TL;DR
This paper introduces SERUP, a novel parser designed to handle self-repairs in Japanese spoken utterances. By translating self-repaired sentences into well-formed versions using linguistic clues, SERUP effectively processes a significant portion of spontaneous speech, demonstrating its utility for natural language understanding systems.
What Problem Does This Paper Address?
Spontaneous human speech is inherently messy, often featuring self-repairs where speakers correct themselves mid-utterance. These "permissible ill-formednesses" are common, with studies showing they occur every few seconds in natural dialogues. For natural language processing (NLP) systems, especially dialogue systems, accurately understanding these self-repaired utterances is a significant challenge. Existing parsers often struggle with this, as they are typically designed for grammatically perfect input. The authors aim to bridge this gap by developing a system that can robustly parse Japanese utterances containing self-repairs, preventing the need for the listener (or system) to interrupt the speaker.
Methods
The core of the proposed system is SERUP (SElf-Repaired Utterance Parser). SERUP operates in two main stages:
- SR-reconstructor: When a standard parser fails, the utterance is passed to the SR-reconstructor. This component identifies self-repairs and translates the utterance into a well-formed version. The translation process involves:
- Removing common editing expressions (e.g., "er," "no," "I mean").
-
Replacing the "reparandum" (the part being corrected) with the "repair part" (the corrected segment). The SR-reconstructor relies on several linguistic clues to detect and resolve self-repairs, including:
-
Repetition: Words or phrases that are repeated around the point of interruption.
- Syntactic Break: Detecting breaks within words or phrases.
- Same Fragment Repetition: When a repair starts with a fragment of the interrupted word.
- Unknown Word: Identifying fragments that don't form valid Japanese words.
- Isolated Word: A fragment that, while potentially a word, is isolated and doesn't fit syntactically.
- Without Repetition of a Stem: Handling cases where repairs occur due to inflectional errors without repeating the base word.
- Fresh Start: A fallback mechanism for complete rephrasing where the preceding utterance fragment is largely ignored.
- Normal Parser: The translated, well-formed utterance is then passed back to a standard parser. If the SR-reconstructor can successfully convert the utterance into a parsable form, the system can proceed with subsequent processing steps without modification. This iterative process allows SERUP to handle multiple self-repairs within a single utterance.
Datasets
The paper uses the ADD corpus, a large collection of one million words of telephone dialogues created at ATR Interpreting Telephony Laboratories. This corpus contains spontaneous dialogues, making it suitable for evaluating a system designed to handle natural speech. The authors analyzed 1,082 self-repairs within this corpus.
Key Insights and Ideas
- Linguistic Clues are Powerful: The paper highlights that even without prosodic information (which is often difficult to obtain or process), linguistic cues like repetitions and word fragments are highly effective for detecting and resolving self-repairs in Japanese.
- Translation as a Preprocessing Step: The strategy of translating self-repaired utterances into well-formed ones before parsing is a key insight. This allows existing NLP pipelines to be used without modification, simplifying system integration.
- Japanese Specifics: The authors acknowledge that Japanese has unique syntactic structures (e.g., left vs. right branching) that might make approaches developed for English less directly applicable, underscoring the importance of language-specific analysis.
- "Changed to Well-Formed" is a Challenge: A significant category of utterances that SERUP struggles with are those where a self-repair results in a grammatically correct but semantically unintended utterance. This suggests that semantic and prosodic information might be crucial for resolving these specific types of errors.
Key Results
SERUP was evaluated on 1,082 self-repairs from the ADD corpus. The system successfully coped with 953 (88.1%) of these self-repairs. The authors estimate that if they could incorporate semantic constraints or prosodic cues, their system could potentially handle 1,064 (98.3%) of the self-repairs.
Limitations and Future Work
Limitations:
- SERUP cannot cope with all categories of self-repairs, particularly those that result in grammatically correct but semantically unintended utterances ("Changed to well-formed").
- The system currently relies on a "well-formed first" method, which limits its ability to leverage prosodic cues that might help resolve certain ambiguities.
- Some categories of self-repairs, such as ambiguous repairs or repairs with different word categories, are difficult to resolve automatically.
Future Work:
- Testing the system with a larger grammar and lexicon.
- Incorporating prosodic processing to improve the handling of challenging self-repair categories.
- Further investigation into semantic constraints to resolve "Changed to well-formed" utterances.
Summary updated