Introduction to Translation and Interpreting Studies by Aline Ferreira & John W. Schwieter

Author:Aline Ferreira & John W. Schwieter [Ferreira, Aline & Schwieter, John W.]
Language: eng
Format: epub
ISBN: 9781119685326
Published: 2022-06-28T09:38:04+00:00

Automatic evaluation: testing the engine with a bilingual corpus that has not been used in the training process. This generates scores in various metrics (see Section 7.3.3).

Retraining the engine (by adding new resources or corpora) and tuning it (by adjusting the weights of the different parameters)

After the system has been trained, decoding takes place to produce the most probable translation for the various phrases.

Both statistical and neural machine training (NMT) are machine learning systems: they are not explicitly programmed to follow pre-defined rules. Instead, they are trained to find solutions “autonomously” by being fed huge amounts of data.

Since the second half of the 2010s, NMT has rapidly become the predominant paradigm, particularly in light of the increased sense of fluency that this technology produces and the feeling that SMT may have plateaued.

Neural machine translation . NMT is an example of deep (machine) learning. One of the differences between SMT and NMT is that whereas the former operates on phrases, the latter works as a next-word predictor on the basis of the co-text of each sentence. SMT tackles the translation problem by breaking it up into various components or “subproblems” and then averages the probabilities given by each component. In contrast, NMT looks at the whole problem all at once and attempts to represent a source sentence (a process called encoding) and then to sequentially produce a target sentence (decoding) based on that representation. Unlike SMT, it does not need frequent human supervision to adjust its components – in particular, the weights of each model or input data – to improve performance.

How does NMT work? As the term suggests, it uses an artificial neural network (ANN), a system that loosely resembles the neurons in our brains and their multiple interconnections. An ANN contains a set of data-processing “neural” units that are connected to each other with varying positive or negative weights (or intensities). Neurons undergo a variable process of activation, that is, they are activated with more or less strength, depending on inputs, interconnected neural units, and how related the information that each neuron processes is to the other units, which is a consequence of previous learning or training.

Neurons are usually grouped into layers: at least an input layer (e.g., for each word in the source segment); an output layer; and, in the middle, one or more hidden layers, each specializing in certain characteristics or processes. They are called “hidden” because this is where the “black-box magic” happens, where all the transformations take place without our knowing exactly how.

In NMT, there are two main processes: encoding and decoding, performed, respectively, by an encoder and a decoder. In the encoder, source and target language words are first converted to a vector representation (a series of numbers) through training with monolingual corpora. These representations are called word embeddings and can be visualized either as a long series of numbers or as a coordinate in a multidimensional vector space. Words that are closer together in this space share similar linguistic, pragmatic, or any other

Download

Copyright Disclaimer:
This site does not store any files on its server. We only index and link to content provided by other sites. Please contact the content providers to delete copyright contents if any and email us, we'll remove relevant links or contents immediately.