The LTI is proud to announce the following PhD Thesis Defense:
Linguistically Informed Language Generation: A Multi-faceted Approach
Eduard Hovy, (Chair)
Alan W Black
Jason Weston, (Facebook AI Research / New York University)
Dan Jurafsky, (Stanford University)
Natural language generation (NLG) is a key component of many language technology applications such as dialogue systems, like Amazon’s Alexa; question answering systems, like IBM Watson; automatic email replies, like Google’s SmartReply; and story generation. NLG is the process of converting computer-internal semantic representations of content into the correct form of a human language, such as English or Korean, so that the semantics are accurately included. One might think that the only information an NLG system would need to produce is that contained explicitly in the utterance.
However, there is a multitude of implicit information not explicitly obvious on the surface. For instance, many different surface sentences have the same meaning but still have slightly different surface outputs. Several kinds of parameters seem to be reflected in variations of an utterance: external knowledge, goals, interpersonal information, speaker-internal information, and more. In this work, we call each individual input parameter a facet. To generate appropriate and natural utterances as humans do, appropriate modeling of these facets is necessary, and the system needs to be effectively guided by these facets.
One of M. Halliday’s linguistic theories, called Systemic Functional Linguistics (SFL), suggests that such parameters could be categorized into three meta-functions, where each contains separate types of information relevant to aural and written communication. We choose three facets of interest, one for each SFL meta-function, and repackaged them into three facet groups: knowledge, structures, and styles. The knowledge facet decides the basic semantics of the topic to be communicated, while the structure facet coherently arranges information guiding the structure of the (multi-sentence) communication. Finally, the style facet represents all the other additional kinds of information that direct the formulation of
We assume that the three facets are, more or less, individual, and they dynamically interact with each other instead of being a sequential process. One can develop a human-like NLG system that effectively reflects the three facets of communication and that simultaneously interact with each other, making a multifaceted system. In such systems, we believe that each facet of language has its own communicative goal, such that the knowledge facet is used to achieve factual goals, the structure facet is used to achieve coherence goals, and the style facet is used to achieve social goals.
To show the necessity and effectiveness of the multifaceted system, we have developed several computing methods for each facet from the following questions: Q1 “What” knowledge must be processed to make the model produce more factual text? Q2 “How” can the model compose multiple sentences coherently? Q3 How can the model produce stylistically appropriate output depending on “who” you are and “whom” you are talking to?
For a copy of the defense thesis please go to the following link.