NTU Computational Linguistics Lab Theses and URECA Projects

2019 * 2018 * 2017 * 2016 * 2015 * 2014 * 2013 * 2012 * 2011

This page lists a selection of Final Year Projects (Undergraduate Theses), Masters and PhD Theses from the computation linguistics group at the Division of Linguistics and Multilingual Studies, Nanyang Technological University, Singapore. It also includes theses done outside NTU but co-supervised by us.

Extraction of neologisms from Japanese corpora Doctoral Dissertation, University of Melbourne, Australia.
Lexicon: http://edrdg.org/jmdict/edict_doc.html

2016

Lee Naomi Elizabeth (2016): An Etymological Study of Singapore Sign Language: The Influence of American Sign Language on Singapore Sign Language, Final Year Project, Linguistics and Multilingual Studies, Nanyang Technological University, Singapore, Abstract (click to toggle)
Though there is no officially recognized national sign language in Singapore, Singapore Sign Language (SgSL) is recognized by the local deaf community. It has influences from Shanghainese Sign Language (SSL) and American Sign Language (ASL), and is continually developing with locally generated signs. This study aims to give an insight into the influence of ASL on SgSL, which serves as a first look into the etymology of SgSL. 14 participants were recruited in this study; 3 participants were given a Swadesh list for sign languages consisting of 100 words, which they were asked to sign in SgSL. The videos of ASL signs for the same words were obtained online and presented alongside the SgSL signs to the other 11 participants, who gave judgments about the similarities of each pair of signs. The signs were also transcribed using the Hamburg Sign Language Notation System, or HamNoSys, and further analyzed based on handedness and four traditional phonological parameters – handshape, location, movement and orientation. This was done by calculating the Levenshtein distances between each pair of transcriptions. The similarity of the signs was then determined after consideration of the participants’ judgments and the analysis of phonological parameters, and it was found that the signs were similar to a great extent, which suggests that SgSL is heavily influenced by ASL.
Lim Jia Ying (2016): Investigation of Classifiers in Singapore Sign Language through Narratives: A Pilot Study, Final Year Project, Linguistics and Multilingual Studies, Nanyang Technological University, Singapore, Abstract (click to toggle)
Several sign language corpora such as the British Sign Language corpus have been set up over the years, providing a platform for extensive research to be done on classifiers. Yet little is known about the Singaporean variety as the Singapore Sign Language (SgSL) Corpus Project only kick-started recently. Focusing on video recordings of 12 SgSL users from varied sociolinguistic backgrounds, this study describes in detail the usage of prototypical classifiers in SgSL in terms of handshape-orientation combinations and structural patterns of occurrences. In addition, it explains why variance in classifier handshapes and morphemic functions occur. Entity classifiers were the most commonly produced classifier type, and together with Instrumental classifiers, the handshapes of the most commonly used classifiers in these two categories were found to be similar to those used in American Sign Language and Hong Kong Sign Language. Handling and Shape and Size Specifier classifiers were also found to be multi-morphemic and grammaticized in SgSL, suggesting that the most commonly used classifiers from each classifier type are highly lexicalized in SgSL, allowing them to remain stable and highly productive. This pilot study on classifiers thus serves to provide an avenue for future research, while highlighting issues of variation in SgSL.
Yeow Jun Jie (2016): Investigating the effectiveness of watching videos with L1 or L2 subtitles on second language acquisition, Final Year Project, Linguistics and Multilingual Studies, Nanyang Technological University, Singapore, Abstract (click to toggle)
This study investigates the effectiveness of watching videos for L2 learning and the effectiveness of L1 and L2 subtitles on L2 learning. Specifically it explores the potential for L2 videos in learning vocabulary and grammar. 60 Japanese L2 learners of various proficiencies were asked to watch clips from a native Japanese television drama with either English subtitles, Japanese subtitles or no subtitles. They were then asked to do a test which includes a comprehension section, vocabulary section and grammar section. A post-study interview was also conducted. Results were tabulated and overall test scores found that English subtitles were more beneficial for beginner learners while Japanese subtitles were more useful for intermediate to advanced learners (referred to as simply advanced learners in this study). Vocabulary scores showed a similar result. English subtitles impeded Japanese grammar learning for all levels of learners. Overall, Japanese subtitles are beneficial for Japanese learning for advanced learners for both vocabulary and grammar learning, while English subtitles are beneficial for beginner learners for vocabulary learning. Results from the study can serve as a pedagogical tool giving language learners a better idea as to how videos can benefit their learning, and as to which subtitles, L1 or L2, are more effective for their current proficiency level. Suggestions for future research have also been included.

2015

Jia Qian Ho (2015): Losing One’s Mind Over Meaning: Analysing the Behaviour of English Possessive Idioms, Final Year Project, Linguistics and Multilingual Studies, Nanyang Technological University, Singapore, Abstract (click to toggle)
Idioms are commonly found in everyday language and reflect the conventionalisations in speech communities. Regarding English idioms, past research have examined the syntactic and semantic analysis of idioms (Villavicencio & Copestake, 2002), along with their decomposability (Nunberg, 1978; Gibbs, 1989a,b) and comprehension (Titone & Connine, 1994; Cacciari & Tabossi, 2014). However, there has been little research on English possessive idioms despite their uniqueness and interesting properties. This thesis thus seeks to analyze the syntax and semantics of possessive idioms and describe their behaviour in terms of their decomposability and plausibility. A total of 514 idioms were categorized into co-indexed and separate possessive idioms and then grouped syntactically in order to be incorporated into new templates in the English Resource Grammar (Flickinger, 2011). Subsequently, the meaning of either each idiom component or paraphrase component was linked to WordNet (Fellbaum, 1998a) by choosing the most appropriate sense. The resulting comprehensive syntactic and semantic idiom descriptions allowed for analyses of their syntax, semantics, decomposability and plausibility. Results demonstrated the interplay between syntax and semantics and revealed novel aspects of possessive idioms, such as alternation and transformation in idioms. Furthermore, results confirmed that a degree of decomposability exists and suggested that possessive idioms could be categorized into four groups according to their projectability. The comprehensive idiom database will be released under an open license where it can be used as a dictionary and to further improve natural language processing applications.
Wenjie Wang (2015): A-not-A Questions in Mandarin Chinese: An HPSG Account, Final Year Project, Linguistics and Multilingual Studies, Nanyang Technological University, Singapore, Abstract (click to toggle)
In this paper, I look at a-not-a questions in Mandarin Chinese, and create an account based on the framework of Head-driven Phrase Structure Grammar (HPSG) and Minimal Recursion Semantics (MRS). While the a-not-a structure has seen extensive research in the past decades, they have largely been movement-based. Thus, this paper attempts to provide a non-movement but constraint-based HPSG/MRS account, which has thus far not been performed. Secondly, I have also begun initial implementation of said account into the HPSG- based Zhong [|] computational grammar for Chinese. While the basic forms of the a-not-a structure were accounted for and implemented, it was found that limitations in the system and formalisation prevented vp-not-vp questions from being successfully implemented.

2014

Sandra Soh (2014): Crosslingual corpus-based analysis of English and Japanese Verbs, Final Year Project, Linguistics and Multilingual Studies, Nanyang Technological University, Singapore, Abstract (click to toggle)
This study looked into divergences and correspondences that occurred in the translation of English verbs to Japanese verbs. A quantitative and qualitative approach was adopted in this study to analyze the distribution of different types of verbs in English and Japanese, based on the parallel NTU Multilingual Corpus (NTU-MC). Verb types examined include English phrasal verbs, Japanese compound verbs, as well as idiomatic verbal expressions and single verbs in English and Japanese. English phrasal verbs and Japanese compound verbs were hypothesized to be correlated. Using a Chi-square test, they were proven to be translation correspondences of each other. Other patterns of translation divergences were also detected, such as lexical gaps and lexical deviations. These divergences were attributed to translation style, but more so on differences in the language systems, such as the system of honorifics in Japanese, of which a similar system is not found in English. Findings from this study can shed some light concerning translation issues of English phrasal verbs into Japanese, and also for learners of either language to understand the semantic behaviours of English phrasal verbs and Japanese compound verbs.
Aiden Lim (2014): Acquiring Predominant Word Senses in Multiple Languages, Final Year Project, Linguistics and Multilingual Studies, Nanyang Technological University, Singapore, Abstract (click to toggle)
The accurate retrieval of the most appropriate sense of a word, in other words Word Sense Disambiguation, is a key process in machine learning and machine translation. Yet oftentimes, the reliabilty and accuracy of a system is a product of the input data which is itself dependent on financial and temporal restrictions. The resulting scarcity also affects the scope of WSD systems by limiting available data to specific languages. This paper proposes a Word Sense Disambiguation system that ignores these restrictions by being able to accept raw text input from any language and extract predominant word senses from it. This is done through a two-pronged approach of mathematically mapping the dis- tribution of a target word and its neighboring words before calculating the semantic similarity of both. By using both the Distributional Similarity Score and Semantic Similarity Score, it becomes possible to mathematically calculate and propose a Predominant Word Sense without any language-specific input. This paper reports how the system was used on English and Mandarin Chinese Wikipedia raw text to extract Predominant Word Senses with a fair degree of accuracy for a successful Word Sense Disambiguation System.

2013

Eshley Gao (2013): Crosslingual comparison of linguistic phenomenon in English, Japanese and Chinese, Final Year Project, Linguistics and Multilingual Studies, Nanyang Technological University, Singapore, Abstract (click to toggle)
Recent trends in computational linguistics tend to focus on how to represent meaning. The availability of parallel corpus has allowed researchers to study how languages convey the same information in different ways. This study adopts a quantitative and qualitative method to study translation shifts in the short novel – The Adventure of the Dancing Men. The parallel tri-text corpus in English-Japanese-Chinese is a sub-corpus of the NTU Multilingual Corpus. We tagged the concepts according to the senses in the WordNets, and annotated relationships between translation correspondents. The results show that 49.60% and 50.87% of distinct synsets in the English source text were linked in the English-Japanese and English-Chinese corpus respectively. Of the total linked concepts, 51.58% and 60.07% of them are exact correspondents of the source language in the English-Japanese and English-Chinese corpus correspondingly. The remaining contribute to evidence for translation shifts, which includes direct differentiation like hyponymy relationship to less straightforward variation like translation equivalents. The study also attempts to describe some of the translation shifts observed in the corpus. We estimate that more than half of the translation shifts were due to language differences, although translating style also played a part in the shifts. Data from this study can be used to train machine translation systems to produce more human-like translations. Second language learners of Japanese and Chinese can also take advantage of the data to learn how the same idea can be transmitted in different ways.
Le Tuan Anh and Sun Ying (2013): Question-Answering Machine based on Deep Linguistic Parsing, Master's Thesis, Institute of Systems Science, National University of Singapore, Singapore
Liling Tan (2013): Examining Crosslingual Word Sense Disambiguation, Master's Thesis, Linguistics and Multilingual Studies, Nanyang Technological University, Singapore, Abstract (click to toggle)
Understanding human language computationally remains a challenge at different levels, phonologically, syntactically and semantically. This thesis attempts to understand human language's ambiguity through the Word Sense Disambiguation (WSD) task. Word Sense Disambiguation (WSD) is the task of determining the correct sense of a word given a context sentence and topic models are statistical models of human language that can discover abstract topics given a collection of documents. This thesis examines the WSD task in a crosslingual manner with the usage of topic models and parallel corpus. The thesis defines a topical crosslingual WSD (Topical CLWSD) task as two subtasks (i) Match and Translate: finding a match of the query sentence in a parallel corpus using topic models that provides the appropriate translation of the target polysemous word (ii) Map: mapping the word-translation pair to disambiguate the concept respectively of the Open Multilingual WordNet. The XLING WSD system has been built to attempt the topical WSD task. Although the XLING system underperforms in the topical WSD task, it serves as a pilot approach to crosslingual WSD in a knowledge-lean manner. Other than the WSD task, the thesis briefly presents updates on the ongoing work to compile multilingual data for the Nanyang Technological University-Multilingual Corpus (NTU-MC). Both the NTU-MC project and the XLING system are related in their attempts to build crosslingual language technologies.
Muhammad Zulhelmy Mohd Rosman (2013): Creating derivational morphology links in Wordnet Bahasa, Final Year Project, Linguistics and Multilingual Studies, Nanyang Technological University, Singapore, Abstract (click to toggle)
Derivational morphology links are created for the Wordnet Bahasa, a combined Indonesian and Malay online lexical dictionary (Nurril Hirfana, Suerya, & Bond, 2011). The focus was to link root words to affixed words as affixation is one of the more apparent word formation processes in Bahasa Melayu. MorphInd, an Indonesian morphological analyser (Larasati, Kubon, & Zeman, 2011), is used to breakdown affixed words into their root form and affixes. Using Python 2.7 with NLTK, a raw mapping is done by matching the analysed words to the root forms. The derivational links in the Princeton Wordnet (PWN) are used to verify if the same links exist in Wordnet Bahasa. Redundant links are removed by the Part-of-Speech (POS) filter and Semantic Super Type filter. The links are then disambiguated using the Lesk algorithm, where the definitions and other components of the sense (e.g. hypernyms, hyponyms and examples) are compared for their similarity. However, the disambiguation process is rendered ineffective because of the high amount of errors still existing in Wordnet Bahasa. The derivational links are released as a separate file and only those with similar derivational links to PWN are added into Wordnet Bahasa. Erroneous entries that were identified using MorphInd are removed from Wordnet Bahasa.
Yu Jie Seah (2013): Contrastive Analysis of Pronouns across English, Mandarin Chinese and Japanese, Final Year Project, Linguistics and Multilingual Studies, Nanyang Technological University, Singapore, Abstract (click to toggle)
A qualitative and quantitative approach was used in this study to examine the distribution of pronouns in three languages, namely English, Mandarin Chinese and Japanese based on the parallel NTU Multilingual Corpus (NTU-MC) with English being the source language while Mandarin Chinese and Japanese translations are aligned to it at the sentence level. The pronouns are extracted from four subcorpora – two short stories, one essay and the other is an online article about Singapore’s tourism. However, due to time and space constraints, only pronouns from one subcorpus — The Adventure of the Speckled Band, a short story from the Sherlock Holmes series, is tagged, annotated and linked in the corpus. The results show that although English has the most number of pronouns, Mandarin Chinese has the highest percentage of referential pronouns. Also, English has more translated counterparts in Mandarin Chinese as compared to Japanese. We attributed this to the difference in usage of pronouns in the languages. Deprominalisation, surprisingly, was even for both corpora. We believed this to be due to influence from the English text. Findings from this study can shed some light concerning translation issues on pronoun usage for learners of the languages and also contribute to pronoun translation across languages.
Zinaida Pozen (2013): Using Lexical and Compositional Semantics to Improve HPSG Parse Selection, Master's Thesis, University of Washington, United States, Abstract (click to toggle)
Accurate parse ranking is essential for deep linguistic processing applications and is one of the classic problems for academic research in NLP. Despite significant advances, there remains a big need for improvement, especially for domains where gold-standard training data is scarce or unavailable. An overwhelming majority of parse ranking methods today rely on modelling syntactic derivation trees. At the same time, parsers that output semantic representations in addition to syntactic derivations (like the monostratal DELPH-IN HPSG parsers) offer an alternative structure for training the ranking model, which could be further combined with the baseline syntactic model score for re-ranking. This thesis proposes a method for ranking the semantic sentence representations, taking advantage of compositional and lexical semantics. The methodology does not require sense-disambiguated data, and therefore can be adopted without requiring a solution for word sense disambiguation. The approach was evaluated in the context of HPSG parse disambiguation for two different domains, as well as in a cross-domain setting, yielding relative error rate reduction of 11.36% for top-10 parse selection compared to the baseline syntactic derivation-based parse ranking model, and a standalone ranking accuracy approaching the accuracy of the baseline syntactic model in the best setup.

2012

Jeanette Tan Yi Wen (2012): Automatic Generation of Multilingual Crossword Puzzles with WordNet, Final Year Project, Linguistics and Multilingual Studies, Nanyang Technological University, Singapore, Abstract (click to toggle)
This project implements an automatic crossword puzzle generator in Python and a JavaScript solving interface that allows users to tailor the puzzle to suit their preferences. The crosswords are multilingual with the hints and solutions in different languages as it makes use of the Open Multilingual Wordnet which has linked the WordNets of many different languages together. It aims to provide a fun and effective way for language learners to acquire vocabulary.

Randomly generated multilingual crosswords: solve one
Ko Tabitha (2012): Chinese-English translations for passive constructions, Final Year Project, Linguistics and Multilingual Studies, Nanyang Technological University, Singapore, Abstract (click to toggle)
Retrieval of accurate translations is crucial in today’s technologically advanced world where intercultural communications are frequent and necessary. Past research surrounding passives and translations has largely focused on English-Chinese translations. Therefore, this paper seeks to provide new insight by concentrating on Chinese-English passive translations. In view of past observations, five hypotheses are proposed: (i) BEI++ hypothesis; (ii) RANG+ hypothesis; (iii) BE+ hypothesis; (iv) GET-control hypothesis; (v) BY-ACTOR hypothesis. For the purpose of this study, a Chinese-English multilingual corpus from Korea Advanced Institute of Science and Technology (KAIST) was used. Three of the proposed hypotheses, namely BEI++ hypothesis, BE+ hypothesis and GET-control hypothesis, were supported. However, the RANG+ hypothesis and BY-ACTOR hypothesis were not supported. Additionally, the reduction of Chinese passives to PAST PARTICIPLE PHRASES and two new types of Chinese passive constructions were noticed. Furthermore, analysis of English translations exhibited other types of English passives previously overlooked. Results also illustrated the influence of both source language (SL) and target language (TL) norms in translations. An examination assessing current machine translations indicated a lack of appropriate translations. Thus, two sets of actions for Chinese-English passive translation have been proposed. Future research exploring the application of the proposed actions is recommended.
Sheefa Samara Sameha (2012): Make Up Your Mind: An Analysis of Idiomatic Possessive Verb Phrase Constructions in English, Final Year Project, Linguistics and Multilingual Studies, Nanyang Technological University, Singapore, Abstract (click to toggle)
Idiomatic constructions, particularly possessive ones, are inadequately described in English grammar. 307 idioms are structurally clustered and their syntactic and semantic aspects discussed. Minimal recursion semantics of idioms indicates the possessive relationships within the expression. Compositionality is found to affect little of idiomaticity. Conceptual metaphors and image schema are suggested as possible means of understanding when literal expressions become non-literal. Findings point to greater shortcomings in available literature than firstly assumed. A novel means of idiom implementation with a focus on easy access and visual representation is proposed.

2011

Yun Rui Kong (2011): Shape or Substance?: A Cross-linguistic Study of Bilinguals in English and Chinese, Final Year Project, Linguistics and Multilingual Studies, Nanyang Technological University, Singapore, Abstract (click to toggle)
According to the Sapir-Whorf Hypothesis, language shapes how we think. When the hypothesis is tested on monolinguals, the method used will be more straightforward and direct compared to bilinguals. For bilinguals, there are many more factors that need to be considered such as language dominance, language proficiency and similarities or differences between the first and second language. Yet, the most important understanding that we need to have is if the bilingual thought a result of two separate language systems or one system that integrates the two languages together. In this study, we examine if and how the presence and acquisition of a second language in English affects the concept of individuation in speakers of Chinese. This is investigated by looking at how Chinese monolingual, Chinese-English sequential bilingual and English-Chinese simultaneous bilingual undergraduates perform categorization in an online triad-matching task. In the task, participants were required to decide if the shape or substance (material) alternative was more similar to the test image. It is hypothesized for the simultaneous bilinguals to have the highest percent shape response followed by the sequential bilinguals then the Chinese monolinguals. But what we found was that the sequential bilinguals had a lower percent shape response than the Chinese monolinguals. This is suggested to be a result of increased cue sensitivity towards English when acquiring English as their second language. In addition, the results of the Chinese monolinguals were also compared to those of Japanese and English monolinguals in the study by Imai (2000) to see how differences in linguistic features of languages can affect thought. From the findings of this study, we gather that the bilingual thought is a result of the integration of two language systems as one since the acquisition of a second language creates shift in cognition. This is illustrated through how bilinguals categorize compared to monolinguals.
Liling Tan (2011): Building the Foundation Text for Nanyang Technological University — Multilingual Corpus (NTU-MC), Final Year Project, Linguistics and Multilingual Studies, Nanyang Technological University, Singapore, Abstract (click to toggle)
The NTU-MC is a multilingual corpus that taps on the availability of multilingual text available in Singapore. The current version of NTU-MC contains a total of ~375,000 words (15,096 sentences) for the NTU-MC in 6 languages (English, Chinese, Japanese, Korean, Indonesian and Vietnamese) from 6 language families (Indo-European, Japonic, Austro-Asiatic, Sino-Tibetan, Austronesian and Korean as a language isolate); all text in English, Chinese, Japanese, Korean and Vietnamese were Part Of Speech (POS) tagged. This project focuses on compiling the foundation text for the NTU-MC and this dissertation describes the motivations, the corpus compilation process and internal and cross-corpora evaluation of the corpus output. The corpus will be made available to the public under the Creative Common – Attribute 3.0 Unported license in Summer 2011.
(First FYP from LMS!)

Last modified: 2020-01-02