TL;DR
This paper introduces a system that automatically generates shallow semantic frame representations from spontaneous, unrestricted speech. By combining partial parsing (chunking) with verb argument mapping from WordNet, the system aims to overcome the challenges of noisy, ungrammatical speech without requiring extensive manual knowledge engineering.
What Problem Does This Paper Address?
Analyzing spontaneous speech, especially from unrestricted domains, presents significant challenges for natural language processing (NLP). Unlike formal written text, spontaneous speech is often filled with dysfluencies (hesitations, repetitions), ungrammaticalities, and imperfectly recognized words, making traditional deep linguistic analysis difficult. Furthermore, building comprehensive semantic knowledge bases for unrestricted domains is a labor-intensive and often impractical task. This paper addresses the need for a robust and efficient method to extract meaningful semantic information from such challenging speech data, enabling its use in downstream NLP applications like summarization or information retrieval.
Methods
The core of the system relies on two main representations:
- Chunks: These are basic, non-recursive phrasal constituents, similar to standard linguistic notions of phrases but without explicit attachments (e.g., a prepositional phrase attached to a noun phrase). A verbal chunk specifically includes the verb complex but not its arguments.
- Frames: These are built from the parsed chunks by mapping verb arguments to subcategorization frames. These frames are automatically extracted from WordNet.
The process involves:
- Speech Recognition: Generating hypotheses from spoken utterances.
- Preprocessing: Cleaning up speech dysfluencies and segmenting turns into "short clauses" (minimal units with at least a subject and an inflected verb).
- Chunk Parsing: Using a POS-based grammar to identify phrasal constituents (chunks).
- Semantic Mapping: A "mapper" component attempts to find the best match between the identified chunks and the subcategorization frames associated with the main verb of the clause, leveraging WordNet for frame information.
Datasets
- SWITCHBOARD (SWBD) corpus: Used for speech data, transcripts, and annotations. This is a standard corpus for conversational speech research.
- WordNet 1.5: Used as a source for extracting subcategorization frames for verbs.
No new benchmarks were introduced; the evaluation was performed on a subset of the SWBD corpus.
Key Insights and Ideas
- Shallow Semantics is Achievable: The paper demonstrates that it's possible to generate useful semantic representations from noisy speech with minimal linguistic knowledge engineering, challenging the notion that deep analysis is always required.
- Robustness to Imperfection: The system is designed to handle dysfluencies and ungrammaticalities, which are common in spontaneous speech, by focusing on partial parsing and robust mapping.
- Leveraging Lexical Resources: WordNet serves as a crucial, readily available resource for extracting verb subcategorization frames, reducing the need for manual knowledge base construction.
- Chunking as a Foundation: The use of chunking provides a flexible intermediate representation that can handle variations in sentence structure common in spontaneous speech.
- Filtering Non-Informative Content: The observation that many frequent chunk patterns are non-informative (e.g., backchannels, hesitations) suggests that a preprocessing step can effectively filter these out before semantic mapping.
Key Results
- Initial Success Rate: In a test set, correct semantic mappings were identified in 21% of the clauses.
- Improved Rate on Cleaned Data: When clauses that were non-mappable (e.g., containing only noises/hesitations) or ungrammatical were removed, the success rate increased to 44.7% for the remaining clauses.
- Frequent Chunk Patterns: The paper identified that a small number of chunk patterns account for a large proportion of short clauses, highlighting potential areas for optimization.
Limitations and Future Work
- Mapper Improvement: The authors acknowledge that the mapper needs significant improvement to handle issues like constituent coordination, split verbs, and infinitive complements.
- Linking Clauses: The segmentation into short clauses can lead to a loss of information regarding the linkage between main and relative clauses, which needs to be addressed.
- Speech Recognizer Errors: Experiments with actual speech recognizer output (rather than clean transcripts) are planned to assess performance degradation due to recognition errors. The accuracy of the "head verb" identification is highlighted as critical.
- Word Sense Disambiguation: The current system relies on the most frequent sense from WordNet; integrating a word sense disambiguation module would improve accuracy.
- Richer Semantic Roles: The paper plans to extend the semantic frame information by processing machine-readable dictionaries to obtain richer semantic role information beyond basic subcategorization.
- Speech Act Integration: Future work may involve incorporating speech act tags to further enhance the representation.
Summary updated