ATR Interpreting Telecommunications Research Labs
Studying the Characteristics of Conversational Language,
and Embodying these in an Automatic Interpreting Device
Implementation of Transfer-Driven Machine Translation
Satoshi Shirai
Conversations between two human beings frequently contain grammatical errors and
fragmented expressions. We conducted research into technologies that can translate
this type of informal speech. Specifically, we have achieved Transfer-Driven Machine
Translation (TDMT), which creates idiomatic translations using expression patterns
comprised of several words as the translation unit. We have created a number of
technologies to increase the system's applicability in conversational translations,
including filling in the clausal subject and partial-translation methods. We have
gathered and organized expression patterns based on examples of model dialogs
from travel conversations, and have proven the viability of multilingual informal
speech translation.
1. Transfer-Driven Machine Translation (TDMT)
We have created a translation method that facilitates responses even when the
expression is grammatically incorrect, by gathering actual conversations and the
results of simultaneous interpretation, and implementing translations based on
an imitation of the corresponding interpretations. We conducted initial investigations
to generalize the distinctive expressions found in example translations, thus
increasing the coverage of expressions that the system can translate. For example,
"Chugoku no kawa" can be generalized into the pattern "X no Y" [Y of, or belonging
to, X]. To prevent the original meaning from being lost as a result of generalization,
the original expression "Chugoku no kawa" and the translation "river in China"
are registered along with the pattern. When a sentence is input into the translation
system, the system searches for patterns that correspond to the expression, selects
the expression that is closest in meaning to the input expression from among the
registered patterns, and creates a translation imitating the selected expression.
If the sentence "Kyoto ni tochaku suru jikan wo oshiete kudasai" [Please tell
me the time I will arrive in Kyoto.] is input, the system will create the following
translation. The first target is the pattern "X ni Y" in "Kyoto ni tochaku suru".
For this pattern, the following expressions might be registered: "Osaka ni tsuku"
[arrive in Osaka]; "tomodachi ni au" [meet a friend]; or "hidoi me ni au" [have
a terrible experience]. The expression "Osaka ni tsuku" is selected as being the
closest in meaning to "Kyoto ni tochaku suru" according to the theasaurus. The
system then creates the expression "arrive in Kyoto", imitating the translation
"arrive in Osaka". It then creates the following phrases in order: "the time when"
for "tochaku suru jikan"; "tell (me) the time" for "jikan wo oshiete"; and "please
tell me" for "oshiete kudasai". These phrases are combined in accordance with
grammatical rules to derive the final translation, "Please tell (me) the time
when (I) will arrive in Kyoto".
Using TDMT, and by adding patterns based on translation examples, the performance
of the system can be Increased with relative ease, although at times it will be
necessary to adjust pattern correlations.
2. Automatic Insertion of Various Omitted Elements
In the above noted translation, "Please tell (me) the time when (I) will arrive
in Kyoto", the elements Iin parentheses are not expressed in the Japanese it is
common not to express these elements, but they are essential in the English equivalents.
If we compare the Japanese and English expressions in the example translation,
we can see what elements need to be supplied. To accommodate this, we created
a decision tree to suggest elements that were omitted from the translated sentence,
and used this to create a method of suggesting the required elements. According
to evaluation experiments, Iit Iis possible to correctly fill in the sentence
subject in over 80% of cases where the sentence subject was unknown, using as
clues such factors as the speaker, politeness level, and the verb governing the
missing element.
We are currently making considerable progress with this technology, and are now
considering contextual processing methods that can suggest, for example, what
Iis being indicated by the Japanese phrase "sono hoteru" [that hotel].
3. Partial Translations
Because it is still difficult for speech recognition technologies to correctly
recover all elements of a conversation, recognition results often contain errors.
The utterance "hoteru wo yoyaku shitai'n dakedo", if correctly recognized, can
be translated as "I'd like to reserve the hotel". But if the word "hoteru" [hotel]
is incorrectly recognized as "hotaru" [firefly], and if this is translated without
being corrected, the result will be incomprehensible, even if the translation
process is correct. In this case, if there is an assumption that the situation
is related to "travel conversations", then the system can make a judgment that
the word "firefly" is unnatural. It will then remove the unnatural element, translate
"yoyaku shitai'n dakedo", and create the phrase "I'd like to reserve...". In a
conversation, one can take measures such as asking for a repetition, so it is
possible to continue the conversation using this partial translation. With regard
to the segment where the unnatural expression appeared, we are also examining
methods of automatically correcting the error by gathering dialogue examples in
advance, and using these to surmise the type of utterance that would normally
be used in this situation.
4. Toward Multilingual Translation
We have applied TDMT to translations from Japanese to English, English to Japanese,
Japanese to Korean, Korean to Japanese, Japanese to German, and Japanese to Chinese,
and have confirmed the system's effectiveness in each case. The translation targets
are limited to travel conversations, but the proportion of crucial information
translated was over 90% in the case of Japanese to Korean and Korean to Japanese,
and over 85% in the case of Japanese to English, English to Japanese, and Japanese
to German. In the future, it will be necessary to accumulate more translation
examples, and at the same time to promote increases in translation quality and
in the efficiency of initial investigations supporting pattern creation.
