The state of the art techniques in Machine Translation (MT) require large amounts of symmetric resources from source and target languages. This is true regardless of whether the approach is Transfer or Interlingua, Symbolic or Statistical. Symmetry within these approaches is necessary to ensure quality, robustness and retargetability. In reality, such symmetry, whether it is in terms of structural transfer lexicons, interlingual dictionaries or parallel corpora, is a major bottleneck in developing any MT system. This dissertation presents an approach to MT that addresses the lack of symmetry by exploiting symbolic and statistical target language resources in source-poor/target-rich language pairs. This approach is called Generation-Heavy Hybrid Machine Translation (GHMT). Expected source language resources include a syntactic parser and a simple one-to-many translation dictionary. No transfer rules or complex interlingual representations are used. Rich target language symbolic resources are used to overgenerate multiple structural variations from a target-glossed syntactic dependency representation of source language sentences. Statistical target-language resource are then used to select amongst the overgenerated translations. The source-target asymmetry of systems developed in this approach makes them more easily retargetable to new source languages.
Figures & tables
Figure 1.1: Generation-Heavy Machine Translation
Figure 2.1: The Nitrogen Generation System
(DEFINE-WORD :DEF_WORD "jog" :CLASS "51.3.2.a.ii" :THETA_ROLES "_th,src(),goal()" :WN_SENSE (01315785 01297547) :LANGUAGE ENGLISH :LCS (event go loc (* thing 2) ((* path from 3) loc (thing 2) (position at loc (thing 2) (thing 4))) ((* path to 5) loc (thing 2) (position at loc (thing 2) (thing 6))) (manner jog+ingly 26)) :VAR_SPEC ((3 :optional) (5 :optional))) (DEFINE-WORD :DEF_WORD "to" :LANGUAGE ENGLISH :LCS (path to loc (thing 2) (position in loc (thing 2) (* thing 6))))
Figure 2.2: Lexicon Entries for jog and to
#
Thematic Role
Definition
0
no thematic role assigned
1
AG
agent
2
TH ,EXP ,INFO
theme or experiencer or information
3
SRC()
source preposition
4
SRC
source
5
GOAL(), PRED()
goal or pred preposition
Figure 2.3: Inventory of Thematic Roles in LCS
26.3 Verbs of Preparing (event cause (* thing 1) (event go ident (* thing 2) (path toward ident (thing 2) (position at ident (thing 2) (property run+ed 9)))) ((* for 17) poss (head) (* thing 18))) Example: John ran the store for Mary. Other verbs: bake boil cook fix grill iron mix prepare roast roll wash …
47.5.1.b Swarm Verbs (Locational) (event act loc (* thing 2) ((* position [at] 10) loc (thing 2) (thing 11)) (manner run+ingly 26)) Example: The dogs run in the forest. Other verbs: bustle crawl creep swarm swim teem …
51.3.2.a.i Run Verbs - (Locational,Theme only) (event go loc (* thing 2) ((* path from 3) loc (thing 2) (position [at] loc (thing 2) (thing 4))) ((* path to 5) loc (thing 2) (position [at] loc (thing 2) (thing 6))) (manner run+ingly 26)) Example: The horse ran into the field from the barn. Other verbs: climb crawl fly jog jump leap race swim walk …
Figure 2.4: Different RLCS Entries for “run”
Figure 2.5: Different CLCS Decompositions into English, Spanish and Arabic
Divergence
Spanish
English
Occurrence
Categorial
X tener hambre (X have hunger )
X be hungry
98%
Conflational
X dar puñaladas a Z (X give stabs to Z)
X stab Z
83%
Structural
X entrar en Y (X enter in Y)
X enter Y
35%
Head Swapping
X cruzar Y nadando (X cross Y swimming )
X swim across Y
8%
Thematic
X gustar a Y (X please to Y)
Y like X
6%
Table 2.1: Translation Divergence Types
1
The entire content of the original text have been translated into an easily readable form.
2
The entire content of the original text have been faithfully reproduced and is easily understood in the translation, but minor post-editing is required.
3
The translation is generally faithful to the original, but some reorganization of the word-order is required.
4
The translation is generally faithful to the original, but there are errors in the relationships between phrases, tenses, singular/plural, position of adverbs, etc. post-editing requires structural transformations.
5
The structure and content of the original work are not well preserved in translation, with potions of the original text incorrectly translated or referents of words and phrases incorrect.
6
The structure and content of the original are poorly translated. Phrases and clauses are missing but valid sentence structure is achieved.
7
The translation reflects neither the structure nor the content of the original text. Valid sentence structure is not achieved because of missing subjects, predicates, etc.
Table 2.2: Nagao’s Fidelity Metric
1
The sentence has a clear meaning, with no possibility of misunderstanding. Appropriate grammar, terminology, and sentence structure are employed, and post-editing is not necessary.
2
The meaning of the sentence is clear, but there are slight problems in the sentence structure, grammar, or terminology. Post-editing would easily be accomplished by human hand.
3
The overall meaning can be grasped, but there are small ambiguities due to grammar or terminology. Post-editing would best be done with reference to the original work.
4
The quality of the translation is poor, and there are many problems of grammar and terminology. Some understanding of the sentence can be inferred after a lengthy study of the translation, but it would be better to undertake the translation anew by human translator than to post-edit the machine translation.
5
No understanding of the contents of the translated sentence is possible. Human translation is required.
Table 2.3: Nagao’s Ease of Understanding Metric
Figure 3.1: Generation-Heavy Machine Translation
Figure 3.2: Generation
Divergence Type
LCS-based MT
Divergence Operations
Categorial
SL and TL lexical entries with shared primitives
categorial variation
Conflational
SL and TL lexical entries shared substructures
conflation
Inflational
SL and TL lexical entries shared substructures
Inflation
Structural
SL and TL lexical entries with argument-position markings
case variation
Head-swapping
TL lexical entries marked for predicate reversal w.r.t SL predicates
inflation followed by conflation
Thematic
TL lexical entries marked for argument reversal w.r.t. SL arguments
case variation
Table 3.1: LCS-based MT vs Divergence Operations
Figure 4.1: Matador and Exerge
Feature
Definition
Values
Parts of Speech
:WORD
Source language word
<STRING>
:SURFACE
Target language phrase translation
<STRING>
:MOD-CC
Coordinated Modifier
<NODE>
:GEN
Gender
(F|M|N)
N , V , PN
:NUM
Number
(SG|PL)
N , V
:PER
Person
(1|2|3)
N , V
Table 4.1: Syntactic Dependency Specification
Figure 4.2: Output of Analysis
abandonar V abandon/desert/forsake/leave/quit abandonar V cede/give_in/give_up/give_way/relinquish/yield cesión N abandonment abandonado AJ forlorn/abandoned/abandonee abandonamiento N indulgence abandonar V forsake/abandon/desert/leave/quit/relinquish/shutter/surrender/give_way/succumb abandono N renunciation/dereliction/failing abdicación N abandonment/job/task aflojar V abandon_ship/abate/loose/loosen/relax/slack/slacken/unfasten/untie/sink deber V owe/should::AUX/must::AUX/have_to::AUX dejar V lets/assume/bequeath/desert/lay/abandon/allow/let/permit/leave desamparado AJ abandonee/helpless desamparar V forsake/abandon descontar V depreciate/reduce/abandon_ship/cash_up/derate descuidar V abandon/neglect expedir V abandon/abdicate/draw_up/give_up/send/dispatch/leave/register renunciar V abjure/give_up/abandon/renounce/waive/relinquish/surrender tense:impf FEAT tense:past tense:pret FEAT tense:past icoOTHERWISEic/icalente OTHERWISE ent ciónOTHERWISEtionamente OTHERWISE ly/ely/ally ía$ OTHERWISE y á OTHERWISE a é OTHERWISE e í OTHERWISE i ó OTHERWISE o ú OTHERWISE u ñ OTHERWISE gn/n
Figure 4.3: A Sample from the Matador Spanish-English Dictionary
Figure 4.4: Output of Translation
Figure 5.1: Exerge: Expansive Rich Generation for English
Thematic Role
Definition
AG
agent
INSTR
instrument
TH
theme
EXP
experiencer
INFO
information
PRED
predicate
Figure 5.2: Inventory of Thematic Roles in Exerge
Figure 5.4: Output of Normalization
RESTRUCTURE (parent) returns list-of-dependencies initialize restructuredParents = NIL #Check for restructuring conditions if (parent.POS equal V and parent.words.POS includes AUX or AV and exist (child in parent.children) such-that child.synRel equal :obj and child.POS equal V) #Restructure then foreach pos in parent.words.POS newParent = replicate parent newParent.words = subset of parent.words with pos newParent.POS = pos if pos equal V then add newParent to restructuredParents else swap newParent and newParent.child combine arguments of newParent and newParent.child add newParent.child to restructuredParents return restructuredParents
Figure 5.5: Pseudocode for Restructuring
Figure 5.6: Output of Restructuring
THEMATIC-LINKING (synRels, thetaGrid) returns linkSets #Create link netweork initialize linkNetwork = NIL foreach synRel in synRels foreach thetaRole in thetaGrid if (synRel-thetaRole in SynThematicLinkingMap) then add synRel-thetaRole to linkNetwork #Extract all possible link sets linkSets = enumerate link sets in linkNetwork foreach linkSet in linkSets linkSet.score = score(linkSet) if (linkSet contains unpaired obligatory thetaRoles) then remove linkSet from linkSets #Rank and Select return all linkSets with maximum score
Figure 5.7: Pseudocode for Thematic Linking
Figure 5.8: Syntactic-Thematic Linking Example
NN-CONFLATION-VARIANTS (parent) returns list-of-dependencies initialize conflatedParents = NIL foreach child in parent.children # Check for conflatability if (exist(child’ = VerbCatVar(child.word)) and child’.lcsPrim equal parent.word.lcsPrim and child’.thetaGrid equal (parent.word.thetaGrid - child.thetaRole)) # Apply conflation then newParent = replicate(parent) remove child from newParent.children newParent.word = child’ add newParent to conflatedParents return conflatedParents
Figure 5.9: Pseudocode for Node-Node Conflation Variant Creation
NR-CONFLATION-VARIANTS (parent) returns list-of-dependencies initialize conflatedParents = NIL foreach child in parent.children # Check for conflatability if (exist(confVerb = VerbCatVar(child.synRel)) and confVerb.lcsPrim equal parent.word.lcsPrim and parent.word.thetaGrid is-subset-of confVerb.thetaGrid and confVerb.thetaGrid.obligatory is-subset-of parent.word.thetaGrid # Apply conflation then newParent = replicate(parent) newParent.children.child.synRel = * newParent.word = confVerb add newParent to conflatedParents return conflatedParents
Figure 5.10: Pseudocode for Node-Relation Conflation Variant Creation
NN-INFLATION-VARIANTS (parent) returns list-of-dependencies initialize inflatedParents = NIL # Check for inflatability if (exist(parent’ = NounCatVar(parent.word) or AdjectiveCatVar(parent.word)) and exist(infVerb) such-that infVerb is-member-of LightVerbs and infVerb.lcsPrim equal parent.word.lcsPrim and exist(thetarole) such-that infVerb.thetaGrid equal (parent.word.thetaGrid + thetaRole)) # Apply inflation then newParent = replicate(parent) newParent.word = infVerb add parent’ to newParent.children in relation thetarole add newParent to inflatedParents return inflatedParents
Figure 5.11: Pseudocode for Node-Node Inflation Variant Creation
NR-INFLATION-VARIANTS (parent) returns list-of-dependencies initialize inflatedParents = NIL # Check for inflatability if (exist(newSynRel = PrepositionCatVar(parent.word)) and exist(infVerb) such-that infVerb is-member-of LightVerbs and infVerb.lcsPrim equal parent.word.lcsPrim and infVerb.thetaGrid equal parent.word.thetaGrid) # Apply inflation then foreach child in parent.children if child.thetaRole is-member-of newSynRel.thetaRoles then newParent = replicate(parent) newParent.word = infVerb newParent.children.child.synRel = newSynRel add newParent to inflatedParents return inflatedParents
Figure 5.12: Pseudocode for Node-Relation Inflation Variant Creation
DEPENDENCY-PRUNE (parent) returns pruned-parent SCORE-WORDS parent prunedParent = SELECT-WORDS parent return prunedParent SCORE-WORDS (parent) calculates scores and preferences foreach child in parent.children SCORE-WORDS child foreach parentWord in parent.words parentWord.score = 0 foreach child in parent.children select childWord from child.words such-that maximize (bigram(parentWord,childWord) + childWord.score) set prefer(parentWord,childWord) = true parentWord.score = parentWord.score + bigram(parentWord,childWord) + childWord.score SELECT-WORDS (parent) returns pruned-parent select parentWord from parent.words such-that maximize parent.word.score parent.words = parentWord foreach child in parent.children SELECT-CHILD-WORDS (parentWord,child) return parent SELECT-CHILD-WORDS (parentWord,child) returns pruned-child select childWord from child.words such-that prefer(parentWord,childWord) foreach grandChild in child.children SELECT-CHILD-WORDS (childWord,grandChild) return child
Machine translation develops computational methods for converting text, speech, and visual content between languages while preserving meaning and purpose.
High-quality machine translation (MT) can scale to hundreds of languages, setting a high bar for multilingual systems. However, compared to the world's 7,000 languages, current systems still offer only limited coverage: about 200 languages on the target side, and maybe a few hundreds more on the source side, supported due to cross-lingual transfer. And even these numbers have been hard to evaluate due to the lack of reliable benchmarks and metrics. We present Omnilingual Machine Translation (OMT), the first MT system supporting more than 1,600 languages. This scale is enabled by a comprehensive data strategy that integrates large public multilingual corpora with newly created datasets, including manually curated MeDLEY bitext. We explore two ways of specializing a Large Language model (LLM) for machine translation: as a decoder-only model (OMT-LLaMA) or as a module in an encoder-decoder architecture (OMT-NLLB). Notably, all our 1B to 8B parameter models match or exceed the MT performance of a 70B LLM baseline, revealing a clear specialization advantage and enabling strong translation quality in low-compute settings. Moreover, our evaluation of English-to-1,600 translations further shows that while baseline models can interpret undersupported languages, they frequently fail to generate them with meaningful fidelity; OMT-LLaMA models substantially expand the set of languages for which coherent generation is feasible. Additionally, OMT models improve in cross-lingual transfer, being close to solving the "understanding" part of the puzzle in MT for the 1,600 evaluated. Our leaderboard and main human-created evaluation datasets (BOUQuET and Met-BOUQuET) are dynamically evolving towards Omnilinguality and freely available.
Most endangered languages lack the parallel data required for machine translation, despite the existence of descriptive grammar books. We introduce a pipeline that uses large language models to extract grammatical rules, example sentences, and lexicons from grammar books and generate synthetic parallel corpora for fine-tuning-rather than feeding grammar content into prompts at inference time, as in prior work. Validated on three typologically diverse low-resource languages-Kalamang (Papuan), Tuatschin (Romance), and Mandan (Siouan)-we show that fine-tuning on synthetic data improves over seed-data baselines in 75% of configurations for Kalamang and 59% for Tuatschin, with best-case ChrF++ gains of +8.8, +5.3, and +3.3 respectively. Through a systematic factorial study across 96 configurations varying target part-of-speech, retrieval granularity, and sample volume, we identify which factor combinations drive gains and where they break down. Our results demonstrate that static linguistic documentation can be repurposed for machine translation fine-tuning, offering a practical path towards translation tools for severely under-resourced languages.
Varun Ghat Ravikumar, Sina Ahmadi, Lena Jäger +1
Department of Informatics, University of Zurich · Department of Computational Linguistics, University of Zurich
Large-scale machine-translation (MT) systems are typically evaluated on random samples from a corpus whose distributional composition is an artifact of how it was assembled. Such a sample inherits the phenomena the collection happens to contain rather than the full space a system must handle, spanning rule-governed conventions (terminology, punctuation, currency formatting) and context-dependent phenomena (tone, honorifics, document-level coherence), and thus provides no coverage guarantee for assessing robustness. We propose TACTICS (Taxonomy-Aware Coverage-opTimized Intelligent Corpus Sampling), which recasts coverage as an explicit objective. TACTICS induces a hierarchical taxonomy from a locale style guide, classifies segments against it, and selects a fixed-budget subset jointly optimizing coverage of rare categories, document-level coherence, and distributional fidelity to the full corpus. Applied to MT evaluation across four translation directions, TACTICS improves coverage of rare categories over lexical and embedding-based selection. By targeting the phenomena that separate systems, TACTICS makes a fixed evaluation budget go further, recovering the true system ranking from far fewer segments than random sampling wherever a real quality gap exists and never signaling a difference where none exists.