Springer出版：机器翻译混合方法深度解析

需积分: 9 157 浏览量更新于2024-07-19 收藏 5.94MB PDF 举报

"Hybrid Approaches to Machine Translation" 是一本由 Springer 出版的208页专业书籍，涵盖了自然语言处理领域内的深度探讨。本书旨在理论与实践相结合，提供机器翻译的最新进展。在过去的十年里，随着大数据的增长、新应用领域的涌现以及跨领域间的紧密联系，自然语言处理(NLP)的发展如日中天，研究内容和方法呈现出多元化趋势。作为 "Theory and Applications of Natural Language Processing" 系列的一部分，这本书特别关注于NLP和语言技术中的特定话题。系列主编包括 Julia Hirschberg、Eduard Hovy 和 Mark Johnson，他们共同确保每一卷都聚焦于最前沿的研究成果。书中深入剖析了计算模型和多语言文本处理的各种方法，包括但不限于语音和文本的处理技术，以适应日益增长的应用场景。在机器翻译方面，"Hybrid Approaches" 提供了混合方法的详细研究，这些方法结合了统计机器翻译（SMT）、神经机器翻译（NMT）以及深度学习等技术。通过这种方法，作者们试图克服单一方法可能遇到的局限性，比如SMT在处理复杂语义和长距离依赖上的不足，而NMT则在端到端学习上表现卓越。本书还可能讨论如何利用语言模型、词嵌入、注意力机制和其他先进的技术来提升翻译质量，同时兼顾效率和准确性。对于学生和潜在用户来说，这本书既是一份宝贵的理论参考，也是进入这个快速发展的领域的重要指南。它为读者提供了系统化的框架，帮助他们理解并应用不同方法之间的相互补充和协同作用，从而推动机器翻译技术在实际应用中的创新与发展。通过详尽的目录和书签，读者可以轻松地找到他们感兴趣的主题，无论是基础理论还是实用技巧，都能在本书中找到相应的阐述和案例分析。 "Hybrid Approaches to Machine Translation" 是一本深入研究当前NLP技术融合策略的宝典，为研究人员、工程师和开发者提供了丰富的工具和思路，以应对日益复杂的语言处理挑战。无论是在学术界还是工业界，这本书都具有很高的参考价值。

6 C. España-Bonet and M.R. Costa-jussà

Grammar rules

Lexicons

RBMT

Constituents

Corpus

SMT

Statistical

Phrases

Extraction

Generation

Enrichment

Fig. 4 Shallow loans and exchanges between the resources of RBMT and SMT paradigms (above)

and their side products (below). Some hybrid systems are build by taking advantage of these

relations

very useful (Fig. 4). Statistical systems can also be used as a pre-processing or post-

processing step in order to help in disambiguation and choose among the different

translation options, but, as we will see in the following section, it is more common

the other way around and rules are sometimes used as a pre-process in SMT systems

to reorder the input sentence or as a post-process to choose grammatically correct

sentences.

2.2 Statistical Machine Translation

SMT systems estimate probabilistic models (language and translation models) by

frequency counts in corpora. For a language model, monolingual data in the target

language is needed; translation probabilities are extracted from relations in both

the target and the source language and therefore parallel data are used to build

the translation model. At the end, the best translation in a statistical system is that

maximising the product of models.

Differences in SMT systems mainly arise from the deﬁnition of the minimal unit

for translation (word vs. phrase vs. constituent). Dictionaries and/or constituents

obtained with RBMT systems can help in the translation as a complement to the

parallel corpus, and rules can be used for reordering the input or the output as a

pre-process or post-process respectively (Fig. 4).

2.2.1 Phrase-Based Systems

Phrase-based translation (PBSMT) (Koehn et al. 2003) is the natural evolution of

word-to-word translation. When considering a phrase instead of a word, the (small)

context and the local reordering of every word is taken into account. A phrase here

is just a sequence of words but it is not necessarily a linguistic element consistent

Hybrid Machine Translation Overview 7

with a word alignment between the source and the target (Koehn 2010: 131). This is

positive because the number of phrases that can be extracted from a corpus is larger

than the number of syntactic elements and it allows more freedom in the translation.

But on the other hand, the extraction of noisy phrases leads to non-grammatical

translations.

2.2.2 Syntax-Based Systems

Syntax-based systems (SSMT) (Wu 1997; Quirk et al. 2005) implement the opposite

approach and translate syntactic elements. The basic idea is to use synchronous

grammars which are able to generate the source and the target simultaneously.

Synchronous grammars are learned from parallel corpus and that makes the

approach very slow in comparison to the PBSMT systems.

2.2.3 Hierarchical Phrase-Based Systems

Hierarchical phrase-based systems (HPBSMT) (Chiang 2005) strike a balance

between pure lexical PBSMT and SSMT. A hierarchical phrase consists of words

and subphrases and this hierarchy is intended to capture reorderings among phrases.

The hierarchical phrase pairs are obtained from a synchronous context-free grammar

(CFG) learned from parallel corpora without syntactic information. As the authors

explain, the system is formally syntax-based as it uses synchronous CFG, but not

linguistically syntax-based, because the parallel data does not include syntactic

information.

3 System Combination

A popular research direction for bringing together different paradigms in MT is

system combination, which consists in selecting the highest quality output among

various options from multiple MT engines. Therefore, this type of combination does

not aim at integrating system architectures, but only the translation outputs from

systems as the ones described in the previous section.

System combination in the ﬁeld of MT has originally been inspired by previous

successful works in the area of speech recognition (Fiscus 1997). Basically, MT

outputs may be combined at three different levels: sentence, subsentencial, or search

graphs. The combination for the sentence and subsentencial levels takes either the

best output sentence of each available system or a list of n-best outputs. In the

following, we make an overview of each category and collect references from all

categories in Table 1.

8 C. España-Bonet and M.R. Costa-jussà

Table 1 System

combination summary

Granularity References

Sentence-level Callison-Burch and Flournoy

(2001), Nomoto (2004), Akiba

et al. (2002), Costa-jussà et al.

(2007), Formiga et al. (2013)

Subsentencial-level Frederking and Nirenburg (1994),

Bangalore et al. (2001), Jayaraman

and Lavie (2005), Matusov et al.

(2006), Sim et al. (2007), Rosti

et al. (2008), He et al. (2008),

Mellebeek and van Genabith

(2006)

Graph-level Li et al. (2009), DeNero et al.

(2010), Duan et al. (2011), Okita

and van Genabith (2012)

3.1 Sentence-Level Combination

This is the level with the lowest granularity. At this level, system combination has

been widely inﬂuenced by the use of the information provided by n-gram language

modelling (Callison-Burch and Flournoy 2001; Nomoto 2004), which assume that

the best translation comes from the most ﬂuent translation. Akiba et al. (2002)

take a step forward and propose two different methods that use a combination

of language and translation models to choose the best translation. The authors in

Costa-jussà et al. (2007) combine the outputs of an n-gram-based and a PBSMT

system by using the scoring of both systems. These scores include language and

translation models as in the previous works, but also the other features of standard

SMT systems. Recent approaches such as that in Formiga et al. (2013)useSVMs

and random forests together with several conﬁdence measures to re-rank an n-best

list of translation outputs.

3.2 Subsentencial-Level Combination

One step further in obtaining the best translation output is done by merging

hypothesis at the level of words or phrases. One of the ﬁrst works in this direction

can be found in Frederking and Nirenburg (1994), which combines three MT

engines from different nature into a chart data structure. This work is even previous

to the ﬁrst references on sentence-level combination systems, and to obtain the

best translations they use several heuristics to estimate the quality of each chunk.

Some years later, Bangalore et al. (2001) used a monotonic alignment to construct

a confusion network from several translation hypotheses. Alignments are crucial in

the combination and, soon after, Jayaraman and Lavie (2005) and Matusov et al.

(2006) propose a non-monotonic alignment. In the latter, the context of the whole

Hybrid Machine Translation Overview 9

document is taken into account. Then, Sim et al. (2007) extracts alignments with the

Translation Error Rate (TER) measure (Snover et al. 2006). Rosti et al. (2008)uses

conﬁdence scores derived from generalised linear models and He et al. (2008)uses

indirect HMM alignments.

Within another line of subsentencial combination, Mellebeek and van Genabith

(2006) do not rely on word alignments of output hypotheses, but prepare the

input sentence for multi-engine processing. They do this by using a recursive

decomposition algorithm that produces simple chunks as input to the MT engines. A

consensus translation is produced by combining the best chunk translations, selected

through majority voting, a trigram language model score and a conﬁdence score

assigned to each MT engine.

3.3 Search Graph-Level Combination

At the top level of granularity, there is the proposal of integrating consensus

translations at decoding time. This is the most recent approach within system

combination and also the one that is closer to HMT. In Li et al. (2009) both

partial and full hypotheses are re-ranked during the decoding phase directly using

consensus between translations from different systems. According to the authors,

this is different from system combination since decoders collaborate by exchanging

partial translation results and not ﬁnal n-best lists, but this is the case for several

systems presented in this section. Also the authors in DeNero et al. (2010)deﬁne

their system as a forest-based technique that uniﬁes consensus decoding and

system combination. The approach is able to deal with systems with heterogeneous

structures, the only request is that every system to be combined can output a

forest or lattice of translations. In both cases, collaborative decoding (Li et al.

2009) and model combination (DeNero et al. 2010), the systems outperform their

corresponding baseline in terms of the BLEU score (Papineni 2002).

Work by Duan et al. (2011) presents hypothesis mixture decoding, a new decod-

ing scheme that performs translation reconstruction using hypotheses generated by

multiple component systems. This decoding involves two decoding stages: ﬁrst,

each component system decodes the source sentence independently; second, a new

search space is constructed by composing existing hypotheses produced by all

systems using a set of rules, and a new set of features is used to seek the ﬁnal

best translation within this new constructed search space. Finally, we mention a

system combination strategy that uses Minimum Bayes Risk (MBR) decoding.

Under certain hypotheses, the authors in Okita and van Genabith (2012)develop

a new MBR decoding strategy that exploits a larger hypothesis space compared

with previous works. The ﬁnal system improves from 1 to 10% on the BLEU score

depending on the language pair.

10 C. España-Bonet and M.R. Costa-jussà

4 Hybridisation Lead by an SMT System

A second way of combining more than one translation engine is by a hybridisation

between approaches. This hybridisation is useful to compensate the weak points of

an approach with the strong points of another one. Statistical systems can be built

whenever there are parallel corpora available and the quality of the translation is

affected by the amount of data. Since lexical selection is modelled from these data,

a good lexical choice and ﬂuency are the strong points of SMT systems when data on

the speciﬁc domain is available. On the bad side, syntax and long distance structures

are more difﬁcult to catch even with lots of data [see Costa-jussà et al. (2012)fora

review].

This section shows how statistical systems try to improve their global syntax

and grammaticality. To do so, SMT has integrated features from RBMT systems

at different stages including at a superﬁcial level such as using rules at pre/post-

processing or at a deep level such as augmenting the phrase table with RBMT

entries. See Table 2 for a summary of this classiﬁcation.

4.1 Pre/Post-processing Integration

There are popular approaches that use pre-processing rules to reorder source

sentences to better match target sentences. These approaches include human written

rules for different languages (Xia and McCord 2004; Collins et al. 2005;Wangetal.

2007; Patel 2013; Hatakoshi et al. 2014) or automatically extracted rules with non-

linguistic knowledge (Costa-jussà and Fonollosa 2006) or with syntactic knowledge

(Khalilov and Fonollosa 2011). In the former studies, speciﬁc rules are designed for

English–French (Xia and McCord 2004), German–English (Collins et al. 2005),

Chinese–English (Wang et al. 2007), English–Hindi (Patel 2013) and Japanese–

English (Hatakoshi et al. 2014). In the latter cases, when extracting automatic rules,

one cannot really say that the system is integrating RBMT knowledge.

A different type of pre-processing rules are those considering normalisation

issues in order to deal, for example, with data noise, chat or informal languages

Table 2 Hybrid systems lead by SMT: summary

Integration level References

Pre/post processing Xia and McCord (2004), Collins et al. (2005), Wang et al. (2007), Patel

(2013), Hatakoshi et al. (2014), Costa-jussà and Fonollosa (2006),

Khalilov and Fonollosa (2011), Formiga et al. (2012), Lewis et al.

(2013), Rudolf (2014), Farrús et al. (2011)

Core system Nießen and Ney (2004), Vogel and Monson (2004), Okuma (2008),

Eisele et al. (2008), Sánchez-Cartagena et al. (2016), Chen and Eisele

(2010), Ahsan et al. (2010), Li et al. (2011)

剩余207页未读，继续阅读

ignite678@126.com

粉丝: 2
资源: 42

Springer出版：机器翻译混合方法深度解析

approaches to translation studies

Uncertainties signal a need for new approaches to mobility.pdf

Learning to Learn Exploring Approaches to Help Machine and People Learn

Equilibrium Approaches to Modern Deep Learning.pdf

Machine Learning and Statistical Modeling Approaches to Image Retrieval

Statistical and Machine Learning Approaches for Network Analysis.pdf

Digital Approaches to Text Reuse in the Early Chinese Corpus.pdf

Relation Extraction using Deep Learning approaches Knowledge Graph.pdf

Multidisciplinary Approaches to Neural Computing

Practical approaches to SOA Governance

最新资源