Page 129 - My FlipBook
P. 129
Brochure 2020
Figure 1 : Sentence Parsing Example.
4. Distributed Word Representation of applications utilizing this toolkit has exceeded 250,
Compared to traditional word meaning representation and more than 1,000 utilize the dictionary alone. We
via symbols, distributional word representation are currently working on a new effective supervision
provides additional computational ability and the method for sentiment and sarcastic tweets collection
advantage of generation, but lacks explanatory ability. to assist the research community and boost research
Thus, fully making use of the strength of each kind progress. We have also established international
of representation is crucial to resolving practical NLP cooperation with Israel and the US. We are now
tasks. We have studied how to infuse information investigating sentiment information model- and
into knowledge bases in the form of distributional system-wide for new emerging and vital research
word representation, and published our work in topics such as lie detection and fake news intervention,
EACL 2017 and IALP 2017, with this latter manuscript in which we anticipate making useful advances.
receiving a best paper award. We have also developed 2. Machine Reading:
a lexical sentiment analyzer using both E-HowNet Since most knowledge is nowadays expressed in text
and distributional word representation to predict form, machine reading has become a very important
sentiment for a given word or phrase. Our system won topic. We are building a Conversational Open Domain
third place overall in the 2017 IALP contest, and first Document-based Natural Speech Q&A system that can
place for prediction of phrasal arousal. read from Wikipages and then answer questions in
natural speech. We plan to start this long-term research
III. Natural Language Applications project by using a Chinese machine reading program
Our Chinese input system, GOING, is used by over one that can be evaluated with reading comprehension
million people in Taiwan. Our knowledge representation tests. Initially, we plan to focus on reading elementary
kernel, InfoMap, has been applied to a wide variety of school texts, and then gradually shift to high school
application systems. In the future, we will design event and real domain-oriented applications (e.g., smart
frames as a major building block for our learning system. Q&A).
We will also develop basic technologies for processing 3. Spoken language processing
spoken languages and music to support various Our recent achievements include discriminative
applications. autoencoders for speech/speaker recognition,
1. Sentiment Analysis and Opinion Mining subspace-based models for spoken language
We have studied opinions, sentiments, subjectivities, recognition, variational autoencoder-based
affects, emotions and views in texts such as news approaches for voice conversion, automatic speech
articles, blogs, forums, reviews, comments, dialogs, assessment models, and audio-visual speech
and short messages, and then developed sentiment enhancement. Our group member Wen-Chin Huang
analysis techniques for both Chinese and English. We won the best student paper award at ISCSLP2018.
have built one of the most popular Chinese sentiment Ongoing research includes Minnan speech recognition,
analysis toolkits, CSentiPackage, which includes cross-lingual speech recognition, and spoken question
sentiment dictionaries, scoring tools, and the deep answering.
neural network module UTCNN. Thus far, the number
127
Figure 1 : Sentence Parsing Example.
4. Distributed Word Representation of applications utilizing this toolkit has exceeded 250,
Compared to traditional word meaning representation and more than 1,000 utilize the dictionary alone. We
via symbols, distributional word representation are currently working on a new effective supervision
provides additional computational ability and the method for sentiment and sarcastic tweets collection
advantage of generation, but lacks explanatory ability. to assist the research community and boost research
Thus, fully making use of the strength of each kind progress. We have also established international
of representation is crucial to resolving practical NLP cooperation with Israel and the US. We are now
tasks. We have studied how to infuse information investigating sentiment information model- and
into knowledge bases in the form of distributional system-wide for new emerging and vital research
word representation, and published our work in topics such as lie detection and fake news intervention,
EACL 2017 and IALP 2017, with this latter manuscript in which we anticipate making useful advances.
receiving a best paper award. We have also developed 2. Machine Reading:
a lexical sentiment analyzer using both E-HowNet Since most knowledge is nowadays expressed in text
and distributional word representation to predict form, machine reading has become a very important
sentiment for a given word or phrase. Our system won topic. We are building a Conversational Open Domain
third place overall in the 2017 IALP contest, and first Document-based Natural Speech Q&A system that can
place for prediction of phrasal arousal. read from Wikipages and then answer questions in
natural speech. We plan to start this long-term research
III. Natural Language Applications project by using a Chinese machine reading program
Our Chinese input system, GOING, is used by over one that can be evaluated with reading comprehension
million people in Taiwan. Our knowledge representation tests. Initially, we plan to focus on reading elementary
kernel, InfoMap, has been applied to a wide variety of school texts, and then gradually shift to high school
application systems. In the future, we will design event and real domain-oriented applications (e.g., smart
frames as a major building block for our learning system. Q&A).
We will also develop basic technologies for processing 3. Spoken language processing
spoken languages and music to support various Our recent achievements include discriminative
applications. autoencoders for speech/speaker recognition,
1. Sentiment Analysis and Opinion Mining subspace-based models for spoken language
We have studied opinions, sentiments, subjectivities, recognition, variational autoencoder-based
affects, emotions and views in texts such as news approaches for voice conversion, automatic speech
articles, blogs, forums, reviews, comments, dialogs, assessment models, and audio-visual speech
and short messages, and then developed sentiment enhancement. Our group member Wen-Chin Huang
analysis techniques for both Chinese and English. We won the best student paper award at ISCSLP2018.
have built one of the most popular Chinese sentiment Ongoing research includes Minnan speech recognition,
analysis toolkits, CSentiPackage, which includes cross-lingual speech recognition, and spoken question
sentiment dictionaries, scoring tools, and the deep answering.
neural network module UTCNN. Thus far, the number
127
