Finally, exactly 2 weeks before the D-Day a.k.a defense day, I managed to finish my thesis and submitted 3 bound copies to the university. Why 3 I’m not sure, one of them will go to university library, and I’m supposed to get one back after the defense, the other one… perhaps because I have co-supervisor in different university.
I really like it that we don’t have to make a hardcover of the thesis, we just have to bind the pages in this special binder, with bright orange color for faculty of computer science. Makes it easy if I need to change few things, I just need to replace the pages *which is possible to be done even after I submitted them ;p*. The price for one binder is 6 something euros, hence including the printing maybe it’s around 20 euros in total for 3 copies. Plus I need to submit 2 special stamp of 14.6 euros. In the end, it costs me quite a lot in order to graduate from here ;p.
Some crazy stories behind…
- My supervisor is 5 months pregnant and I didn’t notice at all! I just thought that, “wow, she really gained weight recently…”, and turns out she’s expecting a baby girl ^^.
- On my way to the faculty administration to submit those bound copies, I checked the email just in case. And she sent me email pointing out that I had a (fatal) typo in one chart. After I double checked, it actually happened in all of the chart -_- I put the wrong label. So I hurriedly went back to the library, fixed it, re-printed the wrong pages, and again… struggled with the binder.
But really, I should thank my supervisor a lot because without her, I probably won’t finish this on time. She pushed me hard, always sets a deadline and expects it to be followed. If I couldn’t make it, she would give the impression that I didn’t try hard enough, and kinda put me down, but at the same time motivated me to do better. My greatest gratitude goes to her…
Anyway, I’m really glad that these weeks in hell of non-stop writing were finally over. The only thing left is the defense *Friday, 12-10-12.. save the date! ^^*, and thanks to this crazy supervisor of mine I’ve already finished the slides, just need to revise few slides. And we will have the dry run next Monday, to practice and get feedback from others.
So here you go, the abstract of my Master thesis, Exploiting CCG Derivations within Distributional Semantic Models, just in case you’re interested…
For the last decade, Distributional Semantic Models (DSMs) have been an active area of research to address the problem of understanding the semantics of words in natural language. The central assumption which functions as the core of DSMs is that the linguistic context surrounding a given word provides important information about its meaning, and hence word co-occurrence statistics can provide a natural basis for semantic representations.
The main idea behind this work is to explore ways to incorporate syntactic information within the distributional semantic models. This approach is known as structured DSMs, as opposed to unstructured DSMs which use a set of unordered words as linguistic context without any syntactic information. There have been some works focusing on structured DSMs which use syntactic categories such as part-of-speech (POS) tags (i.e. nouns, verbs, adjectives, etc.) and dependency relations between words (i.e. subject, direct object, modifier, etc.), to enrich the basis elements of context vector used to represent the meaning of a word in DSMs. Although it has been shown that dependency paths are a good approximation to semantic relations between words, the lack of information about syntactic category of each word in the context might become the shortcoming to fully exploit syntactic information within DSMs.
To this end, we exploit Combinatory Categorial Grammar (CCG) categories that provide a transparent relation between syntactic category and semantic type of the linguistic expression, to enrich the linguistic context used to represent the word meaning in DSMs. CCG categories, which are regarded as supertags, are supposed to be better at defining semantics than the traditional less informative POS tags since they carry more rich syntactic information.
We use the standard framework to build DSMs, which includes defining linguistic context, building the co-occurrence matrix, and applying various weighting schemes to the produced co-occurrence matrix. However, in order to build CCG-based DSM, constructing the co-occurrence matrix requires the use of CCG parsed corpus to extract the CCG categories of each word, hence full CCG parsing is carried out beforehand.
The constructed CCG-based DSM is then evaluated on one of widely known semantics tasks, which is word categorization, by comparing it with other types of DSM. From the experiments we could conclude that CCG-based DSM is generally better than POS-based DSM, and most certainly outperforms unstructured DSM that do not provide any syntactic information in the linguistic context. It is also shown that CCG-based DSM performs better than dependency-based DSM in the case of verb categorization task, demonstrating the importance of syntactic categories in defining the meaning of verbs.
We also investigate the impact of employing different context window on the performance of CCG-based DSM. Finally, we explore the effect of including the function words, which are grammatical words such as determiner, preposition, pronoun, etc. belonged to the group of closed-class words, in the co-occurrence matrix since existing works covering this topic usually use only content words including nouns, verbs, adjectives, and most adverbs.
Key words: Distributional Semantic Models, Combinatory Categorial Grammar, Word Categorization, Clustering