Friday, November 22, 2013

The ALPAC Report: The Failings of Machine Translation

One of the organisations interested in
the potential of machine translation.
Not long ago, we had a look at the birth of machine translation (MT) with the Georgetown-IBM experiment. Following the experiment, optimism was at an all-time high for MT, and the problem was expected to be solved promptly. Today we're looking at the next important milestone in early MT, the ALPAC Report. Unfortunately, our tale includes a lot of government bodies and research groups, so expect a lot of acronyms.

In the US, the Department of Defense, the National Science Foundation, and the Central Intelligence Agency (CIA) were very interested in the prospect of automatically processing languages and MT. In the case of the Department of Defense and the CIA, this was mainly because the US was extremely curious and sceptical of the Russians and wanted to know what they were up to. By 1964 they had promoted and funded work in the field for almost a decade, and together the three organisations founded the Joint Automatic Language Processing Group (JALPG).

In 1964, JALPG set up the Automatic Language Processing Advisory Committee (ALPAC) in order to assess the progress of research. ALPAC was, in essence, founded by the US Government to ensure that funds were being spent wisely.

John R. Pierce, head of ALPAC.
The group was headed by chairman John R. Pierce, an employee of Bell Labs, who was assisted by various researchers into MT, linguists, a psychologist and an artificial intelligence researcher. They worked together in order to produce the 1966 ALPAC report, which was published in November of that year.

Titled "Languages and machines: computers in translation and linguistics", the report would appear to have a focus not only on MT, but also on computational linguistics as a whole. However, the report viewed MT very narrowly, from the perspective of its applications in terms of the US government and military, and how they could use the technology exclusively with the Russian language.

The report showed that since most scientific publications were in English, it would actually be quicker and therefore more cost-effective to learn and read Russian than to pay for translations into English. They also noted that there were an abundance of translators and that the supply of translators outweighed the demand for them, meaning that there was even less demand for research into MT to replace human translators.

While the report evaluated the translation industry in general, it also covered research into MT. It condemned the work done in Georgetown, as there was little evidence to support quality translations from the same place that had spawned the idea that the MT issue was close to being solved.

In fact, Georgetown's MT project had produced no translations of scientific texts, nor had it any immediate plans to do so. The report had defined MT as a process that required no human interaction and the fact that Georgetown's work still required human post-editing left ALPAC to deem it as a failure.

One of the criticisms of the unedited output of the MT was that though it could be deciphered by a human reader, it was sometimes inaccurate or completely wrong. It also criticised the work of Georgetown when compared with the 1954 experiment, stating that the output from 10 years previous were not only better, but showed little progress of the programme after that time.

Though the input for the original experiment was extremely limited and the systems tested by ALPAC were experimental, this did not lead to ALPAC cutting Georgetown any slack. ALPAC did, however, state that MT was not an issue with a foreseeable resolution as the Georgetown-IBM experiment had certainly suggested.

Though ALPAC hardly praised MT, it did appear to approve of the ideas of "machine-aided translation", which effectively refers to translation tools, which are fairly commonplace in today's translation industry. The report assessed that MT had advanced the field of linguistics more than it had the field of computing, and that MT was not deserving of more funding. Before it could receive more funding, certain criteria would have to be met.

In conclusion, ALPAC suggested the following:
  1. practical methods for evaluation of translations; 
  2. means for speeding up the human translation process;
  3. evaluation of quality and cost of various sources of translations;
  4. investigation of the utilization of translations, to guard against production of translations that are never read;
  5. study of delays in the over-all translation process, and means for eliminating them, both in journals and in individual items;
  6. evaluation of the relative speed and cost of various sorts of machine-aided translation;
  7. adaptation of existing mechanized editing and production processes in translation;
  8. the over-all translation process; and
  9. production of adequate reference works for the translator, including the adaptation of glossaries that now exist primarily for automatic dictionary look-up in machine translation
It would be fair to say that given the aim of the report, ALPAC achieved its objective of assessing MT. The downside to the report is that research into MT was effectively suspended for two decades, since all significant government funding was cut.

Perhaps we are little bitter that the ALPAC report was so damning of the work of MT merely because we can still see failings in modern day MT, such as our "favourite" Google translate. However, it would be fascinating to see what MT could have achieved had it been funded with as much fervour during the 60s, 70s, and 80s as it had been in the mid-to-late 50s.

Do you feel we would be better off had MT research continued? Or do you think "Machine-Aided Translation" was the correct avenue to pursue? Tells us your thoughts in the comments below. If you wish to read the 1964 ALPAC report, a full copy can be found here.