📄 A TAG-based noisy-channel model of speech repairs

Summary

The paper tackles speech repairs in spontaneous speech transcripts. In a repair, a speaker abandons a stretch of words (the reparandum), optionally inserts a filler or editing phrase (the interregnum), and then produces a corrected version (the repair). Johnson and Charniak treat repair removal as noisy-channel decoding. A language model generates a clean, repair-free sentence, and a channel model adds repairs to it. The channel model is a stochastic transducer built on Tree Adjoining Grammar (TAG). TAG was chosen because the repair usually echoes the reparandum word by word, which creates crossed dependencies that context-free or finite-state models cannot express.

Search uses a bigram language model to produce 20 candidate analyses. These candidates are then rescored with the syntactic parser-based language model of Charniak (2001). Training and testing use the disfluency-annotated Switchboard section of Penn Treebank III, with partial words and punctuation removed. With parser rescoring, word-level reparandum detection reaches an F-score of 0.797. This beats the bigram and trigram variants and the earlier Charniak and Johnson (2001) classifier, showing that syntax-aware language models help disfluency detection.

Key points

• Data:

  • Training used the disfluency-tagged Switchboard files (dysfl/dps/swbd) on the LDC Penn Treebank III CD-ROM.
  • The split follows Charniak and Johnson (2001): sw[23].dps for main training, sw4[5-9].dps as held-out data, and sw4[0-1]*.mrg from the parsed version for testing.

• Copying evidence: In the training data, more than 60% of the words in a reparandum reappear identically in the corresponding repair. The channel model is designed to assign high probability to such copies.

• Search pipeline:

  • TAG dynamic programming with a bigram language model finds the 20 best analyses.
  • The parser-based language model then reranks them.
  • Because parsing costs n^5, analysis is restricted to substrings of at most 12 words.

• Results (precision / recall / F-score):

  Model	Precision	Recall	F-score
  CJ01 classifier	0.951	0.631	0.759
  Bigram	0.776	0.736	0.756
  Trigram	0.774	0.763	0.768
  Parser	0.820	0.778	0.797

• Negative result: Replacing word bigram dependencies with POS bi-tag dependencies gave nearly identical results.

Interesting ideas

• I1. Cheap search, rich rescoring: A finite-state language model keeps the combined grammar tractable for dynamic programming, and a stronger global model is applied only to a short n-best list. The same split works whenever a good scorer cannot be intersected with the search space.
• I2. Transducer as a grammar over symbol pairs: Each terminal is an observed:source pair, with an empty symbol marking deletions. Any grammar formalism then becomes a transducer whose two sides are read off by projection, a simple trick for alignment-style tasks.
• I3. Mildly context-sensitive formalisms for "rough copy" phenomena: Using TAG to model a second string that echoes the first with crossed links is a principled option for other near-duplicate structures. The authors draw a parallel to genome modelling.
• I4. Local repair odds instead of one global parse: Each bounded substring gets a score: the probability of generating it as a repair versus as non-repair material. The scored substrings are merged afterwards. This is faster, and it can represent a repair that is itself repaired again, which a single global analysis cannot.
• I5. Realistic evaluation and complementary systems: - Removing punctuation and partial words imitates speech-recognizer output, a useful habit for any speech-text pipeline. - The authors observe that windowed classifiers handle short repairs well while the generative model handles long, similar ones. This suggests combining the two approaches.