision is 0.87. I consider these results to be very promising, because they suggest that it is feasible for a machine to reconstruct complex idioms, phrases, and neologisms with good precision without any notion of words. Next, I examine the potential of the pointillism approach for extracting topical trends from microblog posts that are related to environmental issues. Independent Component Analysis (ICA) is utilized to find the trigrams which have the same independent signal source, i.e., topics. Contrast this with probabilistic topic models, which leverage co-occurrence to classify the documents into the topics they have learned, so it is hard for it to extract topics in real-time. However, pointillism approach can extract trends in real-time, whether those trends have been discussed before or not. This is more challenging because in phrase extraction, order information is used to narrow down the candidates, whereas for trend extraction only the frequency of the trigrams are considered. The proposed approach is compared against a state of the art topic extraction technique, Latent Dirichlet Allocation (LDA), on 9,147 labelled posts with timestamps. The experimental results show that the highest F1 score of the pointillism approach with ICA is 4% better than that of LDA. Thus, using the pointillism approach, the colorful and baroque uses of language that typify social media in challenging languages such as Chinese may in fact be accessible to machines. The thesis that my dissertation tests is this: For topic extraction for scenarios where no adequate lexicon is available, such as social media, the Pointillism approach uses timing information to out-perform traditional techniques that are based on co-occurrence.

Language

English

Keywords

Unknown Words Extraction, Trends Extraction

Document Type

Dissertation

Degree Name

Computer Science

Level of Degree

Doctoral

Department Name

Department of Computer Science

First Advisor

Crandall, Jedidiah

First Committee Member (Chair)

Saia, Jared

Second Committee Member

Luger, George

Third Committee Member

Wallach, Dan

Share

COinS