Whereas the December competition showed a focus on the general consistency of topics in chat, these new works dig deeper into aspects of tone, mood, and texture of speech in a chatbot. The most interesting aspect of this week's work is that it marks a departure in the foci of research. Issues include inconsistencies in the facts and logic of sentences and mindless repetition of phrases. Results from a massive competition among chatbots in December, at the NeurIPS conference on machine learning, revealed numerous problems with chatbots even when the best engineers are writing the code.
They may all help make conversations less annoying, but they also fall short of the grandest promises, which are couched in the anthropomorphic language of human "reason" and "emotion."Īs reported by ZDNet in February, the state of the art is pretty terrible in chatbots. Some of the papers introduce new data sets that may help future chatbot work they all claim some improvement over state-of-the-art benchmarks. The papers were prepared for the annual conference of the Association for Computational Linguistics, or ACL, which started Sunday in Florence, Italy.Ī sample of what might be better chat, from Facebook's new EmpatheticDialogues data set for chatbots. Last week, Facebook's artificial intelligence researchers, with the help of academics, unveiled multiple projects to give chatbots a variety of new qualities, ranging from being less repetitive to being more on-topic, to even displaying some semblance of emotion.
So low it shouldn't be too hard to add a bit more intelligence to this dismal technology. The bar has been set pretty low for chatbots, those computer programs that seek to engage a person in back-and-forth dialogue.
0 kommentar(er)
