Functionality and goals.

Summary of the results, published in October 2009 in Communications of the ACM

Overview on IST Results web site

The aim of the TransType project is to develop a new kind of interactive tool to assist translators. The system observes a translator as he or she types a text and periodically proposes extensions to it, which the translator may accept as is, modify, or ignore simply by continuing to type. With each new keystroke the translator enters, TransType recalculates its predictions and attempts to propose a new completion to the current target text segment. The following snapshot shows a TransType screen during a translating session:

There are several ways in which such a system can be expected to help a translator:

  • when the translator has a particular text in mind, the system's proposals can speed up the process of typing it
  • when the translator is searching for an adequate formulation, the proposals can serve as suggestions
  • when the translator makes a mistake, the system can provide a warning, either implicitly by failing to complete the mistake or explicitly by attaching an annotation for later review

The novelty of this project lies in the mode of interaction between the user and the machine translation (MT) technology on which the system is based. In previous attempts at interactive MT, the user has had to help the system analyze the source text in order to improve the quality of the resulting machine translation. While this is useful in some situations (for example, when the user has little knowledge of the target language), it has not been widely accepted by translators, in part because it requires capabilities (in formal linguistics) that are not necessarily those of a translator. We feel that using the target text as the medium of interaction will result in a tool that is more natural and useful for translators .

This idea poses a number of technical challenges. First, unlike standard MT, the system must take into account not only the source text but that part of the target text which the user has already keyed in. The tool should limit its proposals to those in which it has reasonable confidence; and it should adapt its behavior to the current context, based on what it has learned from the translator. Second, the system must be able to operate in real time and be fast enough to keep up with its user. Finally, it must display its proposals clearly but unobtrusively, and provide an efficient mechanism by which they can be accepted.

History, research personnel and partners

The first phase of the project (1997-2000) was funded by the Natural Sciences and Engineering Research Council of Canada (NSERC).

A second phase was launched in March 2002 and ended during the summer of 2005. Two new languages (Spanish and German) and several technological enhancements were added to the initial prototype in order to improve its prediction capabilities and usability. For more on the second phase of the project the TT2 flyer (PDF 100K).

TT2 involved international and Canadian partners.

At the RALI, the principal researchers who were involved in the project are: Guy Lapalme, Philippe Langlais, Elliott Macklovitch, Simona Gandrabur and Ngoc Tran Nguyen.

In Canada, TransType was funded by:

The Technology: realizations and challenges

In order for such a tool to actually speed up the translation process and not hinder it, there are several important factors that come into play. The RALI's past and present work within this project have revolved around all of these axes.

Accuracy and speed of the translation predictions.

The more accurate the predictions are, the better the chances the user will accept them and save typing time. There is often a trade-off to be made between accuracy and prediction speed. Various strategies to tackle this problem have focussed on improving the statistical translation models and decoding techniques.

An innovative Maximum Entropy approach for fast text prediction of sentence fragments was proposed by Foster et al. ([Article 1] [Article 2]). Langlais et al. (publications) explored various word-based and, more recently, some phrase-based and translation memory approaches to modeling, and different decoding strategies. Finally, some experiments have been done with rescoring techniques based on a machine learning approach (publications).

Benefit to the user.

This kind of stochastic tool will generally be able to present the user with numerous translation predictions of various lengths. Given the potential ambiguity of the source text and the inadequacies of current machine translation techniques, translation errors can't be avoided. It is therefore important that the system be able to choose the optimal predictions. In order to assess optimality, we first need to be able to estimate the benefit to the user in terms of typing speed gains versus time lost by reading wrong predictions. Benefit estimations are carried out by means of a user-model built from data collected during user evaluations (publications).

Secondly, we need an accurate estimate of how reliable the individual translation predictions are. This is what we call confidence estimation for machine translation. The RALI was the among the first research laboratories to explore confidence measures and confidence estimation techniques for machine translation (publications). These were inspired by similar techniques that are widely used in Automatic Speech Recognition by which had not yet been applied to other areas of Natural Language Processing. As a result of this innovative topic, two members of the RALI organized a summer workshop at Johns Hopkins University's Center for Language and Speech Processing in 2003.

User-friendliness of the human/machine interface.

The RALI was responsible for the implementation and optimization of the graphical user interface on the TransType2 project. Numerous issues had to be addressed in order to meet user expectations and demands, as well as to integrate the translation engines provided by three different research laboratories within a single GUI. Moreover, the last version of the prototype included a speech interface which allowed a user to interact with the application by issuing certain vocal commands.

User Evaluation

An avant-garde tool like this one has to be tested and its impact evaluated in authentic working conditions, with human translators attempting to use it on realistic translation tasks. The RALI was responsible for coordinating user evaluations and compiling the evaluation results in terms of comprehensive statistics. In order to facilitate this task, we developed a tracing and off-line simulation tool called TTPlayer, which enables us to replay any particular translation session and analyze it in the most fine-grained detail. Data collected from these user evaluations was extremely valuable for assessing the strengths and weaknesses both of the system and the evaluation process itself. Moreover, this data could form the basis for developing other user-adaptive and domain-adaptive translation techniques.

In the project's final evaluation rounds, most of the translators participating in the trials registered impressive productivity gains, increasing their productivity from 30 to 55% thanks to the completions provided by the system. The trials also served to highlight certain modifications that would have to be implemented if the TransType prototype is to be made into a translator's every day tool: e.g. the ability of the system to learn form the users' corrections, in order to modify its underlying translation models.

The final results of TransType2 trials have been published in

Elliott Macklovitch. TransType2: The Last Word. Proceedings of LREC 2006, Genoa, Italy, may 2006.