Standardizing technical language has always been a strong necessity of the technological society. Today, Natural Language Processing as well as the widespread use of computerized document writing can give a tremendous boost in reaching the goal of standardizing technical language. In this paper, we propose two methods for standardizing language. These methods have been applied to the dataset of near misses, collected during the inspections at Major-Accident Hazard (MAH) Industries.1
Standardizing technical language has always been
a strong necessity of the technological society.
Artifacts, objects, measures and so on should have
a clear name and a clear description in order to
assure mutual understanding, which leads to the
reach of important goals in building and
controlling machines. However, language
standardization has always the same problem: language is a
social phenomenon
Nowadays, Natural Language Processing as well as the widespread use of computerized document writing can give a tremendous boost in reaching the goal of standardizing technical language. Language in use can be captured and, then, 1Copyright c 2019 for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0). analyzed. Technical people can be invited to use a standardized dictionary with writing suggestions.
This paper discusses two different methods of
standardizing technical languages, which have
been applied to a dataset of near misses coming
from the inspections at Major-Accident Hazard
(MAH) industries, named also “Seveso”
industries. . The first method aims to help a
standardization agency to propose the standard language for
writing these reports. We proposed to analyze
language in use by word embedding similarity such
that the standardization agency can propose a
language that is close to the one used. The second
method aims to reduce the use of unnecessary
synonyms in compiling reports of near misses. In fact,
using unnecessary synonyms may result in
confusing the report. For this problem, we propose to use
a combination of language modeling derived from
the CBOW model of the word2vec
The rest of the paper is organized as follows. Section 2 describes the application scenario and the dataset. Section 3 shortly reports on the models used in this study and proposes the two tasks. Section 4 reports on a preliminary analysis of the possible results of the system. Finally, Section 5 draws some conclusions and proposes further investigations. 2 2.1
The European “Seveso” Directive deals with the control of major-accident hazards involving dangerous substances, which can cause toxic clouds, fire, or explosion with consequences to people, as
Data (Date): 2007-02-15 Trasudamento OCD da serbatoio di stoccaggio OCD Durante le operazioni di riempimento del serbatoio K2 da nave cisterna, si e` notato un leggero trasudamento diOCD per corrosione del mantello (sottospessore localizzato mantello serbatoio) a quota 6 metri circa lungo il latoovest. Uno degli operatori addetto ai controlli durante la discarica della nave ha evidenziato l’evento. L’operazionedi discarica della nave cisterna e` stata fermata. Non si sono avuti rilasci, a meno del leggero trasudamento. serbatoio olio combustibile (ocd) Descrizione (Description) Fallimento procedure di manutenzione e controllo.
Azioni pianificate (Planned Actions) Fuori servizio e bonifica del serbatoio. sets and environment, also outside the establishments. All European Member States apply this Directive, which foresees periodical inspections by National Competent Authorities; in Italy, Inail is one of these authorities. During the inspection, the operator has to provide the inspectors with the list of near-misses, minor incidents, and accidents occurred in the last ten years. Near misses and minor incidents are events of losses of containment, involving dangerous substances with none or minor consequences, respectively. In Seveso industries, the registration and the analysis of near misses is strongly recommended, as they can be considered as precursors of incidents with serious consequences.
In Italy, under Seveso legislation, there are about a thousand industries, including refineries, petrochemical, and chemical. One of the pillars of the Seveso Directive is the Safety Management System SMS, whose adoption is mandatory for the establishments’ operators, in order to control major accident hazards.
The Safety Management System (SMS), implemented by the establishment’s operator, addresses technical measures and organizational procedures in order to guarantee human, asset and environmental safety, with a view to the prevention of major accident or the mitigation of their consequences.
In the recent inspections, the focus is often toward the study of the incidents and near misses (see Figure 1). The approach based on near-miss discussion is considered more “risk based” as it is able to single out the critical issues of the safety system. 2.2
The dataset refers to the near misses reports provided by the operators of “Seveso” establishments. The collection of reports on near misses, hereafter referred as REP corpus, consists of 1300 documents called ”operative experiences”. These operative experiences span the period from 2006 to 2017 and are related to 320 plants.
Each “operative experience” tells about the events occurred in the recent past (see Figure 1 for an example). Each event is registered by the operators filling in a pre-defined form. The document contains information including the date, a title summarizing the event, a short description, the reference to failed, missing or misapplied technical or procedural barriers, those that stopped the escalation and the recovering actions, and eventually the planned actions for improving the safety.
It is out of scope of this paper to discuss the
different methods used in the literature to
manage near miss information for improving the safety
management system. However, the common
objective is to exploit the valuable information
contained.
Although the form of the document is the same adopted for all operators, the compiling mode varies by the establishments and by the type of event recorded. The accuracy of the documents is not homogeneous and the interpretation of operative experience concept changes from one establishment to another; their carefulness varies on the sector activities, and often reveals the safety culture of the establishment. At a few establishments, just the releases of hazardous substance without consequences are registered. In other cases, reports include anomalies, unsafe situations, failures, and trivial errors; that is, events not directly related to major accident hazard. The documents are various, but represent truthful pictures of deviations occurred inside the establishment. 3
The overall goal is to show that existing methodologies can help in standardizing language in the specific case of reports on near misses on Seveso industries and we aim to perform this standardization with two tools: (1) analyzing similarities among words in current reports; (2) propose a methodology to help in writing these reports. 3.1
The specific case of reports on near misses is particular for several compelling reasons. The first compelling reason is that reports are written by operators belonging to sub-communities of speakers. In fact, people working in each plant can be considered a sub-community, which shares a particular language. Hence, standardizing language of reports means also harmonize sub-languages of different sub-communities, which do not interact. This problem is particularly severe when the aim is to standardize language across the whole Seveso industries. The second compelling reason is the different background of reports’ writers. Reports are in fact written by operators, which may have different knowledge, different school degree, and different cultural background. This reason makes particularly relevant the goal to help writers in compiling reports on near misses. 3.2
To meet the overall goal , we here experiment with
standard and well-assessed models and
methodologies: the notion of word embedding. In fact,
the long tradition of representing word meaning in
vectors is what is needed to: (1) help the
standardization organism to develop a common and
acceptable language; (2) devise ways to suggest more
appropriate words to writers of reports. In this study,
we used two different word embeddings:
General Language Word Embeddings
(GLwe)
We aim to provide the standardization organism, that is, INAIL, the possibility to investigate the language used in these reports on near misses. The possibility we explored is to provide a visual representation of similarity computed using similarity among word embeddings. Giving this visual representation, researchers in INAIL can devise the definition of a standard language that is built on a common and shared language. This idea is similar to what has been done in the past for terminology extraction. The real added value is that similarity among terms is computed according to word embeddings. 3.4
We aim to provide a tool to assist operators while writing reports. We explored the first capability of this tool, that is, avoiding unnecessary use of synonyms while writing. In the Italian tradition, using repeating words is seen as bad writing. Hence, when writing, synonyms are used to introduce a variation. However, for technical documents, unnecessary use of synonyms in core concepts may introduce misunderstanding. Hence, we envisage a tool that helps in reducing use of synonyms.
The algorithm governing the tool works as follows. While writing a report, the algorithm accumulate words in a set W . Whenever a new content word w is added, the algorithms compute the similarity with the words in the set W . If there is a word w0 2 W for which the similarity sim(w; w0) = wT w0 is above a threshold , the algorithm suggests w0 as a possible substitution of w. In this way, the operator is forced to (a) General Language Word Embeddings
Text La perdita non si era evidenziata al controllo dell’area effettuato preliminarmente all’inizio attivita`, ne´ rilevata dal CTM presente in zona area (sim = 0:64)
attivita` (sim = 0:31)
Necessita` di prevedere un piu` elevato grado di protezione contro la perdita di contenimento da fondo serbatoi. La fuoriuscita perdita (sim = 0:54)
contenimento (sim = 0:42) think whether the word w0 that s/he already used is similar to the word s/he is using now. In this case, w0 can be used to replace w and an unnecessary synonym is avoided. For the first task, we experimented with the two dictionaries: the General Language word embeddings (GLwe) and the Domain-adpated word embeddings (Dawe). Similarity spaces for the two word embeddings (see Figure 2) may help in understanding whether unnecessary synonyms are used and, hence, suggest a standardized word that should be used for a group of words.
Using the two dictionaries, we built two similarity spaces (Figure 2) obtained as follows. We selected 10 frequent words in the REP training corpus and, then, we presented in the two figures the top 15 words that are more similar to the 10 selected frequent words. The similarity spaces are built according to GLwe (Figure 2a) and according to Dawe (Figure 2b).
The Dawe similarity space (Figure 2b) gives apparently better hints on how words are used. The dictionary seems to be more tailored to the specific domain. In fact, there is an interesting groups of words such as favvenuto, accaduto, occorso, verificatosi g and f causato, provocatog. These gropus are missing in the GLwe similarity space (Figure 2a). For the second task, we experimented with some sample reports. The algorithm in action is reported in Figure 3. This test has been carried out on existing reports and aimed to show that some words can be replaced with previously used words. In the report #174, the word zona can be replaced with the word area, which has been previously used.
In the report #175, the word focolaio could be replaced with the word incendio. Finally, in the report #109, the word fuoriuscita can be replaced with the word perdita. However, the operator is free to accept or refuse the suggestion if this is not satisfactory. 5
Standardizing language is a need of our
technological society. In this paper, we investigated the
possibility of using modern NLP techniques to reach
this goal in the specific scenario of near misses
in Seveso Industries. Initial results on the corpus
provided by Inail are interesting and leave room
for improvement. Future model should include
the treatment of multi-word expressions by
using compositional distributional semantic models