Negotiating conversational agents

Negotiating is an essential process found in all aspects of day-to-day life, from complex legal disputes to merely discussion weekend plans. The aim of any negotiation is to reach an outcome which all parties are content with, often requiring one or more parties to compromise to reach a suitable solution. However, a mutually beneficial outcome is only one of many, and the observed outcome may differ from the optimal due to the ability of one party to persuade the others.

To effectively persuade, using logic alone is not sufficient; agents must combine logical reasoning with an understanding of human nature and how emotions can affect a person's perspective and their subsequent actions. The most mutually beneficial outcome relies on perfect information between all parties. However, only each person truly knows what they desire, resulting in asymmetric information despite best intentions. These
needs and wants continually shift over the course of the interaction, and to succeed, agents must continually evolve their strategy as the negotiation progresses.

Negotiation strategies are not clearly defined - they are not mutually exclusive and are often used simultaneously alongside at least one other. During a negotiation, agents will likely evolve their strategy over time due to their perceived progress and their perceptions of the other parties. Therefore, the subsequent actions taken are often a result of how an agent has perceived both overt and subtextual information available,
therefore implying a more successful negotiator likely has a greater wealth of experience to draw from. However, despite seemingly endless inputs and countless intermediary states, the result of any negotiation must end in one of two distinct outcomes: the proposal is either accepted or rejected - a process akin to a classification task.

Along with considering how facts and rhetoric are represented, this thesis will also explore the impact emotion has on the natural language used within negotiations and when attempting to persuade another person. From modelling from both closed-environment negotiations and real-world examples of persuasion attempts, agents will be evaluated to explore what features draw its attention and what information can be discovered from
probing its hidden states. Further analysis can explore how a given predisposition towards certain strategies can affect the overall success of the agent in a negotiation and how agents could persevere when faced with adverse situations.

The scope of this thesis is limited to understanding natural language from a strictly text-based standpoint; minimising the introduction of confounding variables from facial expressions, voice tonality, and more. However, this limitation does not impede this thesis from contributing to existing and future work regarding affective understanding using multimodal inputs or across other domains.