Integrating machine learning with linguistic features: A universal method for extraction and normalization of temporal expressions in Chinese texts.

BACKGROUND AND OBJECTIVE

With the rapid development of information dissemination technology, the amount of events information contained in massive texts now far exceeds the intuitive cognition of humans, and it is hard to understand the progress of events in order of time. Temporal information runs through the whole process of beginning, proceeding, and ending of events, and plays an important role in many natural language processing applications, such as information extraction, question answering, and text summary. Accurately extracting temporal information from Chinese texts and automatically mapping the temporal expressions in natural language to the time axis are crucial to understanding the development of events and dynamic changes in them.

METHODS

This study proposes a method integrating machine learning with linguistic features (IMLLF) for extraction and normalization of temporal expressions in Chinese texts to achieve the above objectives. Linguistic features are constructed by analyzing the expression rules of temporal information, and are combined with machine learning to map the natural language form of time onto a one-dimensional timeline. The web text dataset we build is divided into five parts for five-fold cross-validation, to compare the influence of different combinations of linguistic features and different methods. In the open medical dialog dataset, based on the training model obtained from the web text dataset, 200 disease descriptions are randomly selected each time for three rounds of experiments.

RESULTS

The F1 of multi-feature fusion is 95.2%, which is better than the single-feature and double-feature combination. The results of experiments showed that the proposed IMLLF method can improve the accuracy of recognition of temporal information in Chinese to a greater extent than classical methods, with an F1-score of over 95% on the web text dataset and medical conversation dataset. In terms of the normalization of time expressions, the accuracy of the IMLLF method is higher than 93%.

CONCLUSIONS

IMLLF has better results in extracting and normalizing time expressions on the web text dataset and the medical conversation dataset, which verifies the universality of IMLLF to identify and quantify temporal information. IMLLF method can accurately map the time information to the time axis, which is convenient for doctors to intuitively see when and what happened to the patient, and helps to make better medical decisions.