![]() But the more evaluation by a diverse mass of people takes place, the more "human" the answers become. This is interesting because it involves a subconscious, human component - because what makes sense and what doesn't is subjective. does it fit from a logical and human understanding to what was written by the user, so does it make sense? Evaluators should rely on their common sense. Two questions are asked when evaluating conversations. Google has defined two standards: Sensitivity and Specificity. What makes it special is not only the amount of data, but also the criteria used by human "crowdworkers" to evaluate her answers. The model was trained with 341 GB of open data based on social media conversations. ![]() What makes Meena different from other chatbots? What approaches to these two problems does the model take? User: How do we solve the climate crisis? Mitsuku: I would try looking at all the clues and maybe ask someone who has solved them before to help.Įxample 1: This answer makes sense but is not very context-specific source: Interview with chatbot Mitsuku. A real person, on the other hand, can give meaningful answers that make sense and are context-specific. ![]() By using very generic answers like these, the chatbot avoids producing utterances that are completely off-topic. Phrases like "Cool!" or "I don't know" fit very different inputs but are usually not really meaningful. If answers are too unspecific, it disturbs the flow of the conversation or seems unnatural. At the same time, an answer is not necessarily "good" in human understanding just because it makes sense. However, this does not necessarily mean that the answer is good: existing open-domain chatbots struggle with the problem that their answers often don't make sense: utterances aren't always consistent with what they said before throughout the conversation or they lack world knowledge and general understanding of relations. This means that the user can write or ask questions about anything and the CA provides an answer. Like Mitsuku, Meena is a so-called open-domain chatbot. In January 2020, Google has published an article on their developed model, the chatbot Meena, in which they describe an approach to how chatbots could become even more human. Until last year, leading English-speaking representatives were mainly two Conversational Agents (CA): Steve Worswick's Mitsuku and Microsoft's Zo, which went offline in summer 2019. One form of chatbots are so-called virtual companions (or "conversational agents"), which are able to conduct conversations on a wide variety of topics. This could now change with Google's chatbot Meena, as this model sets new standards in human-digital conversations. A few subtypes of chatbots are already capable of generalist conversations, but they are often very error-prone - on the one hand, they lack large amounts of data for "good" conversations, on the other hand, unlike people, they have no cultural sensitivity. However, such chatbots are usually limited to their area of responsibility and are not designed for general conversation. This model could be the next important step for chatbots and natural language user interfaces.Ĭhatbots provide airlines with information about bookings or flight dates, they suggest recipes or take pizza orders. The voice interaction with so-called conversational agents is becoming increasingly human. With two simple criteria, Google's chatbot Meena sets new standards in human-digital conversations.
0 Comments
Leave a Reply. |
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |