TurnTurk-GPT
A text-based turn prediction model fine-tuned from a Turkish GPT-2, using the TurnGPT training objective to predict turn-ending points in dialogue, with particular attention to word-final morphological cues in morphologically rich languages.
Timeline (last 90 days)
TurnTurk-GPT is a text-based Turkish turn predictor, trained by fine-tuning a Turkish GPT-2 model with the TurnGPT objective.
TurnTurk-GPT outperforms widely used multilingual turn-end detectors in both record-disjoint and speaker-disjoint evaluation.
TurnTurk-GPT outperforms the zero-shot Turkish language model only when the test speaker appears in the training set; for unseen speakers, the two are statistically indistinguishable.
Removing the suffix of the turn-final word reduces average precision (AP) by up to 44%, whereas removing the suffix of intra-sentential words reduces it by only 2%.
The TurnTurk-GPT model trained only on synthetic dialogue performs weakest, indicating that real dialogue data is indispensable for capturing natural turn-transition patterns.
When the speaker pauses, the zero-shot baseline model and the multilingual baseline model are close to chance level, whereas TurnTurk-GPT still maintains some performance.
The evaluation design of TurnTurk-GPT uses unpunctuated, lowercased text, makes judgments at word-level decision points, and adopts record-disjoint and speaker-disjoint data splits, nested threshold selection, and a paired cluster bootstrap test.
All Facts (7)
TurnTurk-GPT is a text-based Turkish turn predictor, trained by fine-tuning a Turkish GPT-2 model with the TurnGPT objective.
50%UnverifiedTurnTurk-GPT outperforms widely used multilingual turn-end detectors in both record-disjoint and speaker-disjoint evaluation.
50%UnverifiedTurnTurk-GPT outperforms the zero-shot Turkish language model only when the test speaker appears in the training set; for unseen speakers, the two are statistically indistinguishable.
50%UnverifiedRemoving the suffix of the turn-final word reduces average precision (AP) by up to 44%, whereas removing the suffix of intra-sentential words reduces it by only 2%.
50%UnverifiedThe TurnTurk-GPT model trained only on synthetic dialogue performs weakest, indicating that real dialogue data is indispensable for capturing natural turn-transition patterns.
50%UnverifiedWhen the speaker pauses, the zero-shot baseline model and the multilingual baseline model are close to chance level, whereas TurnTurk-GPT still maintains some performance.
50%UnverifiedThe evaluation design of TurnTurk-GPT uses unpunctuated, lowercased text, makes judgments at word-level decision points, and adopts record-disjoint and speaker-disjoint data splits, nested threshold selection, and a paired cluster bootstrap test.
50%