Hybrid extraction of multi-word terms: an application on vibration-based condition monitoring technique

1. Introduction

Every scientific area makes use of a special vocabulary to convey specialized concepts by means of technical language. This special vocabulary contains a wealth of information that was gathered over the years in a particular domain. Terminologies, which are the lexical components of specialized languages, have value in the way they condense mass of information into single-word, multi-word or compound word units. They are a crucial component of both technical and scientific writing, ensuring more effective communication. Terminology identification and documentation is a time consuming task, requiring manual validation by humans. It is also subjective and depends on the experience and criteria of the experts who validate it.

Condition monitoring (CM) utilizes specific measurement of equipment parameters, such as the vibrations in a machine, its temperature, or condition of its oil, assessing significant changes (if any) that may be indicative of an impending failure. CM offers a multitude of benefits, including 24×7 remote monitoring, early warnings of potential (serious) failures, reduced unplanned downtime, reduction in maintenance costs, and support for ongoing reliability and risk reduction with fewer inspections. The vibration-based CM is a system which uses non-destructive non-destructive sensing and analysis of the characteristics of the system in the time, frequency or modal domains for the purpose of detecting a change that can indicate damage or degradation.

Different technical terms, jargon, acronyms, and concepts are used in vibration analysis domain. There is a lot of semantic confusion in the area of vibration-based CM when general or vague terms like frequency, frequency response, frequency response function, etc. are used. Specifically, vocabularies referring to a specific condition are used interchangeably with terms denoting general dysfunction and vice versa. These confusions may reflect research that has changed dramatically in the last decades. We aim to eliminate this confusion by providing a stable and unambiguous terminology work of vibration-based CM.

While terminology recognition, i.e., development of a list with domain specific vocabulary from a textual corpus, has become widespread presently, the notion itself of term is nevertheless not clear, both from a pure linguistic and a computational perspective [1]. Juan C. Sager [2] defines a term as a depository of knowledge and a unit with specific reference in that it “refers to discrete conceptual entities, properties, activities or relations, which constitute the knowledge space of a particular subject field”. Similarly, Christian Jacquemin [3] and Maria Teresa Pazienza [4] define a term as “a surface representation of a specific domain concept”.

A considerable amount of work has been done in order to define terms for a specific domain. They use statistical methods to extract terminologies from textual corpora, which is a collection of documents in the domain. Many of the results are completely useful. Gerard Salton [5] proposed TF-IDF. It is a statistical measure used to evaluate how relevant a term is to a document within a collection of documents. This is done by considering: (i) how many times a term appears in a document; and (ii) how frequent a term appears across a set of documents in the reverse order. Later, Kenneth Church and Patrick Hanks [6] suggested the mutual information (MI) which is an important concept in information theory. Particularly, it measures how much information, on average, is communicated in one random variable about another. Both variables are sampled at the same time. Ted Dunning [7] proposed a probability measure, which relies on frequency profiling and is based on likelihood ratios. The method produces reasonable results and works for both large and small text samples. Relatively recently, researchers have captured $n$-grams (i.e., occurrences of one or more words) using corpus comparison based on normalized frequency. For instance, Adam Kilgarrif [8] experiments with several measures, among others χ2-test, Mann-Whitney rank, t-test, MI, and TF-IDF and concludes that χ2-test performs best. However, in general, most methods produce endless lists of candidate terms.

In this paper, we present a method for the automatic extraction of multi-word terms from machine-readable textual corpora. We apply the method to a corpus of the vibration-based CM technique domain, which is a keytool in the predictive maintenance of any equipment and machines used in a wide range of industries including transport, oil and gas, processing and manufacturing. The key idea of our approach is to reduce the noisy terms as well as the out-of-domain terms by using two different corpora: a) a domain specific corpus and b) a general purpose corpus. Our experimental results show that our method can extract multi-words terms, i.e., a set of two or more words, with nice correspondence to domain-specific concepts and notions. Thus, the human effort is reduced and it is subjective independent. The rest of this paper is structured as follows: we analyze our terminology extraction method in Section 2. Section 3 presents the results of our experiment as well as the error analysis. Finally, Section 4 concludes our work.

2. Our terminology extraction approach

Automatic term extraction is an open problem of natural language processing. The task is to identify and extract terminological units from textual corpora. In this paper, we extend the methodology proposed by Patrick Drouin [9], where the extraction of terms is done in domain-specific corpora with the help of general corpora. Using TF-IDF and stop-word lists, we restrict terms that are sub-terms of other terms in the list of candidate terms in order to reduce the number of candidate terms.

Our algorithm emphasizes in identification of verb or noun phrases multi-word terms. For verb phrases, the method excludes auxiliary and light verbs and takes their base forms as candidates. For noun phrases, the method extracts structures of noun-noun (NN), adjective-nouns (AN) and its combinations, including conjunctions (C) or determiners (D). It also recognizes unknown words (e.g., words of the domain-specific corpus which are not found in the general corpus) because most technical jargon is unlikely to be included in a general dictionary. To do so, we split the task into three steps: (i) corpus creation, (ii) term identification and, (iii) automatic validation.

3. Results

The list of the candidate terms consists of 20,000 English entries obtaining a total of 9,000 term candidates (i.e., 45 % of the total) after filtering and the automatic validation. After filtering out the unigrams which are out of scope of this research, 4,000 term candidates remained (i.e., 20 % of the total). Table 2 shows the number of n-grams for each step of the automatic term extraction.

Some sample fragments of the extracted terms with a high, medium and low frequency of occurrence in the corpus are shown in Table 3. There are no units, acronyms, or abbreviations such as FRF, FTF, FFT, etc. in the following list.

Table 4 shows some examples of candidate terms, grouped by their structure. The lexical items on this list met the probability threshold used in the identification process. The list needs to be further cleaned by hand. For instance, the candidate term electrotechnical standardization with a high frequency would most likely be removed from the list if it is not a representative term of the domain of vibration analysis. This is considered normal given that we have a corpus of scientific papers with standard phraseology used by authors. Consequently, all candidate terms and in particular those of high frequency, such as electrostatic effect and absolute maximum, have to be further validated by experts in the vibration-based CM domain to be representative of their domain.

Validation is done by taking into account: (i) if the term is representative of the domain, (ii) if the term is neutral and free of judgements, (iii) if the term has been deprecated, and, (iv) if the term is not ambiguous. Metadata are attached to each term; namely, a list of contexts, the POS, abbreviations, the frequency, etc. (see Fig. 1). The experts could either approve or reject the candidates based on their experience in the area. In addition, they assign definitions to the terms and add more metadata, if necessary.

Fig. 1Metadata displayed for the cacidate term acoustic emission

For this reason, 2 hours of training was offered for the experts to introduce them to the terminography. It should be noted that we acknowledge and recognize the valuable terminology work made by the International Organization for Standardization with the help of the International Electrotechnical Commission (IEC) [16] as well as by the Vibration Institute (https://www.vi-institute.org/vibration-terminology-project). We use definitions from both contributions in our work and expand the metadata for the terms.

Based on their relationship to concepts, terms are classified according to the principles of terminography; (i) synonyms are grouped together in a single entry and (ii) homonyms and polysemes (i.e., words with the same pronunciation and/or spelling) are presented separately (different entries) since they represent different concepts. Furthermore, the term formation shall be concise and as neutral as possible, avoiding connotations, especially negative ones [17].

To prepare for and familiarize themselves with the documentation process, experts started with the validation of 20 ambiguous but typical terms in the domain of vibration-based CM [18]. These terms have broad meanings, and are employed in a variety of disciplines including music, physics, acoustics, electronic power transmission, radio technology, and other domains. For example, among others, the term node (http://www.electropedia.org/iev/iev.nsf /6d6bdd8667c378f7c12581fa003d80e7?OpenForm&Seq=2), according to IEC’s electropedia is used across the domains of acoustics and electroacoustics, circuit theory, mathematics, telecommunication networks, teletraffic and operation, electric traction. In the same way, the term filter (http://www.electropedia.org/iev/iev.nsf/SearchView?SearchView&Query=field +SearchFields+contains+filter+and+field+Language=en&SearchOrder=4&SearchMax=0) is used across the domains of radiology and radiological physics, oscillations, signals and related devices, electrical and magnetic devices, lighting, etc. It is important to note that a term from one domain within the same language can be borrowed and attributed to another concept within another domain. Thus, a term may have a conflicting meaning with its general language meaning, or it may confuse the multiplicity of technical meanings.

After a term is verified, it is converted into machine-readable XML (According to [19]) format. The term entry harmonic is shown in Fig. 2. More specifically, in the subject domain of vibration CM, the English term harmonic is formed as a single-word term, via terminologization of the ordinary term harmonic, meaning “repeating signals, such as sinusoidal waves” in order to render the concept “harmonic vibration, the frequency of which is an integral multiple of the fundamental frequency”. It is a single, masculine noun, and its deprecated term is overtone because “the term “overtone” has frequently been used in place of harmonic, the $n$th harmonic being called the ($n–1$)th overtone”. An example of how to use this term in the domain of vibration is “Some of these harmonics have a dominant value in the vibration spectrum due to interaction of machine flu harmonics and the mechanical structure of the machine”.

Fig. 2Metadata displayed in XML format for the candidate term harmonic

As it is mentioned above, experts with strong knowledge of the domain of the vibration-based CM validate manually the term candidates. At the same time, we perform an error analysis so as to better understand the behavior of our method. For that purpose, we took 5 % random examples (400 candidate terms) from each $n$-gram category and manually analyzed each of the errors made by the algorithms. Our analysis first reveals that the rejected terms make up 4.5 % of the sample.

As for the actual errors, we observe that nearly a third of them correspond to out-of-domain. For instance, terms like network error, limited $\alpha$-frequency range, angle asymmetries, etc. are not relevant to vibration-based CM domain. We also observe a few errors that are related to nested terms. In our work we filter out candidate substrings of a candidate term taking into consideration its frequency by itself in the corpus (i.e., within different surrounding words) [14]. However, not all cases are covered. For instance, in the term candidate list there are examples like wind turbine vibration and wind turbine vibration signal/wind turbine vibration data, transient elastic wave and transient elastic wave generation. For the remaining errors, terms that are syntactically interrupted by [A] are rejected by the experts. For instance, angular frequency, angular samplic frequency and angular supply frequency.

All in all, our error analysis reveals that our method performs well and based on the feedback the algorithm will be able to filter out the extracted list. Moreover, it shows that the quality of the multi-word term candidates is much better than what the statistical approaches encouraging the extension of the algorithm in other applications or the integration of other techniques.

4. Conclusions

In this work, we propose a method for domain-specific terminology extraction of groups of words that commonly occur together. We use an advanced filtering strategy to exclude out-of-domain term candidates and/or nested terms. Thy hybrid methodology we designed, benefits from both statistical and linguistic approaches in order to provide a more accurate and precise term selection. In addition, we use two different types of corpora, (domain-specific vs. general-purpose corpus) to remove the noise. Our algorithm emphasizes in identification of verb or noun phrases multi-word terms. For verb phrases, it excludes auxiliary and light verbs and takes their base forms as candidates. For noun phrases, it extracts structures of [N N], [A N] [A A], and [V] or [V N] or combinations, including [C] or [D]. In addition, it takes into consideration the identification of neologisms and technical jargons.

Our study based on a vibration-based CM domain corpus shows that our method can export high quality domain-specific terminology from a small corpus. Through the error analysis results, we can see that most extracted terms have nice correspondence to the domain of CM concepts and notions. Hence, our method reduces the human effort as well as helps in a wide-range and representative term selection for a given domain.