Moderating a dialogue between Humans and Machine
University: Umeå Institute of Design
Duration: 1 week, spring 2018
Industries have invested heavily into natural language processing in the last decades. Since then, it has gotten more and more precise in various applications, reaching close to 99% comprehension of words you say. Yet, the idiosyncrasies of human dialogue seem to be a tough area natural language processing has to deal with. Humans speak with high variety, depending on to whom you talk to and in which region you live in. What if you enable computer devices (with different language settings) to play the Broken Telephone game (via the Google translate application)?
The modified version of the children’s game broken telephone starts with a given sentence which will be “verbally” passed from one device and to another via Google translate. Each device is enabled to listen and speak in a pre-selected language. For example, one device listens to Spanish and continues to output the Spanish sentence in Russian. The next device would pick up the Russian translated sentence and translates it again into another language and so forth. At the end, the last device would either output the correct given input or outputs a distorted input.
Google translate seem to have difficulties in outputting the correct sentence, when playing the modified version the broken telephone game (with use of multiple languages). It is interesting to see how inherently tricky human speech and voice as mediums are, as well as how unnatural and/or creative and outputs can sound.
Although some translations sounded out of place, Google translate was occasionally creating comprehensible sentences. Which led to the question, how it would be, when a human takes part in a conversation with the modified telephone game as a conversation partner. Would that human understand anything? How long would it take for the human to identify and realise that its Google translate who they are talking to and not an actual other human being. Questions like these moved the exploration further to the idea of forming dialogues between humans and machines.
The idea is to create conversational flows between humans and machines. In this project scope, humans are represented as the Amazon Turk workers hired on Amazon Mechanical Work Platform. Whereas the machine side is represented by the Google translate application, or rather the modified telephone broken game. Both parties come together online and converse with each other in turns. The dialogue is edited and moderated (formed) in a way, so the conversation still makes sense, allowing it to flow.
Experiments with the Google translate show for example that, the more languages you use in that game setup (6 instead of 3 languages), the more comprehensible sentences you would get for building the dialogue, so it appears.