MorphologyBased TenseAspect Disambiguation
for Telugu English Translation

Telugu, a Dravidian language spoken by over 80million people, encodes a rich set of tenseaspect (TA) distinctions through agglutinative morphology. When translating Telugu sentences into English, the surface form alone often does not make the intended TA clear. Accurate translation therefore requires a systematic disambiguation step that interprets morphological markers before generating the English clause. This page surveys the linguistic background, the computational challenges, and the current morphologydriven strategies that enable highquality TeluguEnglish translation.

1. Morphological Overview of Telugu

Telugu verbs are built from a lexical root followed by a sequence of suffixes that convey:

• PersonNumberGender (PNG) agreement with the subject.
• Tense: present, past, future.
• Aspect: habitual, progressive, perfective, perfectprogressive.
• Mood: indicative, imperative, desiderative, conditional.
• Voice: active, passive, causative.

Because these suffixes appear in a fixed order, a single inflected form may carry multiple TA cues. For example:

Here - (taunnardu) is a progressive aspect marker fused with presenttense agreement. The same root (caduvu read) can appear with very different TA values, making morphological analysis a prerequisite for any rulebased or statistical translation system.

2. Tense and Aspect in Telugu

While English uses auxiliary verbs and adverbials to express TA, Telugu relies heavily on suffixes. The most common markers are:

TA Category	Telugu Marker	Typical English Rendering
Simple Present	- / -	does / does (polite)
Simple Past	- / -	did
Future	- / - (with )	will
Progressive	- / -	is ing
Habitual	- (contextual)	usually / habitually
Perfective	- / -	has ed
PerfectProgressive	- after perfective	has been ing

Because the same suffix can belong to more than one TA class, disambiguation often depends on surrounding context or on syntactic cues such as auxiliary verbs in the source sentence.

3. Why Disambiguation Is Needed for Translation

• Lexical ambiguity: could mean it fell (past) or it is falling (present progressive) without extra clues.
• Aspecttense interaction: The progressive marker - can combine with future intent when accompanied by a particle , yielding will be ing.
• Crosslingual mismatch: English sometimes splits a single Telugu suffix into a verb + auxiliary (was eating) while Telugu may encode the same notion in one word.

A failure to resolve these ambiguities leads to mistranslations such as He ate instead of He was eating, which can change the temporal flow of a narrative dramatically.

4. MorphologyBased Disambiguation Approaches

4.1 RuleDriven Morphological Parsing

Traditional systems first segment the verb into root + suffix chain using a finitestate transducer (FST). The FST encodes a deterministic mapping from suffix sequences to TA categories. For example:

Rules then resolve conflicts. If both perfective - and progressive - appear, a preference hierarchy (Perfective>Progressive) is applied, unless a discourse cue overrides it.

4.2 Statistical Morphological Tagging

Conditional Random Fields (CRF) and BiLSTM models trained on manually annotated corpora can predict TA tags directly from raw characters. These models capture morphophonemic alternations (e.g., vowel harmony) that are difficult to enumerate in rule tables.

4.3 Hybrid Systems

Modern pipelines combine an FST for reliable segmentation with a neural tagger for ambiguity resolution. The FST guarantees that every token is split into meaningful morphemes; the tagger then decides which TA interpretation best fits the surrounding word sequence.

5. Integrating Disambiguation with Machine Translation

Once a TA label is assigned, the translation component can be steered in two ways:

1. Preprocessing: Insert explicit English auxiliaries into a pseudosource sentence (e.g., convert to readPROG3SGMASC), then feed the enriched sequence to a standard neural MT model.
2. Postediting: Let a baseline TeluguEnglish model produce an initial translation, then replace the verb phrase with a TAconsistent rendering derived from the morphological tagger.

Both strategies have been shown to reduce BLEU score errors related to tense and aspect by 812% on test sets of parallel TeluguEnglish sentences.

6. Example Analyses

6.1 Simple Past vs. Progressive

6.2 Future Progressive with Modal Particle

7. Evaluation Metrics

Beyond conventional MT scores, TAspecific assessment is crucial. Common metrics include:

• TAAccuracy: Percentage of verbs whose tense and aspect match the reference.
• Temporal Cohesion Score: Measures how well translated events maintain chronological order.
• Human Rating: Native speakers rate the naturalness of temporal expressions on a 5point scale.

In a recent shared task, the best hybrid system achieved 92% TAAccuracy while maintaining a BLEU of 41.2.

8. Future Directions

Research is moving toward tighter integration of morphological analysis with large language models (LLMs). Promising avenues include:

• Prompttuned LLMs that condition on explicit TA tags to generate more faithful English verbs.
• Joint learning where the same neural architecture simultaneously learns segmentation, tagging, and translation, sharing representations across tasks.
• Crosslingual transfer using highresource languages with similar agglutinative patterns (e.g., Tamil, Kannada) to bootstrap annotators for Telugu.

9. Conclusion

Telugus morphology offers both a challenge and an opportunity for precise tenseaspect disambiguation. By decomposing verbs into root and suffixes, applying rulebased or statistical TA tagging, and feeding the resulting information into a translation engine, developers can produce English renditions that respect the temporal nuances intended by the original speaker. Continued advances in hybrid parsing, neural tagging, and LLM integration promise even higher fidelity for TeluguEnglish MT, ultimately supporting better communication across the two language communities.

For further reading, see the works of Kumar etal. (2022), Rao & Bhat (2021), and the EMNLP 2023 Shared Task on TAaware Translation.