Minimum Error Rate Training in Statistical Machine Translation

By Franz Joseph Och. 2003

3,328 citations recorded by the source corpus.

Often, the training procedure for statistical machine translation models is based on maximum likelihood or related criteria. A general problem of this approach is that there is only a loose relation to the final translation quality on unseen text. In this paper, we analyze various training criteria which directly optimize translation quality. These training criteria make use of recently proposed automatic evaluation metrics. We describe a new algorithm for efficient training an unsmoothed error count. We show that significantly better results can often be obtained if the final evaluation criterion is taken directly into account as part of the training procedure.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition, Natural Language Processing, Optimization & Training

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Computer Vision and Pattern Recognition

In Natural Language Processing

In Optimization & Training

Citations in the corpus

Sequence-Level Knowledge Distillation (2016, background): “, 2011), and minimum risk training (Och, 2003; Shen et al., 2016).”
Findings of the 2016 Conference on Machine Translation (2016, methodology): “the development set, optimizing TER/BLEU with Minimum Error Rate Training (Och, 2003).”
Findings of the 2015 Workshop on Statistical Machine Translation (2015, positive): “Finally, the APE system was tuned on the development set, optimizing TER with Minimum Error Rate Training (Och, 2003).”
A Knowledge-Grounded Neural Conversation Model (2017, background): “In many respects, this second paradigm can be characterized as an extension of conventional dialog models with or without statistical modelling, e.g., (Oh and Rudnicky, 2000; Ratnaparkhi, 2002; Banchs and Li, 2012; Ameixa et al., 2014; Nio et al., 2014).”
A Diversity-Promoting Objective Function for Neural Conversation Models (2015, methodology): “We optimize γ and λ us-ing MERT (Och, 2003) on N-best lists of response candidates.”
An Actor-Critic Algorithm for Sequence Prediction (2016, background): “Another popular approach is to replace the domain over which the task score expectation is defined with a small subset of it, as is done in Minimum (Bayes) Risk Training (Goel and Byrne, 2000; Shen et al., 2015; Och, 2003).”
Survey of Hallucination in Natural Language Generation (2022, methodology): “To mitigate this problem, Wang and Sennrich [162] propose substituting MLE as a training objective with minimum risk training (MRT) [118].”
A Neural Attention Model for Abstractive Sentence Summarization (2015, methodology): “We follow the statistical machine translation setup and use minimum-error rate training (MERT) to tune for the summarization metric on tuning data (Och, 2003).”
A Neural Network Approach to Context-Sensitive Generation of Conversational Responses (2015, methodology): “The log-linear weights are estimated by running one iteration of MERT.”
Learning to Write with Cooperative Discriminators (2018, background): “statistical machine translation (Koehn et al., 2003; Och, 2003; Watanabe et al., 2007; Chiang et al., 2009) and remain common with neural machine translation, even when an extremely large amount of data is available (Wu et al.”
Neural Approaches to Conversational AI (2018, neutral): “luation, and this would presumably require model parameters to be publicly available. 63 for example in the frequent cases where automatic evaluation metrics are used directly as training objectives (Och, 2003; Ranzato et al., 2015) as unintended “gaming” may occur unbeknownst to the system develope”
From captions to visual concepts and back (2014, methodology): “Our final stage uses MERT [31] to re-rank the sentences.”
A Persona-Based Neural Conversation Model (2016, methodology): “We optimize γ and λ on N-best lists of response candidates generated from the development set using MERT (Och, 2003) by optimizing BLEU.”
Abstractive Sentence Summarization with Attentive Recurrent Neural Networks (2016, methodology): “Additionally, we compare to an enhanced version of their system (ABS+), which relies on a range of separate extractive summarization features that are added as log-linear features in a secondary learning step with minimum error rate training (Och, 2003).”
METEOR: An Automatic Metric for MT Evaluation with High Levels of Correlation with Human Judgments (2007, background): “Bleu is fast and easy to run, and it can be used as a target function in parameter optimization training procedures that are commonly used in state-of-the-art statistical MT systems (Och, 2003).”
deltaBLEU: A Discriminative Metric for Generation Tasks with Intrinsically Diverse Targets (2015, methodology): “We also included four variants of that system, which we tuned with MERT (Och, 2003).”
Modeling Coverage for Neural Machine Translation (2016, methodology): “Following the success of Minimum Risk Training (MRT) in SMT (Och, 2003), Shen et al.”
Neural Summarization by Extracting Sentences and Words (2016, methodology): “We incorporate the features in a log-linear reranker whose feature weights are optimized with minimum error rate training (Och, 2003).”
Minimum Risk Training for Neural Machine Translation (2015, methodology): “…a new training objective: where S ( x ( s ) ) ⊂ Y ( x ( s ) ) is a sampled subset of the full search space, and Q ( y | x ( s ) ; θ , α ) is a distribution deﬁned on the subspace S ( x ( s ) ) : Note that α is a hyper-parameter that controls the sharpness of the Q distribution (Och, 2003).”