This paper describes a prototype system capable of generating text tailored to a particular listener’s interests or sympathies. The system requires information about each potential listener’s social status, interests, sympathies, antipathies, and emotional state as part of its knowledge base. A story representation is then given as input to the system which generates different text for different hearers. The text can be focused on either a particular listener’s interests or sympathies.

The paper includes a good example of the system in use. The example story, one used elsewhere in conceptual dependency-oriented research, is about an act of terrorism carried out by the IRA, in which a British soldier and a female bystander are shot and killed, and a 12-year-old girl is injured. The system then generates text for the IRA terrorist, the terrorist’s wife, a British soldier, and a neutral American by first focusing on each listener’s interests and then on his or her sympathies.

The system selects words which will best play on the emotions of a particular listener. In the example, the IRA man is referred to as a terrorist in the British version, a freedom fighter in the IRA version, and a gunman for the neutral American. The system orders the sentences of the text in such a way that the particular listener’s interests and/or sympathies are presented first. In addition to word selection and sentence order, the system may decide to expand on or omit certain information depending on a particular listener’s interests or sympathies. For example, the text generated for the British soldier describes the gun used in great detail since the soldier has an interest in guns. The other versions do not describe the gun in detail, reflecting the lack of interest on the part of the other listeners.

One problem associated with this system, which is not addressed by the authors, is the collecting of the required data on each listener’s interests, sympathies, antipathies, and emotional state. This data is different not only for each listener but for each listener in each situation. In the example, the emotional state of the American is considered neutral and relaxed. However, if the woman shot and killed in the story happened to be the American’s vacationing wife, his emotional state would be anything but relaxed and neutral. This situation-dependent analysis is required for each text generation problem and may be difficult to automate. This leads to a second problem, namely: How does one go about objectively assigning the interests of particular listeners? The interest of an IRA terrorist may be “British losses” from a neutral American’s point of view, but from an IRA terrorist’s viewpoint IRA interests are “freeing Ireland from English oppression.” A British observer may offer a third viewpoint by assigning IRA interests to “the destabilization of British-Irish relations.” The problem occurs when trying to assign objective values in subjective situations.

The paper raises many questions in the area of custom tailored text generation and is a good introductory study of the subject. It is recommended reading, but keep in mind the unaddressed problems in such a system.