--- task_categories: - text-classification language: - de pretty_name: Intent Classification for Robot Assisted Disaster Response size_categories: - 100K This dataset consists of conversations recorded during the training sessions in the emergency response domain. The conversations are typically between several operators controlling the robots, a team leader and a mission commander. The data have been transcribed and annotated during the following projects: [TRADR](http://www.tradr-project.eu/) and [ADRZ](https://rettungsrobotik.de/home). The dialogues are split into turns and each turn is annotated with a speaker and intent. ## Dataset Details ### Dataset Description - **Curated by:** DFKI, [Talking Robots Group at MLT](https://www.dfki.de/en/web/research/research-departments/multilinguality-and-language-technology/tr-team) - **Language(s) (NLP):** German - **License:** [More Information Needed] ## Dataset Structure ### Data Instances ``` { 'id': '1235', 'speaker': 'UAV', 'text': 'wir haben einmal den Akku gewechselt, bis jetzt noch kein Rauch festzustellen ...', 'label': 2 } ``` ### Data Fields ``` id: the id of the dialogue turn, an `int` feature speaker: the speaker of the turn, a `string` feature text: the utterance of the turn, a `string` feature label: the label of the turn, an `int` feature ``` ### Data Splits This dataset contains 3525 dialogue turns in total. The data are split as follows: 2610 turns for training, 310 for development and 605 for test. The data represent a continuous conversation, i.e., the previous id refers to the previous turn in the dialogue. ### Label Description and Statistics | label | meaning | train | percentage | example | | --- | --- | --- | --- | --- | | 0 | disconfirm | 35 | 1.3% | `Ist negativ, noch nicht.` | | 1 | order | 216 | 8.3% | `Für Sie Erkundungsauftrag: Gesamtüberblick über die Einsatzstelle. Kommen.` | | 2 | info_provide | 979 | 37.5% | `Ich verlasse das Erdgeschoss und gehe ins erste Obergeschoss.` | | 3 | info_request | 238 | 9.1% | `Frage: Erkundungsergebnis aus der östlichen Seite des Gebäudes, kommen.` | | 4 | call | 487 | 18.7% | `RobLW an Zugführer, kommen.` | | 5 | call_response | 370 | 14.2% | `Ja, hier ist Zugführer, kommen.` | | 6 | other | 43 | 1.7% | `Einen Augenblick, ich melde mich gleich.` | | 7 | confirm | 242 | 9.3% | `Ein Lagebild von oben, komplette Lage, und ein Lagebild zwischen den beiden Türen, verstanden.` | ## Dataset Creation ### Curation Rationale The dataset is based on the recordings from the emergency response domain that use radio communication protocol. The goal of the conversation is to coordinate rescue operations in a robot-assisted disaster response. ### Source Data The data are based on human-human communication in robot-assisted disaster response. The dialogues are task-oriented, focused on collaborative execution of a mission by a team that uses robots to to explore some area, find hazardous materials, locate fires, damage or victims. #### Data Collection and Processing The initial audio recordings were collected during the [TRADR](http://www.tradr-project.eu/) and [ADRZ](https://rettungsrobotik.de/home) projects, transcribed and annotated by the [Talking Robots Group, DFKI](https://www.dfki.de/en/web/research/research-departments/multilinguality-and-language-technology/tr-team) ### Annotations The annotations include dialogue intents relevant for communication in the emergency response domain: `call`, `call_response`, `info_request`, `info_provide`, `confirm`, `disconfirm`, `order` and `other`. Note the interpretation of the intent depends on the context. E.g., the following examples illustrate how very similar responses ("Warten", "Wait") are annotated differently depending on the previous turn: ``` (1) disconfirm - Können wir weitermachen? (Shall we continue?) - Warten. (Wait.) (2) confirm - Hast du die Möglichkeit, das Fass näher zu identifizieren, was da drin ist? (Can you inspect the barrel closer to identify what is inside?) - Ja, warten. (Yes, wait.) (3) order - Werde aber jetzt auch mal die rückwärtige Seite des Fasses erkunden. (I will inspect now the back side of the barrel.) - UGV 1, damit warten. (UGV 1, wait.) (4) other (pausing to check) - Frage: kommen meine Fotos an? (Question: do you receive my photos?) - Warten. (Wait.) ``` #### Annotation process The recordings were manually transcribed and annotated with emergency response intents. There are 3525 dialogue turns in total with 6.3 tokens per turn on average. #### Who are the annotators? All annotations were done by the research assistants of the [Talking Robots Group, DFKI](https://www.dfki.de/en/web/research/research-departments/multilinguality-and-language-technology/tr-team) #### Personal and Sensitive Information The dataset does not include any real names, addresses or other personal information. The recordings were done during training sessions with simulations of the emergency situation. ## Bias, Risks, and Limitations The dataset covers only a subset of possible emergency situations, focusing mainly on fire, building collapse and chemical leakage. It does not address many other situations, e.g., traffic accidents, floods or explosions. ## Citation Part of this dataset has been introduced in the following paper. However, the current version includes more annotated turns due to additional data collection. **BibTeX:** ``` @inproceedings{anikina-2023-towards, title = "Towards Efficient Dialogue Processing in the Emergency Response Domain", author = "Anikina, Tatiana", editor = "Padmakumar, Vishakh and Vallejo, Gisela and Fu, Yao", booktitle = "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 4: Student Research Workshop)", month = jul, year = "2023", address = "Toronto, Canada", publisher = "Association for Computational Linguistics", url = "https://aclanthology.org/2023.acl-srw.31", doi = "10.18653/v1/2023.acl-srw.31", pages = "212--225", abstract = "In this paper we describe the task of adapting NLP models to dialogue processing in the emergency response domain. Our goal is to provide a recipe for building a system that performs dialogue act classification and domain-specific slot tagging while being efficient, flexible and robust. We show that adapter models Pfeiffer et al. (2020) perform well in the emergency response domain and benefit from additional dialogue context and speaker information. Comparing adapters to standard fine-tuned Transformer models we show that they achieve competitive results and can easily accommodate new tasks without significant memory increase since the base model can be shared between the adapters specializing on different tasks. We also address the problem of scarce annotations in the emergency response domain and evaluate different data augmentation techniques in a low-resource setting.", } ``` **APA:** ``` Anikina, T. (2023). Towards Efficient Dialogue Processing in the Emergency Response Domain. Annual Meeting of the Association for Computational Linguistics. ``` ## Glossary Abbrevations used for the speakers: UGV: Unmanned Ground Vehicle UAV: Unmanned Aerial Vehicle MC: Mission Commander TL: Team Leader RobLW: Robotikleitwagen (robotic lead vehicle) ZF: Zugführer (fire brigade commander) GF: Gruppenführer (group leader) ELW: Einsatzleitwagen (emergency command vehicle) GW-DUK: Gerätewagen-Daten-und-Kommunikation (vehicle for transporting robots and equipment)