A Study of the Vowel Context Effect on Initial Stops of Mandarin Produced by Native and Nonnative Speakers

This study, based on corpus materials, investigates the “voice onset time” (VOT) of Mandarin word-initial stops in isolated syllables according to the effect of vowel contexts produced by native and nonnative speakers. Here, 1250 monosyllables of word-initial plosives /p/, /t/, /k/, /p[Formula: see text]/, /t[Formula: see text]/, and /k[Formula: see text]/ were uttered in combination with the vowels /a/, /i/, and /u/ in four tone contexts except /ki/ and /k[Formula: see text]i/ that are phonetically illegal in Mandarin by 20 participants (10 native Chinese speakers and 10 Urdu learners of Chinese). Results show that for native Chinese speakers, the VOTs of aspirated stops followed by the high vowels /i/ and /u/ are significantly longer than those followed by the low vowel /a/, and unaspirated stops followed only by the high back vowel /u/ are significantly longer than those followed by the low vowel /a/. For native Urdu speakers, the mean VOTs of word-initial stops in Mandarin monosyllables have no significant effect for both aspirated and unaspirated ones in combination with different vowels. Understanding the results of this study will be of assistance to second language learning and teaching.

A Cross-Lingual Speech Synthesis System of Malay and Indonesian Based on HMM–DNN

English as a second language is widely used in countries such as Malaysia and Indonesia, and it is common for English words to appear in Malay and Indonesian sentences. Malay and Indonesian have high homology and relatively few electronic language resources. We combine the corpus datasets of these two similar languages to design and implement a HMM–DNN-based cross-lingual speech synthesis system for Malay (including English words) and Indonesian (including English words). The methods used include: sharing synthesis units between Malay, Indonesian, and English, designing unified context attributes and question set in the process of acoustic model training, speaker-adaptive training with speech corpus of these three languages, and synthesizing speech using speaker-dependent Malay and Indonesian acoustic models. Experimental results show that the speech synthesis quality of the system is better than the traditional Hidden Markov model-based cross-lingual speech synthesis system.

Reading Process of Arab Children: An Eye-Tracking Study on Saudi Elementary Students

Many previous eye-tracking studies were conducted to examine how adult readers process different written languages. Relatively, only few eye-tracking studies have been conducted to observe the reading process of Arab children. This study investigated the influence of orthographic regularity on Saudi elementary grades’ English and Arabic words recognition. The eye movements of 15 grade-four students and 15 grade-six students were recorded while they read words that differ in frequency and regularity. Analysis of the visual information from the word-recognition process shows differences in the students’ eye movements for the two languages. There were statistically significant differences in the total fixation duration and fixation count between the two languages and between both groups. All the students showed longer processing time for English sentences than Arabic ones. However, Arabic-speaking students were influenced by English orthography with more processing difficulty for English irregular words. The visual information shows that more cross-linguistic differences are found in grade-four students’ results. Grade-four students transferred their first language (L1) reading strategies to read English words; however, Arabic reading methods cannot be effectively applied to reading irregular orthographies like English. This explains the increased eye-movement measurements of grade-four students compared to grade-six students, who fixated more on unfamiliar English words. Although orthographic regularity had a major effect on the word-recognition process in this study, the development of the students’ Arabic and English orthographic knowledge affected the progress of their visual word recognition across the two levels.

DSC_NEXT: Document-Level Sentiment ClassificatioN with EXTensional-Information

Document-level sentiment classification is to assign an overall sentiment polarity to an opinion document. Some researchers have already realized that, in addition to document texts, extensional-information such as product features and user preferences can be quite useful. Many previous studies represent them as ID-type extensional-information and incorporate them into deep learning models. However, they ignore the descriptive extensional information that is also useful for document representations. This paper covers the following aspects: (1) introduces the Description of Opinion Target (DOT), a new extensional-information for document-level sentiment classification, (2) builds the Document-level Sentiment ClassificatioN with EXTensional-information (DSC_NEXT) dataset which consists of three datasets: IMDB_NEXT, Yelp_NEXT and CMRDB_NEXT and (3) validates the effectiveness of DOT by performing experiments based on current state-of-the-art (SOTA) document-level sentiment analysis methods. Implications for using extensional-information in neural network models are also considered.

Sentiment Analysis Technologies in AliMe — An Intelligent Assistant for E-commerce

In this paper, we introduce a sentiment analysis framework and its corresponding key techniques used in AliMe, an artificial intelligent (AI) assistant for e-commerce customer service, whose fundamental ability of sentiment analysis provides support for five upper-layer application modules: user sentiment detection, user sentiment comfort, sentimental generative chatting, user service quality control and user satisfaction prediction. Detailed implementation of each module is demonstrated and experiments show our framework not only performs well on each single task but also manifests its competitive business value as a whole.

Analyzing Subword Techniques to Improve English to Sinhala Neural Machine Translation

Neural machine translation (NMT) is a remarkable approach which performs much better than the Statistical machine translation (SMT) models when there is an abundance of parallel corpus. However, vanilla NMT is primarily based upon word-level with a fixed vocabulary. Therefore, low resource morphologically rich languages such as Sinhala are mostly affected by the out of vocabulary (OOV) and Rare word problems. Recent advancements in subword techniques have opened up opportunities for low resource communities by enabling open vocabulary translation. In this paper, we extend our recently published state-of-the-art EN-SI translation system using the transformer and explore standard subword techniques on top of it to identify which subword approach has a greater effect on English Sinhala language pair. Our models demonstrate that subword segmentation strategies along with the state-of-the-art NMT can perform remarkably when translating English sentences into a rich morphology language regardless of a large parallel corpus.

Speaker Identification and Its Application to Social Network Construction for Chinese Novels

Character is one of the three elements of a novel, and conversation is an important way to describe characters. The personality, emotion, and interpersonal relationships of characters are reflected in conversations. Thus, extracting conversations, speakers and other conversation elements from novels are crucial for character analysis and content understanding. We start with Jin Yong’s novels, annotate the largest Chinese corpus for speaker identification with 9721 quotes, and analyze language styles of different characters based on quotes. We then propose a machine learning-based speaker identification method, and design feature templates that show a good performance. For the application of speaker identification, we construct the social network of characters in Jin Yong’s novels based on dialog chain, which lays the foundation for the analysis of the relationship between characters in novels.

Development and Application of Leader Identification Model Using Multimodal Information in Multi-Party Conversations

Predicting the roles of participants in conversations is a fundamental task to build a system that provides assessment results and feedback for each participant. Various role recognition models have been proposed. Nonetheless, most studies have only utilized verbal or nonverbal features even though people usually express what they think or feel with the combination of language, gestures, and tone of voice. In this paper, we aim to realize a high-performance role recognition model by combining features from various modalities. We design nonverbal features that can be extracted from video and audio data. Then, we construct a multimodal leader identification method that fuses nonverbal features proposed by us and verbal features proposed by a previous study. In our experiments, our multimodal model outperforms the baseline model that utilizes only verbal features. We also conduct some analysis, such as statistical tests and ablation studies, and verify the effectiveness of each modality and feature. In the end, we build a prototype of a feedback system and demonstrate how our study can be applied to the discussion assessment/feedback systems.

Text Classification of Gujarati Newspaper Headlines

Text classification is an extremely important area of Natural Language Processing (NLP). This paper studies various methods for embedding and classification in the Gujarati language. The dataset comprises of Gujarati News Headlines classified into various categories. Different embedding methods for Gujarati language and various classifiers are used to classify the headlines into given categories. Gujarati is a low resource language. This language is not commonly worked upon. This paper deals with one of the most important NLP tasks - classification and along with it, an idea about various embedding techniques for Gujarati language can be obtained since they help in feature extraction for the process of classification. This paper first performs embedding to get a valid representation of the textual data and then uses already existing robust classifiers to perform classification over the embedded data. Additionally, the paper provides an insight into how various NLP tasks can be performed over a low resource language like Gujarati. Finally, the research paper carries out a comparative analysis between the performances of various existing methods of embedding and classification to get an idea of which combination gives a better outcome.

Supervised Word Sense Disambiguation on Polysemy with Neural Network Models: A Case Study of BUN in Taiwan Hakka

While word sense disambiguation (WSD) has been extensively studied in natural language processing, such a task in low-resource languages still receives little attention. Findings based on a few dominant languages may lead to narrow applications. A language-specific WSD system is in need to implement in low-resource languages, for instance, in Taiwan Hakka. This study examines the performance of DNN and Bi-LSTM in WSD tasks on polysemous BUNin Taiwan Hakka. Both models are trained and tested on a small amount of hand-crafted labeled data. Two experiments are designed with four kinds of input features and two window spans to explore what information is needed for the models to achieve their best performance. The results show that to achieve the best performance, DNN and Bi-LSTM models prefer different kinds of input features and window spans.