The U.S. government spent 4.5 billion USD from 1990 through 2009 on outsourcing translations and interpretations [BIBCITE%%0%]. If these translations were automated, much of this money could have been spent elsewhere. The research field of Machine Translation (MT) tries to develop systems capable of translating verbal language (i.e. speech and writing) from a certain source language to a target language.
Because verbal language is broad, allowing people to express a great number of things, one must take into account many factors when translating text from a source language to a target language. Three main difficulties when translating are proposed in [BIBCITE%%1%]: the translator must distinguish between general vocabulary and specialized terms, as well as various possible meanings of a word or phrase, and must take into account the context of the source text.
Machine Translation systems must overcome the same obstacles as professional human translators in order to accurately translate text. To try to achieve this, researchers have had a variety of approaches over the past decades, such as [BIBCITE%%2%]. At first, the knowledge-based paradigm was dominant. After promising results on a statistical-based system ([BIBCITE%%3%]), the focus shifted towards this new paradigm.
[Could not find the bibliography file(s)
Note: This is an essay I wrote for the subject Philosophy of Cognitive Science that was part of my bachelor’s course. I think it might be interesting to others, so I’ve decided to publish it here. The format is adapted slightly to be more suitable for this blog; the content is unchanged.
In the field of artificial intelligence, humans are often used as prime examples of adaptable agents with general intelligence. The goal of some artificial intelligence researchers is to arrive at an artificial general, or human-level, intelligence. These agents should be able to perform many of the same tasks with the same adaptability as humans are able to. One of the few empirical certainties in the endeavour of creating such intelligent agents is that the natural, human intelligence works. Thus, there is merit to artificial intelligence research that strives to mimic human intelligence by modelling human mechanisms.
An intriguing and far-from-settled debate concerns the origin of human knowledge, skills, abilities and thought in general. The major theories can be identified as lying somewhere between the two extremes of full-blown nativism and full-blown empiricism [BIBCITE%%4%]. Nativistic theorists would argue for innate knowledge; at least some of our capabilities arise from hard-wired pathways in our nervous system that are available at birth. In contrast, empiricists would argue that these capabilities are learned from experience utilizing the brain’s various capacities for learning. For example, a baby’s suckling after birth is likely innate, whereas the behavioural pattern of brushing your teeth is likely learned. It is still unknown which combination of these extremes in this seemingly easy distinction is correct.
When striving to model human capacities in an artificial intelligence, knowing which parts of human intelligence and other capabilities are hard-wired and which parts arise from experiences should be of particular interest to artificial intelligence researchers. In the following, we will look at the innateness (or lack thereof) of language parsing and acquisition. From this, recommendations will be made regarding the high-level design of an artificial natural language processor.
An artificial neural network (ANN) is a type of machine learning model. It is made up of a number of simple parts called units, or neurons. By combining a large amount of these simple units, ANNs can solve real-world problems. For example, the main network that was used in my bachelor thesis research consisted of over 12,000 units. The name artificial neural network is slightly misleading: they’re mostly related to biological neural networks through the fact that both artificial and natural neural networks are made up of simple parts. Other than that they’re quite unrelated.
During the training of several neural networks for my bachelor’s thesis (more on that later, maybe!) I noticed something fun. The used networks’ weights (in this case classification function parameters) are initialized with numbers drawn from the standard normal distribution, meaning the initial network state is random. Such randomness by its very nature has no actual structure, and thus has high entropy. This means that compressing the information to save it to disk is less effective than on other, more structured, information. Initially, saving one such network’s weights required approximately 19.5MB of disk space.
As the networks’ training progressed, the file sizes shrunk! After a day of training, the space required for saving this network’s weights had decreased to 18.0MB; a 1.5MB decrease from the original value. I hadn’t thought about it before, but once I noticed it I soon realized what was happening. The whole act of training networks is exercised to find structure in data. A neural network does this by learning some sort of representation of the data through continuously updating its weights while training — in other words, the weights are getting more structured as the network is getting smarter! When the weights’ structure increases the entropy decreases, making compression more effective and our disks happier. Or unhappier, perhaps.