Context-Aware Module Selection in Modular Dialog Systems

Proceedings of Recent Advances in Natural Language Processing, pages 785–791 Varna, Sep 4–6, 2023 https://doi.org/10.26615/978-954-452-092-2_085 785 Context Aware Module Selection in Modular Dialog Systems Jan Nehring German Research Center for Artificial Intelligence (DFKI) Alt-Moabit 91c 10559 Berlin, Germany jan.nehring@dfki.de Ren´e Marcel Berk, Stefan Hillmann TU Berlin Straße des 17. Juni 135 10623 Berlin, Germany firstname.lastname@tu-berlin.de Abstract In modular dialog systems, a dialog system consists of multiple conversational agents. The task “module selection” selects the appropriate sub-dialog system for an incoming user utter- ance. Current models for module selection use features derived from the current user turn only, such as the utterances text or confidence values of the natural language understanding systems of the individual conversational agents, or they perform text classification on the user utterance. However, dialogs often span multiple turns, and turns are embedded into a context. Therefore, looking at the current user turn only is a source of error in certain situations. This work pro- poses four models for module selection that include the dialog history and the current user turn into module selection. We show that these models surpass the current state of the art in module selection. 1 Introduction Dialog systems (DS) often consist of multiple sub- dialog systems or modules. There are multiple rea- sons for such a combination: The designer of a DS might want to combine several existing DS without a reimplementation. Sometimes a DS spans multi- ple departments and cannot be merged into a single, unified system. A hybrid system is a possible solu- tion when a DS consists of multiple incompatible subsystems, e.g., a task-oriented DS and a question- answering system. Although this architecture is frequently used in practical applications, it is a gap in scientific research. The modular dialog system (MDS) (Nehring and Ahmed, 2021) describes a framework to combine several dialog systems. In an MDS, a central com- ponent called “module selection” (MS) selects the appropriate sub-DS that generates the answer for an incoming user utterance (Nehring et al., 2023). MS is a classification task to choose one sub-DS from a list of sub-DS for a given user utterance. Figure 1: Example dialog between a user and a modular dialog system that consists of two chatbots. Based on the current utterance, the module selection can easily select the proper agent for the first user utterance. How- ever, the module selection requires dialog context to classify the second user utterance correctly. Current solutions for MS, such as G¨orzig et al. (2023) or Nehring et al. (2023), focus on the current user utterance only. They use the text of the user utterance, confidence values of the models NLUs, or additional features such as detected named enti- ties for the models. However, both works showed that the text of the user utterance is the essential feature for high performance in MS (G¨orzig et al., 2023; Nehring et al., 2023). In some cases, more than the current user utterance and other derived features are needed to find the appropriate sub-DS. Figure 1 shows …

Content truncated to 500 words