Mining Word Boundaries from Speech-Text Parallel Data for Cross-domain Chinese Word Segmentation

TL;DR

This paper tackles cross-domain Chinese Word Segmentation by mining word boundaries from speech-text parallel data. It leverages character-level alignment via the Montreal Forced Aligner to extract pauses as potential word boundaries, then filters these boundaries using a probability-based strategy guided by a BERT-CRF baseline. A Complete-Then-Train (CTT) approach converts partial annotations into full ones and combines source-domain data with completed target-domain data to improve CWS performance, demonstrated on ZX and AISHELL2 with notable F1 gains. The approach shows that carefully filtered pauses fed through CTT effectively augment training data for cross-domain CWS, offering a scalable way to leverage naturally annotated data. Limitations include dependence on alignment quality and under-exploration of other speech cues beyond pauses, with future work aiming to expand dataset scale and richer multimodal signals.

Abstract

Inspired by early research on exploring naturally annotated data for Chinese Word Segmentation (CWS), and also by recent research on integration of speech and text processing, this work for the first time proposes to explicitly mine word boundaries from speech-text parallel data. We employ the Montreal Forced Aligner (MFA) toolkit to perform character-level alignment on speech-text data, giving pauses as candidate word boundaries. Based on detailed analysis of collected pauses, we propose an effective probability-based strategy for filtering unreliable word boundaries. To more effectively utilize word boundaries as extra training data, we also propose a robust complete-then-train (CTT) strategy. We conduct cross-domain CWS experiments on two target domains, i.e., ZX and AISHELL2. We have annotated about 1,000 sentences as the evaluation data of AISHELL2. Experiments demonstrate the effectiveness of our proposed approach.

Figures (4)

• Figure 1: An example of speech-text alignment data. The correct segmentation result is "有/人/在/细细/地/倾听", translated as "some people is carefully listening".
• Figure 2: Statistics of pauses regarding probability/accuracy of being boundaries and duration distribution. Probabilities are grouped into four bins, i.e., ${[0.0, 0.1)}$, ${[0.1, 0.9)}$, ${[0.9, 1.0)}$, and ${1.0}$. The overall percentage means the proportion of pauses belonging to a given probability bin against all pauses. Pause durations are divided into four bins, i.e., ${[10,50)}$, ${[50,150)}$, ${[150,500)}$, and ${[500,\texttt{INF})}$, in the unit of $\textit{ms}$. Given a probability bin, the internal percentage means the proportion of pauses belonging to a given duration bin against all pauses in the probability bin. For the ZX data, accuracy means the proportion of pauses that are really word boundaries according to further verification.
• Figure 3: Constrained label space for the sentence in Figure \ref{['img:example_extract_pause']}, in which we obtain two boundaries "有人/在细细地/倾听". Illegal labels are marked as gray. The red thick lines present a legal path. In this context, the character "人 (people)" is constrained to be either a single-char word ("S") or the end of a word ("E") due to the pause after it. This constraint is based on the assumption that the pause indicates a clear word boundary, preventing "人" from being labeled as the beginning ("B") or middle ("M") of a multi-char word.
• Figure 4: The CTT training strategy.