Indirect Supervised Learning of Strategic Generation Logic

The Strategic Component in a Natural Language Generation (NLG) system is responsible for determining content and structure of the generated output. It takes a knowledge base and communicative goals as input and provides a document plan as output. The Strategic Generation process is normally divided into two subtasks: Content Selection and Document Structuring. An implementation for the Strategic Component uses Content Selection rules to select the relevant knowledge and Document Structuring schemata to guide the construction of the document plan. This implementation is better suited for descriptive texts with a strong topical structure and little intentional content. In such domains, special communicative knowledge is required to structure the text, a type of knowledge referred as Domain Communicative Knowledge. Therefore, the task of building such rules and schemata is normally recognized as tightly coupled with the semantics and idiosyncrasies of each particular domain. In this thesis, I investigate the automatic acquisition of Content Selection rules and the automatic construction of Document Structuring schemata from an aligned Text-Knowledge corpus. These corpora are a collection of human-produced texts together with the knowledge data a generation system is expected to use to construct texts that fulfill the same communicative goals as the human texts. They are increasingly popular in learning for NLG because they are readily available and do not require expensive hand labelling. To facilitate learning I further focus on domains where texts are also abundant in anchors (pieces of information directly copied from the input knowledge base). In two such domains, medical reports and biographical descriptions, I have found aligned Text-Knowledge corpus for my learning task. While aligned Text-Knowledge corpora are relatively easy to find, they only provide indirect information about the selected or omitted status of each piece of knowledge and their relative placement. My methods, therefore, involve Indirect Supervised Learning (ISL), as my solution to this problem, a solution common to other learning from Text-Knowledge corpora problems in NLG. ISL has two steps; in the first step, the Text-Knowledge corpus is transformed into a dataset for supervised learning, in the form of matched texts. In the second step, supervised learning machinery acquires the CS rules and schemata from this dataset. My main contribution is to define empirical metrics over rulesets or schemata based on the training material. These metrics enable learning Strategic Generation logic from positive examples only (where each example contains indirect evidence for the task).