End-to-End Chinese Speaker Identification

Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2274 - 2285 July 10-15, 2022 ©2022 Association for Computational Linguistics End-to-End Chinese Speaker Identification Dian Yu1 Ben Zhou2 Dong Yu1 1Tencent AI Lab, Bellevue, WA 2University of Pennsylvania, Philadelphia, PA {yudian, dyu}@tencent.com, xyzhou@seas.upenn.edu Abstract Speaker identification (SI) in texts aims to iden- tify the speaker(s) for each utterance in texts. Previous studies divide SI into several sub-tasks (e.g., quote extraction, named entity recogni- tion, gender identification, and coreference res- olution). However, we are still far from solv- ing these sub-tasks, making SI systems that rely on them seriously suffer from error prop- agation. End-to-end SI systems, on the other hand, are not limited by individual modules, but suffer from insufficient training data from the existing small-scale datasets. To make large end-to-end models possible, we design a new annotation guideline that regards SI as span ex- traction from the local context, and we annotate by far the largest SI dataset for Chinese named CSI based on eighteen novels. Viewing SI as a span extraction task also introduces the pos- sibility of applying existing storng extractive machine reading comprehension (MRC) base- lines. Surprisingly, simply using such a base- line without human-annotated character names and carefully designed rules, we can already achieve performance comparable or better than those of previous state-of-the-art SI methods on all public SI datasets for Chinese. Furthermore, we show that our dataset can serve as additional training data for existing benchmarks, which leads to further gains (up to 6.5% in accuracy). Finally, using CSI as a clean source, we design an effective self-training paradigm to continu- ously leverage hundreds of unlabeled novels. 1 Introduction Speaker identification (SI) aims to identify the cor- responding speakers for utterances in texts (Zhang et al., 2003; Glass and Bangay, 2007). Most exist- ing SI datasets (He et al., 2013; Chen et al., 2021) provide ground-truth character aliases and utter- ance spans as inputs. However, such annotations are unavailable in realistic settings, under which SI is usually divided into interrelated sub-tasks (Pan et al., 2021; Yoder et al., 2021) (e.g., utterance identification, named entity recognition, corefer- ence resolution, and candidate speaker generation). However, this pipeline faces several challenges. First, these modules are imperfect, and they in- evitably introduce errors that propagate and se- riously affect the final performance (C-I). For example, the performance of the state-of-the-art coreference resolution model is about 80.3% in F1 (Kirstain et al., 2021). Second, classical SI datasets and approaches assume that the speaker to be linkable to one of the named entities, which cannot handle the more realistic settings where the speakers are not humans or only exist as nomi- nals (e.g., “a young girl" or “smartwatch") (C-II). Third, features (e.g., speech verb list and position information) and rules are usually carefully cre- ated and selected by experts for a certain language, which may make these resources difficult to be used for other languages (C-III). Finally, one of …

Content truncated to 500 words