Japanese Named Entity Recognition from Automatic Speech Recognition Using Pre-trained Models

Japanese Named Entity Recognition from Automatic Speech Recognition Using Pre-trained Models Seiichiro Kondo1 Naoya Ueda1 Teruaki Oka1 Masakazu Sugiyama2 Asahi Hentona2 Mamoru Komachi1 1Tokyo Metropolitan University 2 AI Shift {kondo-seiichiro, ueda-naoya}@ed.tmu.ac.jp, teruaki-oka@tmu.ac.jp {sugiyama_masakazu, hentona_asahi}@cyberagent.co.jp, komachi@tmu.ac.jp Abstract Japanese named entities extracted from auto- matic speech recognition frequently contain speech recognition errors and unknown named entities due to abbreviations and aliases. One possible solution to this problem of the named entity extraction task is to use a pre-trained model trained on a large quantity of text to ac- quire various contextual information. In this study, we performed named entity recognition on the logs of a task-oriented dialogue system for road traffic information in Fukui, Japan, using pre-trained BERT-based models and T5. In our experiments using our prepared data, the F1 scores of BERT and T5 are higher than that of string match by 20.2 point and 21.1 points, respectively. The results confirmed that these pre-trained models exhibited significantly higher accuracies on unseen entities than meth- ods based on dictionary matching. 1 Introduction In interactive voice response services, it is neces- sary to accurately perform entity extraction and mention detection from the speech recognition re- sults of user’s speech to carry out dialogue via voice. However, this input contains a variety of noise, including speech recognition errors, com- pared to written text. Moreover, users’ utterances may include personal representations for a certain entities. In our full research, we plan to address these issues in two steps: named entity recognition (NER) and entity linking (EL). The first step, NER, recognizes a part of the text that is likely to repre- sent a named entity. In the second step, we then attempt to link it to a prepared set of entities even if it contains errors. With this architecture in mind, in this study, we tackle the first step, which is NER from speech recognition results. This study focused on named entity recognition (NER) in the context of a task-oriented dialogue system that provides information in response to the user’s requests pertaining to road traffic. In Figure 1: Flowchart of our envisioned spoken dialogue system. This study focuses on the NER component. The goal is to extract named entities even from noisy text caused by speech recognition errors or abbreviations. our system, NER is accomplished by linking the automatic speech recognition (ASR) text with a dic- tionary created for each task (see Fig. 1). One way to achieve more accurate recognition is to extract the named entities before the linking. Although we enforce the inputs to include loca- tion names by system-driven conversation, speech recognition errors and unknown named entities from abbreviations and aliases may occur. For example, “いちごっぱ” (ichi-go-ppa, 158) is a colloquial expression for “国道158号” (kokudou- hyaku-goju-hachi-gou, Japan National Route 158). In addition, there are many cases in Japanese where the sounds are close, but the surface forms and meanings are entirely different. Therefore, in this setting, NER using conventional methods is diffi- cult, especially for rule-based methods. To improve text processing functionality, …

Content truncated to 500 words