A few months ago, I waited in anticipation as my computer installed the English module of Antidote. I had been tracking its progress for several years, and finally the release day had arrived. As a long-time user of Antidote’s French module, I had become concerned that my perception of its capabilities might not be entirely accurate. I was also interested in assessing the performance of the new English module. So I decided that it would be beneficial to conduct an experiment to test Antidote’s capabilities.
The Antidote software package includes a spelling/grammar checker, several dictionaries and detailed language guides. Although not specifically designed for translators, it is a useful revision tool when used with caution and realistic expectations. Over two decades of design and development have culminated in the current version, Antidote 9, which is available for Windows, Mac and Linux. The software integrates into various Microsoft Office programs including Outlook email, placing an Antidote button on the menu bar, and a visual editor is offered as an add-on. Mobile applications of the dictionaries and language guides are available, although they do not give access to the grammar checker.
Correction software has become an integral tool for many translators, and is a standard component of translator workbench systems. Antidote is quite flexible in that its parameters are adjustable. For example, the software is customizable in regard to fluency levels, national varieties of language, regional expressions, and personal dictionaries. However, it is crucial that the user not be lulled into a false sense of security.
While the software has some helpful analysis tools (pragmatic, stylistic, semantic, lexical and logical), the scope of my evaluation was limited to the capabilities of its spelling/grammar checker. Antidote’s official website makes the following claim: “D’un seul trait, le correcteur d’Antidote souligne toutes les fautes, de l’accent oublié à l’accord difficile, de la virgule malvenue au pléonasme bête.” The purpose of the experiment I conducted was to determine if the software actually lives up to that declaration.
It seemed to me that a test of several common categories of errors should provide a good idea of Antidote’s capacity to recognize errors. Knowing that some language rules are easier to follow than others, I proceeded on the hypothesis that the software’s effectiveness varies according to the type of error in question. After thinking about the factors involved in different types of errors (such as specific rules, possibility of ambiguity, need for analysis beyond the word level, etc.), my speculation was that Antidote’s performance would be: almost perfect for typography; high for verb conjugation, agreement and punctuation, and low for prepositions and tricky grammar rules.
For the experiment, I chose an excerpt from an English book and its accompanying French translation. I inserted 25 errors of a certain type in each text and tested Antidote’s response. Then, using the original excerpt, I inserted 25 errors of a different type to test them separately. I then summarized the software’s capacity of recall expressed as a percentage: the number of errors it identified divided by the number of errors that had been inserted in the text. The results were somewhat surprising, and are noted on the chart below. It was disappointing to realize that, overall, the long-awaited English module functions less efficiently than the French module. Even more unsettling was the fact that Antidote seems to be completely unaware of English prepositional errors.
|Recall Capacity of Antidote||French||English|
|Agreement (adjectives, determiners, etc.)||94%||86%|
|Tricky Grammar Rules||75%||28%|
|(French subjunctive; English “ing” as gerund or participle)|
My remarks must be considered in the context of the specific types of errors (and the limited number thereof) that were tested during this evaluation, and not be construed as a comprehensive analysis of Antidote’s recall capacity. However, the results are a good indication of its capabilities, revealing that grammatical correction software is limited by its incapacity to perform certain tasks that are difficult to automate at the sentence level.
From my evaluation, I feel that I can state unequivocally that Antidote software does not fulfill its unrealistic claim of being able to point out every error in a text. However, the French module definitely performs better than the English module. It would be interesting to see what conclusions would emerge from a test of other types of errors. Time did not permit me to study such things as: direct and indirect verbs, verb tenses besides the indicative and subjunctive, reflexive verbs, missing and incorrect capitalization, past participle agreement, relative pronoun agreement, incomplete negatives, and pleonasms, for example.
It appears that Antidote’s weak points fall into the categories of punctuation and prepositions: these results were far below my expectations. It seems odd that the French software has the capacity to deal fairly well with a complicated issue such as the use of the subjunctive tense, yet fails to recognize blatant punctuation errors, such as an exclamation mark inserted randomly in the middle of a sentence. The latter result is quite ironic, considering the fact that “la virgule malvenue” is specifically mentioned in Antidote’s website assertion. I would like to see the developers integrate more specific punctuation guidelines to improve this area.
As well, even though prepositions are difficult to master, it seems that the software could be programmed to at least detect fixed expressions. The English module in particular needs to be improved with respect to prepositions. For instance, the phrases “put my ideas at print”, “my final year on university” and “give perspective at this subject” did not even generate a blip on Antidote’s radar screen. The capacity to think on the sentence level is obviously still out of reach for a machine. So, fellow translators, we are not in any immediate danger of losing our livelihood. There is no remedy more effective than a conscientious revision, but Antidote can help facilitate that task.
Liane Johnston Grant received her BA and MA in translation at Concordia University. In 2011, she founded The King’s Translators, a group of volunteers doing religious translation. Liane is the administrator and chief editor of the group, using the experience to forward her doctoral research at Université de Montréal on quality assurance for non-professional translation projects.