Słownik AI
Kompletny słownik sztucznej inteligencji
Machine Translation (MT)
Computational process that automatically translates text from a source language to a target language without human intervention, using algorithms and linguistic models.
Neural Machine Translation (NMT)
Machine translation approach that uses deep neural networks to directly model the conditional probability of translation, surpassing traditional statistical systems.
Statistical Machine Translation (SMT)
Translation method based on statistical models learned from large bilingual corpora, primarily using probabilistic language and translation models.
Attention Mechanism
Technique that allows NMT models to selectively focus on different parts of the source text when generating each target word, significantly improving translation quality.
Transformer Architecture
Neural network architecture entirely based on attention mechanisms, eliminating recurrence and convolutions, becoming the standard for modern NMT systems.
BLEU Score
Automatic evaluation metric for machine translation that compares generated translations to human references by measuring n-gram precision.
Parallel Corpus
Collection of bilingual texts where each sentence in the source language is aligned with its corresponding translation in the target language, essential for training MT systems.
Back-Translation
Data augmentation technique where a trained NMT model translates target monolingual texts to the source language, creating synthetic data to improve performance.
Zero-shot Translation
Capacity of multilingual models to translate between language pairs never seen during training, using a pivot language as an implicit intermediary.
Subword Segmentation
Process breaking down words into smaller units (BPE, WordPiece) to handle open vocabulary and improve generalization on rare or unknown words.
Beam Search
Decoding algorithm simultaneously exploring multiple partial translation hypotheses, selecting the most probable at each step to balance diversity and quality.
Post-editing
Human review of machine translations to correct errors and improve quality, combining automation efficiency with human linguistic expertise.
Multilingual NMT
Single system capable of translating between multiple language pairs, sharing parameters to benefit from inter-language knowledge transfer and reduce complexity.
Language Model
Probabilistic model assigning probabilities to word sequences, used in NMT as a decoding component to evaluate fluency and coherence of generated translations.
Encoder-Decoder
Neural architecture composed of an encoder processing source text into vector representation and a decoder sequentially generating the target translation.
Byte-Pair Encoding (BPE)
Data compression algorithm adapted for subword tokenization, creating an optimal vocabulary by iteratively merging the most frequent symbol pairs.