High Level Memory Structures and Text Coherence in Translation
Proceedings of the Seventh International Joint Conference on Artificial Intelligence. Natural Language. Pages 47-47.
© 1981 International Joint Conferences on Artificial Intelligence
Abstract
Various memory organization schemes have been proposed in the last five years. Lots of intelligent computer systems have been experimenting with memory schemes like script a, plana, goals, [4] and HOPs [5] in the domain of text understanding [3] Ill and information retrieval [2]. In this paper, the focus is on the problem of translating sentences that involve lexical items which don't have equivalent counterparts in the target language. Examples are drawn from translations between English and Mandarin Chinese.
- REFERENCES Dyer, Michael G. and Lehnert, Wendy G. Organization ~ ~ Processes !QL Narratives. Technical Report 175, Yale University. Department of Computer Science, 1980.
- Kolodner, J. Retrieval ~ Orgapizatiopal Strategies in Conceptual ~: A Computer ~• Technical Report 187, Yale University. Department of Computer Science, 1980.
- Lebowitz, H. GeperaliJation ~ ~ ~A£ Iptegrated Understapdipg ~• Technical Report 186, Yale University. Department of Computer Science, 1980.
- Schank, Roger and Abelson, Robert. Scripts, ~. ~. ~ Upderstanding.
- Schank, R. Reminding ~~Organization: An Introducton 1Q MOPs. Technical Report 170, Yale University. Department of Computer Science, 1980.
References were extracted automatically from the PDF and may contain errors.
Full text extracted automatically from the PDF; figures, tables and equations may differ. The PDF is the version of record.
High Level Memory Structures and Text Coherence in translation
CJ Yang
Department of Computer Science Yale University New Haven, CT
Various memory organization schemes have been proposed in the last five years. Lots of intelligent computer systems have been experimenting with memory schemes like script a, plana, goals, [4] and HOPs [5] in the domain of text understanding [3] Ill and information retrieval [2]. In this paper, the focus is on the problem of translating sentences that involve lexical items which don't have equivalent counterparts in the target language. Examples are drawn from translations between English and
Mandarin Chinese.
Consider the following sentence: 1
- la John went to a restaurant and ordered a hamburger.
- lb JOHN CHIU' YI GIA TSAN TIEN DIAN< LEH
- lc John go one restaurar.t order
YI GEH HAMBURGER.
one hamburger.
Notice that "and" in la does not appear in lc. However, in 2a below:
- 2a John read a book and wrote a letter.
- 2b JOHN DWU- LEH YI ~EN< SRWU BEB-
- 2c John read one book and
SHIYE< LEH YI FONG SHIEN'.
write one letter.
the word "and" IS translated (i.e. What then determines when "and" as "HEH-"). should be translated in Chinese? Although both la and 2a consist of two events, the two events in la are causally related and the ones in 2a are not. In la, each verbal phrase corresponds to an event in the restaurant script [4]. Recall a script is actually a giant causal chain. So, if two events are part of the same sc·ript, their causal connections can easily be inferred. The two events in 2a, however, have no such connections at all.
1 The following conventions will be used throughout this paper:
a refers to the English sentence.
b refers to the Chinese sentence
(in upper case).
c refers to the word-for-word
translation of b.
Nouns like "John" and "hamburger" are not translated since their translations are phonetical.
So our claim is:
- IF the causal events in s inferable, connection Chinese between two sentence is
THEN the two verbal phrases representing these two events are connected without the use of HEB-. (In fact, its use is forbidden)
Verbal phrases that satisfy the above condition are said to be conjoinable.
Without the knowledge at the script level, both la and 2a will just be understood as 2 separate events. So they should be translated exactly the same way. If this is the case, then it is impossible to determine why "and" is translated differently in lb and 2b. Therefore, understanding la at the script level is crucial to its correct translation.
However, we are not saying that all events in a script can be expressed in a sentence without using connectives. Consider the next exantple:
- 3a John went to a restaurant where he ate a hamburger and drank a cup of tea.
- 3b JOHN CHIU' TSAN TIEN TSRE LEH YI GEH
- 3c John go restaurant eat one
HAMBURGER HEH- HEH LEH YI BEl TSRA-.
hamburger and drink one tea.
Although eating a hamburger and drinking a cup of tea are events in a restaurant script, they are not causally related. So the connective "BEB-" must be used in 3a. The conclusion we should draw here is that scripts together with causal chains provide a definite criterion for conjoinability of verbal phrases in Chinese.
Now let's consider the following more complex
story:
- John went to a restaurant for lunch.
- He left without paying the check.
The translation of the first sentence is quite straightforward and irrelevant to our discussion here. Let's focus on the translation of the second sentence:
4a. He left without paying the check.
4b. TA MEl- FWU' CHIAN- GIO' DZO< LEH.
4c. He not pay money 1111 leave.
According to the dictionary, "GIO'" means "at that moment". But actually it means more than that. We understand that John shouldn·t leave the restaurant without paying because we have the knowledge about restaurants (i.e. the restaurant script). 4a doesn't explicitly express that a violation of the restaurant knowledge has occured. The Chinese translation of this story, however, uses "GIO'" to indicate that "leaving the restaurant" should not occur before "not paying the check". Although "GIO'" can·t be directly translated from English, this doesn· t mean "GIO"" should never appear in a Chinese translation of English.
Let's see how "GIO"" can be generated using the knowledge at the script level. First, 4a is parsed into two COs (Conceptual Dependency r epresentations [4)):
- 1 (PTRANS (ACTOR John) (OBJECT John) (FROM restaurant))
- 2 (~TRANS (ACTOR John)
(OBJECT money)
(FROM John)
(MODE neg))
With the knowledge of the restaurant script, the generator realizes that 1 followed by 2 is not a normal instantiation of the script. So "GIO"" must be generated to indicate this violation.
Usually, words are generated from some specific nodes in the memory or slot fillers. For example, DZO< (means "to leave") is generated from "PTRANS from current place", FWU' (means "to pay") is generated from "ATRANS money", and so on. However, we can't find such memory nodes or slot fillers that correspond to words like "GIO"". It is used to characterize certain states or aspects of the memory structures. In other words, certain patterns of memory structure can also trigger generation besides static memory nodes. ''GIO "" should be generated for any memory structures that satisfy the condition we discussed above. Failure to do so will result in serious misunderstanding. For example, if "GIO'" is left out in 4b, then 4b becomes 4b' as shown below:
4b' TA MEl- FWU' CHIAN- ZO< LEH. 4c' He not pay money leave.
Native Chinese speakers will think that 4b' does not make sense. We arP going to show.· why they will think so by the conjoinability criterion discussed earlier. In 4b, the two verbal phrases representing "not paying ·the check" and "leaving the restaurant" are not connected by any connectives. This implies that these two events are causally related according to the conjoinability criterion. Thus, 4b' should be interpreted as "John didn't pay the check so he left the restaurant" which obviously violates our knowledge about restaurants. Therefore, 4b' won't sound right to Chinese speakers. The point is that without the knowledge of the restaurant script, 4b' will be generated to as the translation of 4a.
As a concluding remark, we want to emphasize that high level memory structures are responsible for maintaining text coherence in translation. Our discussion has just scratched the surface. There are msny other kinds of memory structures besides scripts. Currently, we are working on how different high level memory structures are related to the generation of coherent text.
ACKNOWLEDGEMENTS
I thank Hike Dyer for his comments on the first draft of this paper.
@inproceedings{ijcai1981p10,
title = {High Level Memory Structures and Text Coherence in Translation},
author = {Yang, C. J.},
booktitle = {Proceedings of the Seventh International Joint Conference on
Artificial Intelligence, {IJCAI-81}},
publisher = {International Joint Conferences on Artificial Intelligence},
editor = {Patrick J. Hayes},
pages = {47--47},
year = {1981},
month = {8},
note = {Natural Language},
url = {https://ijcai.org/proceedings/1981/10},
}
TY - CPAPER TI - High Level Memory Structures and Text Coherence in Translation AU - Yang, C. J. T2 - Proceedings of the Seventh International Joint Conference on Artificial Intelligence, IJCAI-81 PB - International Joint Conferences on Artificial Intelligence PY - 1981 SP - 47 EP - 47 UR - https://ijcai.org/proceedings/1981/10 ER -Copied
Machine-readable versions of this paper, for tools and AI agents:
Welcome
Future Conferences
AI Journal Division
IJCAI Proceedings
About IJCAI
IJCAI Awards