Two authors explain how A.I. could learn it
One of the most challenging areas of A.I. is surely Natural Language Processing which is improving the ability of an Artificial Intelligence to use language to communicate, write, process unstructured data, analyze them, and translate.
However, the basic challenge of language understanding has yet to be solved.
2 knowledge systems
Language is riddled with ambiguity. Humans resolve ambiguities using context, which we establish utilizing clues from the speaker’s tone, preceding words, and sentences, the conversation’s general setting, and a basic understanding of the world. And we make questions when our intuitions and knowledge fail us. The process of determining context is simple for humans. However, defining the same process in a computable manner is more difficult than it appears.
There are generally two approaches to dealing with this issue.
One is knowledge-based systems that define each word’s role in a phrase and extract context and meaning. A huge number of features about language, context, and the world are used by knowledge-based systems. This data might come from a variety of places and must be calculated in a variety of methods.
Language analysis is made more reliable and understandable by using knowledge-based systems. However, the engineering of features, the creation of lexical structures and ontologies, and the development of software systems that brought all of these components together needed far too much human effort to keep using them. The manual effort of knowledge engineering was seen as an obstacle by researchers, who looked for alternate ways to deal with language processing.
Machine learning models, instead, are knowledge-lean systems that use statistical relationships to solve the context problem. This second approach processes massive corpora of text during training and fine-tunes their parameters depending on how words appear close to one another. The statistical relationships between word sequences, not the meaning behind the words, determine the context in these models. Naturally, the larger the dataset and the more diverse the samples, the better those numerical parameters can capture the range of ways words can occur next to each other.
The problem of computing meaning
Deep learning models can now generate article-length text sequences, answer science test questions, write software source code, and respond to basic customer service inquiries. The majority of these domains have progressed as a result of improved deep learning architectures and, more significantly, larger neural networks every year.
However, while increasingly deep neural networks can enhance specific tasks incrementally, they do not address the wider problem of natural language processing.
Some experts believe that by continuing to scale neural networks, the issues that machine learning encounters may eventually be solved. However, McShane and Nirenburg (authors of Linguistics for an AI Age) argue that more basic issues must be addressed.
Machine learning, according to McShane, who’s a cognitive scientist and computational linguist, must overcome numerous obstacles, the first of which is the lack of meaning.
“The statistical/machine learning (S-ML) approach does not attempt to compute meaning”, McShane said. “Instead, practitioners proceed as if words were a sufficient proxy for their meanings, which they are not. In fact, the words of a sentence are only the tip of the iceberg when it comes to the full, contextual meaning of sentences. Confusing words for meanings is as fraught an approach to AI as is sailing a ship toward an iceberg”.
Machine learning algorithms, for the most part, avoid dealing with the meaning of words by restricting the goal of expanding the training dataset. However, even if a big neural network can preserve coherence over a long period of time, it still doesn’t understand the meaning of the words it generates. That’s the case of GPT-3.
Because their parameters can’t capture the limitless complexity of daily life, all deep learning–based language models start to break as soon as you ask them a series of minor but linked queries. Adding more data to the problem isn’t a solution to the challenge of direct knowledge integration in language models.
Language-endowed intelligent agents (LEIA)
McShane and Nirenburg provide a strategy for addressing the “knowledge bottleneck” of natural language processing without resorting to pure machine learning–based methods that require massive amounts of data.
The concept of language-endowed intelligent agents (LEIA) is at the center of Linguistics for the Age of AI, and it has 3 fundamental characteristics:
- LEIAs are able to understand language’s context-sensitive meaning and navigate through ambiguous words and sentences.
- LEIAs can explain their ideas, activities, and decisions with their human counterparts.
- LEIAs, like humans, can learn throughout their lives as they interact with humans, other agents, and the environment. Lifelong learning eliminates the requirement for ongoing human effort to develop intelligent agents’ knowledge base.
LEIAs go through 6 phases of processing natural language, from determining the role of words in sentences through semantic analysis to situational reasoning. The LEIA can use these steps to resolve conflicts between multiple interpretations of words and phrases, as well as to integrate the sentence into the larger context of the environment in which the agent is functioning.
LEIAs assign confidence levels to their interpretations of language utterances and understand where their abilities and knowledge end. In such circumstances, they resolve ambiguity by interacting with their human counterparts (or intelligent agents in their environment and other available resources). As a result of these contacts, they are able to acquire new things and broaden their knowledge.
LEIAs turn sentences into text-meaning representations (TMRs), which are comprehensible and actionable definitions of each word in a sentence. LEIAs assess which language inputs need to be followed up on based on their context and aims.
LEIAs lean toward knowledge-based systems, but they also incorporate machine learning models into the process, particularly in the early stages of language processing while parsing sentences.
Does natural language understanding need a human brain replica?
Integrating knowledge bases, reasoning modules, and sensory input is one of LEIA’s core features. Currently, fields like computer vision and natural language processing have very little overlap.
Humans use their extensive sensory experience to fill in the gaps in their linguistic utterances in the real world. Humans continue to build mental models of one another, which they use to make assumptions and omit features in language. Any intelligent agent that communicates with us in our own language should have similar abilities.
Meanwhile, LEIAs do not need to replicate the human brain in order to achieve human-like behavior.
In Linguistics for the Age of AI, according to McShane and Nirenburg, mimicking the brain will not help AI achieve its aim of explainability. “[Agents] operating in human-agent teams need to understand inputs to the degree required to determine which goals, plans, and actions they should pursue as a result of NLU [Natural Language Understanding]”, they write.
In conclusion, the ability to understand Artificial Intelligence will increasingly get better over the years as well as its ability to simulate comprehension will appear to us something real and ever more similar to human understanding capabilities. But it will always be a simulation that could also deceive humans from what real humanity is.
Source venturebeat.com
