Search

This paper proposes to develop a model to detect the text semantic similarity in Thai by using a Siamese neural network with MaLSTM. As the text's intent is varied and hard to analyze, thus making the comparison between two sentences for the semantics similarity challenging. Our project wants to find similar intent of two questions by comparing the question between frequently asked questions (FAQs) and input questions from customers via Facebook Messenger of computer engineering at the King Mongkut's University of Technology Thonburi (CPE KMUTT). The data gather manually from the Pantip, CPE KMUTT FAQ, and Wikipedia, which create a dataset and data corpus. Although, the overall score performance of the model is not as we expected. Our model shows promising results, as it can detect between two questions' intent similarity. Results suggest that with the increase of the data, the model should have great potential for finding semantic similarity.

Word segmentation is an important preprocessing step in natural language processing applications, particularly in languages with no demarcation indicators including Thai. A simple method like dictionary-based segmentation does not consider the context of the sentence. This paper proposes an attention-based deep learning approach for Thai word segmentation. With the help of attention, the model can learn character correlations across the entire sentence without gradient vanishing or gradient explode problems and tokenize them into word vectors. The goal of this research is to test three different types of attention mechanisms to determine the effectiveness of word tokenization. The visualization of attention for each attention mechanism is also included as an outcome.

Latent Dirichlet Allocation(LDA) is one of the powerful techniques in extracting topics from a document. The original LDA takes the Bag-of-Word representation as the input and produces topic distributions in documents as output. The drawback of Bag-of-Word is that it represents each word with a plain one-hot encoding which does not encode the word level information. Later research in Natural Language Processing(NLP) demonstrate that word embeddings technique such as Skipgram model can provide a good representation in capturing the relationship and semantic information between words. In recent studies, many NLP tasks could gain better performance by applying the word embedding as the representation of words. In this paper, we propose Deep Word-Topic Latent Dirichlet Allocation(DWT-LDA), a new process for training LDA with word embedding. A neural network with word embedding is applied to the Collapsed Gibbs Sampling process as another choice for word topic assignment. To quantitatively evaluate our model, the topic coherence framework and topic diversity are the metrics used to compare between our approach and the original LDA. The experimental result shows that our method generates more coherent and diverse topics.

Prosody is an essential aspect of speech, as it carries both lexical and non-lexical information. A conventional approach for studying speech prosody is to collect and analyze F0 data based on certain hypotheses and then develop a theory based on the observation as the final conclusion of the study. This process is however far from complete, as the developed theory has not been actually tested for its ability to predict actual acoustic data. This paper presents PENTATrainer2, a prosody modeling tool based on the parallel encoding and target approximation framework. PENTATrainer2 can facilitate prosody studies in testing hypotheses and theories against speech data by using an automatic analysis-by-synthesis and stochastic learning algorithm. Users can flexibly design the annotation scheme based on their own hypotheses and determine whether the hypothesized categories can lead to accurate synthetic F0 contours. PENTATrainer2 consists of three main components: multi-layer annotation, target approximation, and stochastic optimization. First, acoustic data are annotated in parallel layers, each of which corresponds to a functional category that may affect F0 contours. These layers are then compiled into unique functional combinations. The combinations represent underlying invariant representations of communicative functions and their interaction with each other. Target approximation parameters of each combination are then learned through analysis-by-synthesis and stochastic optimization. Pilot tests of PENTATrainer 2 have been conducted on Thai, Mandarin and English. The results demonstrate not only high accuracy of the synthesized F0 contours but also distinctive contrasts in the distribution of pitch target parameters. This indicates the effectiveness of PENTATrainer2 in modeling speech prosody.

Named Entity Recognition (NER) is a handy tool for many natural language processing tasks to identify and extract a unique entity such as person, location, organization and time. In English and Chinese, NER has been thoroughly researched and is able to be applied in more practical settings. Its development in Thai is still limited because of rare resources and language difficulties such as the lack of boundary indicator for words, phrases and sentences. In this paper, we present an application of Bi-LSTM-CRF with word/character level representation, to solve this problem. Firstly, we prepared texts by tokenizing a sentence to a bunch of words. We then prepared word representation and Bi-LSTM character representation. In the end, we built a recurrent neural network combined with CRF to learn the sequence of text and extract the knowledge to build NER recognition to overcome this problem. Our model was evaluated by the NER opensource corpus from a Facebook group ThaiNLP. The results of our model yielded precision, recall, and F1 at 91.79%, 91.51% and 91.65% respectively.

This paper presents the design and implementation of the ALICE O2 Control System prototype. The design of our system is based on the ALICE O2 requirements. The system allows the crew operating the experiment to send simple commands in order to control the ongoing processes running in the computing farm. The sent command is capable of reacting to changing conditions and operational the requirements of experiments' processes. The system supports general commands such as pausing or resuming an ongoing activity, applying a new configuration or terminating the current operation and moving it to a standby status. The topics learnt from this study include communication patterns, multi-level communication, distributed systems as well as state machine controlling. The system is based on the C++ programming language.

Named Entity Recognition (NER) is very important in many natural language processing tasks, especially information extraction. The problem of NE extraction in Thai is much more complicated than English because Thai language lacks orthography and boundary indicator between words. In this paper, we presented a research work in the field of NER with the emphasis on person name recognition (PNR) in Thai text. Our proposed method consists of 4 steps. First, text is tokenized into a set of words. Second, a part-of-name probability is computed for each word using Odds with Laplace smoothing and Logistic function. Third, name candidates are selected based on the likelihood probability. Finally, the end point of name is identified using a set of rules and a drop rate threshold. We then evaluated out method using 1,700 online news from the InterBEST 2009 corpus. The results show that the proposed method yields average precision, recall, f-measure and accuracy at 75.21%, 98.10%, 85.15%, and 81.05% respectively.

This paper proposes a new hybrid method for Thai word segmentation using crowd-sourced dictionary integrated with word bi-gram model. The main dictionary is extracted into basic and compound word dictionaries to improve dictionary based algorithm performance. The word segmentation process begins with heuristic exhaustive matching algorithm using basic word dictionary to generate all possible basic word sequence candidates from an input string. Then, the best candidate is selected by word bi-gram model to solve ambiguity problem. Finally, the sequence of basic words is combined into compound words with compound word dictionary. Another part of this work is applying crowdsourcing paradigm. We implemented a web application for training bi-gram model and dictionary updates from user feedbacks. This process improves the lexical knowledge of the platform over the time. The algorithm was evaluated with two corpora. With InterBEST 2009 corpus, the proposed algorithm yields average precision, recall and f-measure at 97.52%, 97.70%, and 97.63%. With social network corpus, the proposed method yields average precision, recall and f-measure at 98.47%, 98.59%, and 98.54% respectively.

This paper reports the development of a quantitative target approximation (qTA) model for generating F(0) contours of speech. The qTA model simulates the production of tone and intonation as a process of syllable-synchronized sequential target approximation [Xu, Y. (2005). "Speech melody as articulatorily implemented communicative functions," Speech Commun. 46, 220-251]. It adopts a set of biomechanical and linguistic assumptions about the mechanisms of speech production. The communicative functions directly modeled are lexical tone in Mandarin and lexical stress in English and focus in both languages. The qTA model is evaluated by extracting function-specific model parameters from natural speech via supervised learning (automatic analysis by synthesis) and comparing the F(0) contours generated with the extracted parameters to those of natural utterances through numerical evaluation and perceptual testing. The F(0) contours generated by the qTA model with the learned parameters were very close to the natural contours in terms of root mean square error, rate of human identification of tone, and focus and judgment of naturalness by human listeners. The results demonstrate that the qTA model is both an effective tool for research on tone and intonation and a potentially effective system for automatic synthesis of tone and intonation.

This paper proposes an acoustic-to-articulatory inversion using the Particle Swarm Optimization (PSO). We present a schematic and a detailed design of the acoustic-to-articulatory inversion system. The system is implemented by using Praat script and Java, with VocalTractLab as a speech synthesizer. The target data of our synthetic utterance are 5 disyllabic utterances consisting of 9 Thai monophthongs. For each syllable, the synthetic utterances are synthesized from 15 articulatory parameters by which their values are estimated using PSO with inertia weight. The fitness values of the system are evaluated in the term of the sum of the squared errors (SSEs) of these articulatory parameter values. To assess the results, the original and the synthetic utterances are compared in the forms of spectrograms, and F1–F3 formant frequency contours. For our system results, the good agreement between the original and synthetic utterance was achieved for F1 and F2.

Relationships among gene expression levels may be associated with the mechanisms of the disease. While identifying a direct association such as a difference in expression levels between case and control groups links genes to disease mechanisms, uncovering an indirect association in the form of a network structure may help reveal the underlying functional module associated with the disease under scrutiny. This paper presents a method to improve the biological relevance in functional module identification from the gene expression microarray data by enhancing the structure of a weighted gene co-expression network using minimum spanning tree. The enhanced network, which is called a backbone network, contains only the essential structural information to represent the gene co-expression network. The entire backbone network is decoupled into a number of coherent sub-networks, and then the functional modules are reconstructed from these sub-networks to ensure minimum redundancy. The method was tested with a simulated gene expression dataset and case-control expression datasets of autism spectrum disorder and colorectal cancer studies. The results indicate that the proposed method can accurately identify clusters in the simulated dataset, and the functional modules of the backbone network are more biologically relevant than those obtained from the original approach.

The vast amount of data on gene expression that is now available through high-throughput measurement of mRNA abundance has provided a new basis for disease diagnosis. Microarray-based classification of disease states is based on gene expression profiles of patients. A large number of methods have been proposed to identify diagnostic markers that can accurately discriminate between different classes of a disease. Using only a subset of genes in the pathway, such as so-called condition-responsive genes (CORGs), may not fully represent the two classification boundaries for Case and Control classes. Negatively correlated feature sets (NCFS) for identifying CORGs and inferring pathway activities are proposed in this study. Our two proposed methods (NCFSi and NCFS-c) achieve higher accuracy in disease classification and can identify more phenotype-correlated genes in each pathway when comparing to several existing pathway activity inference methods.

This paper proposes an effective method for identifying functional modules of the weighted gene co-expression network using a minimum spanning tree (MST) approach coupled with network neighborhood connectivity. The MST-based gene co-expression network was reconstructed to serve as the backbone of gene co-expression network. Highly connected hub genes were identified based on the connectivity of the backbone network. All sub-networks were extracted by expanding from the hub genes to their neighborhood genes. Finally, functional modules were identified by integrating sub-networks with similar gene expression profiles. We tested the method with both simulated and autism spectrum disorder microarray data sets. The results show that our approach is better in highlighting the hub genes and can effectively identify functional modules with highly enriched pathways.

Feature selection (FS) plays a crucial role in machine learning to build a robust model for either learning or classification from a large amount of data. Among feature selection techniques, the Relief algorithm is one of the most common due to its simplicity and effectiveness. The performance of the Relief algorithm, however, could be dramatically affected by the consistency of the data patterns. For instance, Relief-F could become less accurate in the presence of noise. The accuracy would decrease further if an outlier sample was included in the dataset. Therefore, it is very important to select the samples to be included in the dataset carefully. This paper presents an effort to improve the effectiveness of Relief algorithm by filtering samples before selecting features. This method is termed Sample Filtering Relief Algorithm (SFRA). The main idea of this method is to discriminate outlier samples out of the main pattern using self organizing map (SOM) and then proceed with feature selection using the Relief algorithm. We have tested SFRA with a gene expression dataset of interferon-?(IFN-?) response of Hepatitis B patients that contains outlier data. SFRA could successfully remove outlier samples that have been verified by visual inspection by experts. Also, it has better accuracy in separating the relevant and irrelevant features than other feature selection methods considered.