Interactive Vocabulary Acquisition in XTRA
Proceedings of the Tenth International Joint Conference on Artificial Intelligence. Natural Language. Pages 715-716.
© 1987 International Joint Conferences on Artificial Intelligence
Abstract
This paper describes a _j)l"actical solution to online dictionary update in X1RA, a machine transla,.. tion system developed by Xiuming Huang at the Computing Research Laboratory of New Mexico State University. The focus of the discussion in on IVES -- an Interactive Vocabulary Enrichment System built by this writer for X1RA, It reflects an ongoing effort at the laboratory to build embedded learning mechanisms in machine translation systems. Two types of learning are discussed, word lear(ling and word sense learning. Each type of learning undergoes three routine processes: detection, acquisition, and evaluation. The emphasis of this paper is on the use of semantic preference violations in the detection of the need to learn new word senses.
- Carter, D. M. "On the Fass and Wilks froposal to use 'polysemy rules'." Computatwna Lmguistics 10: 2 (1Q84) 147.
- Cohen, P. R. and E. A. Feiganbaum The handbook of artificial intell~ence. Stanford, Heuristech & Los Angels, Wilham Kaufmann, 1Q82.
- Fass, D., and Y. Wilks, "Preference Semantics, Ill- formedness, and Metaphor." Computatwnal Linguistics Q: 3-4 (1Q83) 178-187.
- HuaJ!g, X-M. "Machine Translation in the SDCG Formalism." In Proceedings o{ the Conference on Theoretical and Methodologica Issues in Machine Translatwn of Natural Languages.
- Michalski, R. S., J. G. Carbonell, and T. M. Mitchell (Eds.). Machine Learning: an Artificial Intelligence Approach Volume f!. Los Altos, Calif., Morgan Kaufmann, 1Q86.
- Ortony, A. "The Role of Similarity in Similes and Metaphors." In: Ortony, A., (Ed.) Metaphor and Thought. New York, Cambridge University Press, 1Q7Q.
- Richards
- I. A. The Philosophy of Rhetoric. New YorK, Oxford University Press, 1Q36.
- Wilks, Y. Grammar, Meaning and the Machine Analysis of Language. Routledge, London, and Boston, 1Q72.
- Wilks, Y. "Making Preference More Active.'' Artificial Intelligence 10:1 (1Q78) 1-11.
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.
IN1ERACTIVE VOCABULARY ACQUISITION IN XTRA
Cheng-ming Guo Computing Research Laboratory, Dept. 3CRL New Mexico State University Box 30001, Las Cruces, NM 88003-0001, USA ( 505) 646-5466 csNEt: guo@ nmsu
information gained during attempts to perform the task goes back to the learning element. In the con text of word and word sense learning in X1RA the environment is the system's interaction with' the user; the learning element is IVES; the knowledge base is t~e system d!ctionaries; and the performance element .'s the ~nghsh ~entence _parser. Depending on the kmd of mformatwn supphed by the environ ment to the learning element, there exist five different learning strategies. These are rote learning learning by instruction, learning by deduction learn~ in~ by analogy, and learning by induction. (Michal ski, Carbonell, and Mitchell, 1986, p. 14).
Word learning in X1RA belongs to rote learn ing .. In rote l~ar~ing the information supplied by the environment IS directly accepted by the learning sys tem. The strategy IS elementary, not powerful enough to accomplish intelligent learning on its own but it is an 'inherent and important part of any learn~ ing system' (Cohen and Feigenbaum, 1982, p. 335).
The detection of the need to learn new words in XTRA is facilitated by X1RA's preprocessing rou tine. Whenever a certain word in the input sentence is found to be missing from XTRA's dictionaries the preprocessing routine issues a warning and prinu; out the missln_K word. Whenever a word is detected as missing, IVES is called, and an interactive session with the system user begins.
. During the acquisition phas,~ information is eli Cited from the user to expand .x.1RA's dictionaries. IVES, as it stands now, handles four major categories of open system words, i.e., nouns, verbs, adjectives, and adverbs. All input into the system dictionaries is checked for its val1dity. Invalid entries are automatically rejected, and a second request for the same piece of information follows the rejection. Although the default value of the target language is Chinese, IVES could handle any target language.
When the acquisition phase is over, parsing con tinues from where it left off. Often, it succeeds as a result of dictionary expansion. However, it may also fail for reasons that will be discussed in the next section.
III WORD SENSE LEARNING
Word sense learnin& can be cate~orized as a spe cial type of learning by mstruction. It undergoes the same three processes as word learning does, i.e., detection, acquisition, and evaluation. The difference between word learning and word sense learning is that the detection process for the latter is much more complicated than that for the former. Also, in the course of the detection process,
ABSTRACT
This paper describes a _j)l"actical solution to on line dictionary update in X1RA, a machine transla,.. tion system developed by Xiuming Huang at the Computing Research Laboratory of New Mexico State University. The focus of the discussion in on IVES -- an Interactive Vocabulary Enrichment Sys tem built by this writer for X1RA, It reflects an on going effort at the laboratory to build embedded learning mechanisms in machine translation systems. Two types of learning are discussed, word lear(ling and word sense learning. Each type of learning undergoes three routine processes: detection, acquisi tion, and evaluation. The emphasis of this paper is on the use of semantic preference violations in the detection of the need to learn new word senses.
I IN1RODUCTION
Machine translation systems, built using an AI approach, could be equipped with embedded learn ing mechanisms so that the systems would con stantly update themselves and modify their behavior. Building such embedded systems should speed up the early realization of machine translar tlon on microcomputers. This is because, exce_pt for the closed system of function words, an MT pro gram cannot be expected to have a 'complete' voca,.. bulary at the time of installation. And 1t is almost impossible for the customer to purchase such sys tems that are adequate to start with. Learning mechanisms could be a remedy in this regard. It is usually easier to find a bilingual person than a software engineer to upgrade a system.
The focus of the discussion is on IVES -- an Interactive Vocabulary Enrichment System built by this writer for X1RA. Two types of learning are dis cussed, word learning and word sense learning. Each type of learning undergoes three routine processes: detection, acquisition, and evaluation. The emphasis of this paper is on the use of seman tic preference violations m the detection of the need to learn new word senses.
II WORD LEARNING
According to a model of machine learning advo cated by Cohen and Feigenbaum (1982), machine learning has four important elementsl ~a) the environment, (b) the learning element, c) the knowled)$e base, and (d) the performance e ement. The env1ronment supplies information to the learn ing element; the learnmg element uses this informar tion to make improvements in the knowledge base; the performance element uses the acquired informar tion to carry out its task; finally, feedback responses from the user are not learned by rote and used by the system. Instead, they are taken as _prag matic instructions and carried out accordingly. Word sense learning uses the same learning module as word learning to acquire new word senses. The acquired word senses are always evaluated by XTRA.
The need to learn new word senses in XTRA is detected when a semantic preference violation occurs. In semantic primitive based systems, seman tic primitives and semantic preference rules embody world knowledge, and they play an extensive role in word sense disambiguations (Wilks, 1 972; Huang, 1985). In fact, the same mechanism can oe employed in detecting the need to learn new word senses. Suppose XTRA has the word 'bridge' in the system wh1ch means 'a physical structure built on the river', and it also knows the word 'play' as in 'play basketball'. When processing the mput sen tence 'John plays bridge', the system would fail as a result of semantic preference violations. In this case, the semantic specification of the expected object of the predicate verb (play) is 'game' whereas the semantic specification of tile actual object head noun (bridge) is 'grain', which means 'any kind of struc ture'. What happens is that 'grain' . 1s not anywhere close to the semantic specifications of the expected object head noun of 'play', hence a semantic prefe~ ence violation. This particular type of semantic preference violation indicates the lack of a required word sense, and therefor; the need to learn.
Unfortunately, not all violations of semantic preference rules signal the need to learn. Only some of the prefere~ce violations are learning sign~ls. :r~e non-learning s1gnals are symptoms of some hngUJstJc complications, such as ill-formedness. In word and word sense learning ill-formedness often involv~s misspelling and the figurative use of language. It IS not surprising that part of the game of word and word sense learning is the recognition and handling of metaphors and misspellings. IVES will not request new vocabulary information until the last recorded semantic preference violation is found unaccounted for, i.e. when no metaphors or misspellings are identified in any pair of recorded word senses involved in semantic preference viola tions.
I
A. Recognition and Handling of Metaphors
Many solutions have been suggested to handle ill-formedness incurred 1:>y metapho:s. Four stra tegies -were proposed by Fass and W1lks (1983) for semantic primit1ve based systems. They are the pas sive relaxation strategy, the Change The Data ( CTD) strategy, the Change The Expectations ( CTEJ strategy, and the active strategy that involves the use of pseudo-texts (Wilks, 1978) . The strategy adopted by XTRA is to relax the preference of the predicate and accept the semantic representation with the conflict unresolved. Relaxation is a useful heuristic to handle metaphors, but not without flaws. To illustrate the unpredictable nature of the relaxa tion strategy, look at the following sentence:
The chopper drank gasoline.
'Chopper' can either be a 'helicopter' or an 'ax'. In this particular sentence, more than likely it means a 'helicopter'. What will happen if the system dic tionary has (a) the 'helicopter' sense of choRper only, (b) the 'ax' sense only, or (c) both the hel icopter' and the 'ax' senses? Since XTRA relaxes the preference of 'drank' for an animate subject in all three cases, we obtain the desired sense of chopper under (a); the less than desired sense of chopper under Cb); and unpredictable results under (c). In other words, 'relaxation' can help produce the correct parse sometimes, but may not do so at all times.
As Carter (1984) pointed out, to recognise metaphors 'recourse to a richer source of knowledge like Wilks' pseudo-text is necessary.' A pseudo-text is a preference semantic representation of factual and functional information about a concept. The form of representation used in IVES incorporates the pseudo-text idea in a form that integrates linguistic knowledge with general world knowledge in one unit of representation. It is called the Integrated Seman tic Unit (ISU). Each ISU is an integrated representa tion of the meaning of a word sense. The building blocks of an ISU are also word senses. The general form of an ISU is as follows:
isu(Wordsense, belong( [Oassj), ) .ik( integrated_knowledge)
where 'isu' stands for integrated semantic unit; 'Wordsense' is a word sense of any entry word in Longman Dictionary of Contemporary English (LDOCE); 'belong' introduces a hierarchical rela tionship between 'Wordsense' and 'Class'; 'Class' represents a superordinate word sense in the hierar chy of word senses in LDOCE; 'ik' introduces integrated linguistic and world knowledge associated with Wordsense. Although the general form of an ISU is uniform across all four open-class categories of English words, i.e., nouns, verbs, adjectives and adverbs, the actual specification of their respective superordinate word senses, and integrated knowled~e varies from category to category. An example 1s given below:
isu( chopper!, belorig( [tooll]), ik(parf{[hanCilel, blade!]) property( [[cut1 1 meatl]]j, part_of_s~h~[n]) root_forni( [chopperf.), inflection8t_form( [choppers]), feature( [I), co_exishn~enses( [3]), sem_head t ing]), ). r_phrase( )
IVES uses ISUs to recognize metaphors. In fact the recognition process be~ins with the recording of all semantic preference v10lations detected during parsing. Some of these violations represent a search mg process and are eliminated. Only interesting semantic clashes remain. IVES looks into each one of these clashes for indications of metaphorical use and misspelling before it requests to acquire any new word senses. Among the semantic clashes that IVES can handle are subject/verb mismatches, verb/object mismatches, adjective/noun mismatches, and sub.]ectfpredicative adjective mismatches.
To recognize metaphorsJ IVES first tries to iden tify three important parts ol a metaphor, the tenor, the vehicle, (Richards, 1Q36) and the salence (Ortony, 1Q7Q). Ortony takes salence' to mean an estimation of 'prominence of a particular attribute with respect to a concept to which it does or could apply.' (p. 162.) In 'The chopper drank gasoline', the tenor is man, a human-being who can drink liquid; the vehicle is 'chopper' the actual subject of the sen tence; the salence is found in the resemblance between a man's act of drinking and whatever a chopper can do. Computationally, the expected sub .ject of 'drank' is taken as the tenor; the actual sub JeCt 'chopper' is taken as the vehicle; and 'drank', the main verb of the sentence, is used to find the salence. IVES examines the properties associated with both the tenor and the vehicle to see if their properties are both related to the act of drinking. Once a metaphor is found, IVES would check with the user for confirmation. To handle metaphors, the CTE strategy as mentioned earlier is used. If no metaphors are spotted, IVES would start looking for spelling errors. The recognition and handling of misspellings are discussed in the next section.
B. Recognition and Handling of Misspellings
IVES recognizes two types of spelling errors in collaboration with the user. They are context independent errors and context-dependent errors. Context-independent errors can be spotted without examining the context where the error occurs. These spelling errors typically involve missing, surplus, or wrong letters, or letters with their proper ~s1tions switched. 'convenence', 'conveemence' conviniece', and 'conveneince' for 'convenience! would be examples in point. These errors would be picked up by XTRA's preprocessing routine simply because the misspelled forms are not found in the system dictionaries. The system would treat them as missing words and call the learning module to input the required information. To recognize and handle context-independent errors, the learning module checks on the spelling of any word that calls for new vocabulary information. The user is requested to do the actual checking, and is given a chance to correct him/herself when a spelling error is spotted. How ever1 it so happens that sometimes the misspelled wor<l takes the form of the correct spelling of another word which is already available in XTRA's dictionaries. We call this second type of misspellings context-dependent errors. Their recognition requires tl:ie examination of the context where the error occurs. More often than not, the spelling error would cause a semantic preference violation, and IVES would undergo the same processes as described in subsection A above to process the semantic mismatch. When IVES fails to identify any meta phors, it will start looking for spelling errors. Again the user is requested to do the actual checking, and is given a chance to correct him/herself when a spel ling error is spotted. Two research questions remain open and unanswered at this stage:
- What if the user is locked in an undesireable mind set and fails to see the spelling error when he or she is given a chance to?
- What if a context-dependent spelling error does not cause any semantic preference violations?
IV CONCLUSION
On-line vocabulary acquisition is not a recent phenomena either for machme translation in particu lar or. for na~ural }anguage . processing in general. What Js new m th1s paper JS the use of semantic primitives and semantic preference violations in the detection of the need to learn new word senses. The mechanism originally designed to disambiguate mul tiple word senses in an MT system is taken advan tage of and made into part of a comflex of mechan isms intended to detect the lack o required word senses and therefore, the need to learn. To cope with the inescapable problem of metaphor recogni tion in the process of detecting the need to learn new word senses, a new form of representation, ISU that integrates lin~uistic knowledge with ~eneral world knowledge 1s introduced. The motivation behind such introduction stems from the following observation:
Knowledge-based systems typically ignore the rich linguistic knowledge found in conventional dic tionaries while conventional dictionaries never express the rich commonsense knowledge assumed of the reader explicitly. Yet, linguistic knowledge and general world knowledge are not ultimately separable. What natural language processing really needs is an integrated form of representation that com hines the two kinds of knowledge in one unit of representation (Wilks, 1Q78). The present research represents an effort in the application of the integrated approach to the area of dictionary update in machine translation.
ACKNOWLED CEMENTS
My deep appreciation is due to all members of our natural language processing group 1 particularly to Xiuming Huang, Dan Fass, and Yorick Wilks, for their encouragement, advice and help.
@inproceedings{ijcai1987p142,
title = {Interactive Vocabulary Acquisition in XTRA},
author = {Guo, Cheng-ming},
booktitle = {Proceedings of the Tenth International Joint Conference on
Artificial Intelligence, {IJCAI-87}},
publisher = {International Joint Conferences on Artificial Intelligence},
editor = {John McDermott},
pages = {715--716},
year = {1987},
note = {Natural Language},
url = {https://ijcai.org/proceedings/1987/142},
}
TY - CPAPER TI - Interactive Vocabulary Acquisition in XTRA AU - Guo, Cheng-ming T2 - Proceedings of the Tenth International Joint Conference on Artificial Intelligence, IJCAI-87 PB - International Joint Conferences on Artificial Intelligence PY - 1987 SP - 715 EP - 716 UR - https://ijcai.org/proceedings/1987/142 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