Human-machine interaction has always been about making the life simpler and easier for humans. However, machines misinterpret words of human and cause trouble, let alone make their lives easier. Virtual assistants are one of the best examples in which the machine accepts the voice inputs from humans and perform necessary actions such as booking a table at a restaurant or telling which movie of their favorite actor will be releasing soon. The extensive research has been conducted by researchers and scientists at technology companies and institutes to make virtual assistants available in multiple languages and understand the local dialect better to perform the correct task. However, there have been limitations to the understanding of the words as the words used in idioms and phrases may be misunderstood and unexpected actions would be carried out.
If the words have multiple meanings and different meaning in different context, it will be difficult for a machine to understand the context and give required output. For instance, the word “king” can be used in different contexts such as Chess, the one who is the king of an empire, or an expression to be used to convey affection. If we say while playing Chess, “I’m coming for your king, beware!”, this is the context of the Chess piece. If we say, “The King Felipe VI of Spain will address the people by 12 noon”, this refers to the king of Spain. This is what machines could not understand. Because the word “king” has been used in different contexts. The ability of words used in different contexts depicting different meanings refers to polysemy. This is the rule. The meaning of the word is determined by the phrasing with which it is used.
Getting back to the machine learning of these words. The extensive research would make those systems flexible to the words and depict the meaning based on the context. A new system known as ELMo (Embeddings from Language Models) has been invented. This system determines the context of the word and leads systems to better understanding. It understands the polysemy with ease. The training data is used to identify if the word has multiple meanings and how those meanings are used in the language. For instance, in determining the meaning of the word “king”, the system would determine the context. The basic assumption of compressing a single word into single vector has been questioned and the system compressed a single word into an infinite number of vectors.
The paper published on this system was awarded the best paper honors at North American Chapter of the Association for Computational Linguistics (NAACL) last week. Mathew Peters, the lead author of the paper, said, “We were looking for a method that would significantly reduce the need for human annotation. The goal was to learn as much as we can from unlabeled data.”
ELMo would try to understand the context from the entire sentence in which the word has appeared. Moreover, as the whole sentence is considered, it eases up the mapping of the structure and labeling clauses along with parts of speech. Though the system will not be as efficient as the human with years of experience in parsing language, it has improved natural language algorithms by nearly 25 percent. This marks a significant milestone in the field. Furthermore, this system is easy to integrate into existing commercial systems. Microsoft has already begun using the algorithm into Bing. It helps in reading the query accurately and responding as per requirement. The good news is ELMo is open source. So, the companies with the requirement of natural language processing should check this system out.