Abstract
Named Entity Recognition (NER) and Entity Linking (EL) play an essential role in voice assistant interaction, but are challenging due to the special difficulties associated with spoken user queries. In this paper, we propose a novel architecture that jointly solves the NER and EL tasks by combining them in a joint reranking module. We show that our proposed framework improves NER accuracy by up to 3.13% and EL accuracy by up to 3.6% in F1 score. The features used also lead to better accuracies in other natural language understanding tasks, such as domain classification and semantic parsing.
Introduction
Understanding user intent in voice-controlled environments requires precise identification and grounding of entities. However, spoken queries present unique challenges compared to written text. They are typically short, providing minimal context, and are frequently corrupted by noise. This noise stems from two primary sources: Automatic Speech Recognition (ASR) errors (e.g., "Bohemian raspberry" instead of "Bohemian Rhapsody") and user errors, such as non-canonical nicknames or mispronunciations.
Traditional Natural Language Understanding (NLU) pipelines process Named Entity Recognition (NER) and Entity Linking (EL) sequentially. This approach suffers from error propagation: if the NER module fails to identify a span correctly, the EL module cannot recover. This paper introduces a Named Entity Understanding (NEU) framework that jointly optimizes these tasks through a reranking mechanism.
Methodology
The proposed architecture moves away from a rigid pipeline by utilizing a Joint Reranking module. Instead of passing only the top NER result, the system generates the top 5 NER hypotheses via beam search.
Core Components
- NER Module: Employs a biLSTM-CRF model with character and word-level (GloVe) embeddings to produce candidate entity labels.
- Candidate Retrieval: A retrieval engine uses n-grams, aliases, and phonetic matching to find potential matches in a large-scale Knowledge Base (KB).
- Neural Scoring & Reranking: A Multi-Layer Perceptron (MLP) scores hypotheses based on several features:
- Entity Popularity: Prior probability of an entity being requested.
- Character-level CNNs: Robustly compare spoken mentions to KB entries, mitigating transcription errors.
- Relational Features: Leverage KB relationships (e.g., song-artist) to disambiguate multiple entities within a single query.
Experimental Results
The framework was evaluated on a dataset of 600,000 anonymous English transcripts across three domains: Music, Sports, and Movie & TV.
Overall Performance
The joint reranking approach consistently outperformed the baseline across all domains. NER F1 scores improved by up to 3.13%, while EL F1 scores saw gains of up to 3.6%.
Robustness to Noise
A key strength of the NEU framework is its resilience to noisy input. On a subset of data containing ASR or human errors (approximately 13% of the total), the inclusion of the MLP-based reranker and CNN features provided significant boosts. In the music domain, performance on noisy data improved by over 10%.
Downstream Impact
The benefits of improved entity understanding extend beyond NER and EL. The features generated by the NEU system also enhanced performance in Domain Classification (DC) and Semantic Parsing (SP), demonstrating the value of high-quality entity grounding for the entire NLU stack.
Conclusion
By integrating NER and EL into a joint reranking framework, we significantly improve the robustness of voice assistants to real-world noise. The use of relational features and character-level similarity allows the system to recover from errors that would typically derail a standard pipeline. Future work will explore zero-shot linking for emerging entities and the unification of domain-specific ontologies.
Summary updated