International Symposium on Chinese Spoken Language Processing
August 23-24, 2002
Challenges and Advances in Semantic Representation and Interpretation
Department of Electrical Engineering, National Cheng Kung University
One of the greatest challenges in developing natural language understanding lies in semantic
representation and interpretation (SRI). Semantic representation concerns how elements of sentences
represent semantic constituents and what kind of relationships semantic constituents are. Semantic
interpretation concerns how to interpret these constituents and relationships semantically.
The field of semantic representation and interpretation has witnessed a number of significant
advances in the past years. The traditional methods include first-order predicate calculus (FOPC), case
frames and grammars, and logical form (LF). Recently, unified natural language (UNL), flat and deep
ABox, ontologies, and latent semantic acquisition (LSA) have been developed for the shared formal
conceptualizations of particular domains. These methods, as specifications of the concepts in a given
field, and of the relationships among those concepts, provide insight into the nature of information
produced by that field and are an essential ingredient for any attempts to arrive at a shared
understanding of concepts in a field.
However, some design issues must be considered to achieve a good SRI. For this reason, this
lecture introduces four evaluation criteria – scope portability, accuracy, facility, and efficiency (SAFE).
After comparing previous SRI techniques under SAFE, we find several methods are deficient in scope
portability and efficiency, while others are lack of facility or accuracy. And these drawbacks confuse
new researchers and various computer systems when dealing with natural language understanding.
For these drawbacks, we have proposed a devised Acting Role Table (ART) and an efficient
automatic ART construction algorithm for language understanding. The semantics of each sentence is
represented in the devised Acting Role Table. The ART consists of acting roles of the sentences, i.e.,
action, agent, instrument, theme, location, and time, together with their associated modifiers. In this
step, we use verb-driven syntax analysis to determine the acting roles in the sentences; and use the
semantic analysis to constrain the acting roles based on the features defined in well-known knowledge
database e.g. WordNet and HowNet. In addition, we applied ART in question answering for primaryschool
textbook, spoken dialogue system for mobile information retrieval, template expansion for
example-based machine translation and language translation for travel planning and illustrate a good
direction for language understanding.
James Allen (1995) NATURAL LANGUAGE UNDERSTANDING. Reading, The Benjamin/
Cummings Publishing Company, Inc.
- Wolfgang Wahlster (2000) Verbmobil: Foundation of Speech-to-Speech Translation. Reading, Springer.
- Daniel Jurafsky and James H. Martin (2000) Speech and Language Processing. Reading, Prentice Hall.
- Jhing-Fa Wang and Hsien-Chang Wang, J-Y Huwang(2000) Domain-unconstrained language
understanding Based on CKIP-AutoTag, How-net, and ART. 6th International Conference of Spoken
- Jhing-Fa Wang and Hsien-Chang Wang and Chin-Nan Lee (2000) Domain Unconstrained Language
Understanding Based on How-net. PACLIC 14, Japan.
- Jhing-Fa Wang and Shun-Chieh Lin (2002) Bilingual Corpus Evaluation and Discriminative Sentence
Vector Expansion for Machine Translation. ICAIET 2002, Malaysia.
Wang, Jhing-Fa (2002):
"Challenges and advances in semantic representation and interpretation",
In ISCSLP 2002, paper INV4 (abstract).