Telemedicine as a special case of machine translation

Highlights

• We created Machine Translation system for medical text domain.

• We adapted it for PL-EN language pair.

• We improved translation quality by adaptation of additional training data and interpolating language models.

• Numbers of experiments on different methodologies were performed.

• The high quality translations were obtained in evaluation process.

Abstract

Machine translation is evolving quite rapidly in terms of quality. Nowadays, we have several machine translation systems available in the web, which provide reasonable translations. However, these systems are not perfect, and their quality may decrease in some specific domains. This paper examines the effects of different training methods when it comes to Polish–English Statistical Machine Translation system used for the medical data. Numerous elements of the EMEA parallel text corpora and not related OPUS Open Subtitles project were used as the ground for creation of phrase tables and different language models including the development, tuning and testing of these translation systems. The BLEU, NIST, METEOR, and TER metrics have been used in order to evaluate the results of various systems. Our experiments deal with the systems that include POS tagging, factored phrase models, hierarchical models, syntactic taggers, and other alignment methods. We also executed a deep analysis of Polish data as preparatory work before automatized data processing such as true casing or punctuation normalization phase. Normalized metrics was used to compare results. Scores lower than 15% mean that Machine Translation engine is unable to provide satisfying quality, scores greater than 30% mean that translations should be understandable without problems and scores over 50 reflect adequate translations. The average results of Polish to English translations scores for BLEU, NIST, METEOR, and TER were relatively high and ranged from 7058 to 8272. The lowest score was 6438. The average results ranges for English to Polish translations were little lower (6758–7897). The real-life implementations of presented high quality Machine Translation Systems are anticipated in general medical practice and telemedicine.

Introduction

Statistical Machine Translation (SMT) is the translation of the text by a computer, with no human involvement. SMT systems have no knowledge of language rules. Instead, they “learn” to translate by analyzing large amounts of data for each language pair. They can be trained in specific industries or disciplines using additional data relevant to the sector needed. Typically, SMT systems deliver fluent-sounding but less consistent translations.

The machine translation is evolving quite rapidly in terms of quality. Currently, several machine translation systems are available on the web that provides reasonable translations. Developed systems are not perfect, and their quality may decrease in some specific domains. In addition to this, the scientific community is involved in machine translation. It must be pointed out that the scientific organizations, conferences, and events dedicate great effort to its improvement. One of the biggest advantages of machine translation is that most users do not require perfect translations [1]. Users may only be interested in roughly understanding a text simply to get an idea of what the text is about. However, other users may not be that flexible. For example, the correctness and beauty of writing in medicine may not be important, but the precision and adequacy in the translated message is crucial. In medical communication, a translation error between the patient and the physician, or an error in communication regarding treatment or a diagnosis may have serious consequences for a patient's health [2].

Recently, due to the growing success of an interest in language technologies, machine translation has been applied to the field of medicine. For example, one study [3] analyzed the feasibility of post-editing machine translations of health-promotional English documents from local and national public health websites in the USA. It was assumed, a priori, that machine translation would not provide a high enough quality for the documents to be used as official versions. Despite that, language technologies are steadily increasing in quality. It should be expected that, in the not-too-distant future, machine translation will be capable of translating any text in any domain with the required quality.

The medical data domain is, in our opinion, a very narrow, but relevant and promising field of research for language technologies. MT systems can be used for translation of medical records of any kind. Accessing and translating a foreign patient's medical history might even save their life. Preparation of direct speech-to-speech translation systems is also possible. The foreign patient's speech is recognized using an Automated Speech Recognition (ASR) system. After recognition, the speech is translated into another language and synthesized in real-time. For example, the EU-BRIDGE project aims at developing automatic transcription and translation technology that will permit the development of innovative multimedia captioning and translation services of audiovisual documents between European and non-European languages [http://www.eu-bridge.eu].

Obtaining and providing medical information in comprehensive ways appears to be of crucial importance for both patients and physicians [4], [5], [6], [7]. For example, as emphasized by Healthcare Technologies for the World Traveler (HTH) [8], a foreign patient may require an explanation and description of their diagnosis and comprehensive information about available treatment options. In several countries, many residents and immigrants communicate in languages other than the official one.

According to Karliner et al. [9], it is necessary to analyze how human translators could enhance access to health care, including improvement of its quality [10]. Nevertheless, human translators experienced in telemedicine information are very often unavailable for both patients and medical professionals [11]. Although existing machine translation capacities are imperfect [11], machine translation must ensure the reduction of costs associated with medical translation. On the other hand, it is necessary to increase its availability and quality [12].

Medical professionals, researchers, and patients require adequate access to the abundance of telemedicine information on the Internet [6], [13]. This information can potentially improve our health and well-being. Sharing medical information could improve medical research, as well. English is the most dominant language used in medical science, but not the only one.

Polish is considered to be the one of the most challenging West-Slavic languages, due to its complexity. It is a tough language for an SMT system. For example, Polish grammar, includes complicated rules and elements, including an immense vocabulary (thanks to its complex declension). Nearly free word order in sentences is also problematic. All of these are the main reasons for its challenging character. In addition, the Polish language includes 7 cases and 15 gender forms for both nouns and adjectives.

As expected, these facts strongly influence the data and data structure used in statistical translation models. The lack of available and appropriate resources necessary for data input to SMT systems presents another problem. SMT systems give the best results for concrete and narrow text domains. The proper quality of the parallel data, including the required domains, has inadequate availability. On the other hand, Polish and English differ strongly in syntax. Above all, English is a positional language. This means that the syntactic order, which includes the word order of one sentence, has an invaluable significance, especially because of the limited inflection of words (for example, lack of declension endings). Sometimes, the position of the word in a sentence is the only indicator of the sentence meaning. As far as English sentences are concerned the subject comes before the predicate. Therefore, a sentence is structured in Subject–Verb–Object (SVO) word order. In contrast, Polish simply has no particular word order. Additionally, the word order itself has no decisive impact on a sentence's meaning. In Polish, one can express the same idea in many ways, which is simply not possible in English. For instance, the sentence “I have bought myself a new car.” can be expressed in Polish as “Kupiłem sobie nowy samochód”, or “Nowy samochód sobie kupiłem.”, or “Sobie kupiłem nowy samochód.”, or “Samochód nowy sobie kupiłem.” As one can see, changes in word order influence the complexity of the translation process.

As a consequence, the development of SMT systems for the Polish language has been considerably slower in comparison to English and other languages. The primary goal of this research is to develop an SMT system for translation from Polish to English language and vice versa, with an emphasis on medical data. This paper has the following structure: Section 2 contains an introduction to the preparation of Polish data. Section 3 presents the English language issues. Section 4 describes the methods associated with translation evaluation. Section 5 presents the results. Sections 6 Discussion and conclusions, 7 Future work provide the summary of potential implications and opportunities for future work.