For a Human/agent and agent/agent communication

Classification of Message Levels and

the Use of Machine Translation Technique

for the Flexible Communication

between Agents

Key-Sun Choi, Jun-Sik Park, and Namil Kim, Woon-Jae Lee, Seongyong Kim,

Korea Advanced Institute of Science and Technology

Center for Artificial Intelligence Research

Department of Computer Science

kschoi@cs.kaist.ac.kr

1. Abstract

In these days, reasoning and grasping the user's intention in the view of user interface are the core of agent technologies. Natural language processing technology is important because it enables intelligent understanding of the user's intention and the user's convenience. But agents make it difficult to use a uniform language. Therefore we propose a concept of interface-level of message between agents. In addition, we underscore the importance of machine translation technology which helps a user handle exceptional situation and propose FMT(Flexible Machine Translation) as a machine translation technology.

1. Introduction

Agent technology will need amounts of supplements of long duration because there is no agent system which works fully or is verified by the end-user. When a user deals with exceptional situations, he can not use natural language. When a user contacts a foreign agent directly, he must communicate with the agent in foreign language. However, FIPA specification regards the national adaptation as an enabling technology and specifies roughly.

But, assume the followings in the real world.

• Agent technology is not perfect. (Some malfunction may arise.)
• Agent power does not always follow the situation.
• Agent interface is local to its country.

According to the above assumptions , it may occurs direct contact between a user and an agent. Therefore the user needs help, especially for language problems. Our proposal considers those problems and proposes a concept of message levels as a solution and Flexible Machine Translation as a message translation technology.

1. Message Level

Though the interfaces for user/agent and agent/agent are specified, there are problems in the relationship between the interfaces. The interface for agent/agent does not require complex language format and a user will not use agent-oriented expression, abandoning the expressive power of natural language. For this problem, we propose the concept of interface level.

The interface levels consist of four classes.

• Primitive(reserved) : message on this level is used if in the case of worst situation or the simplest message form is required. The simplest words are used. For example, a user is informed of the cases of ERROR, WARNING or EMERGENCY with a dialogue box interface.(level 1)

• Intermediate(interactive): In most cases, a fixed dialogue or a situation are repeated. It can be thought as the word filling for each variable portions in a template. The following example shows this case.(level 2)

Variables are in ().

• Constrained(controlled): Restricted natural language can be used in this level. It is processed efficiently because of restriction on vocabulary and syntax. It gives users flexibility also.(level 3)

• Human - oriented(Full natural language): It is general natural language and used for communication between human and agent .(level 4)

Level 1 and 2 are sufficient for agent/agent communication. If human is involved, level 3 and 4 is desirable although level 2 can be used. Of course, the higher the level is, the more expensive the processing cost is. Figure 1 shows levels.

Figure 1 interface level and its cost

Level 1 and 2 do not require amount of linguistic processing. The literal word-to-word concept solves most of them. But level 3 and 4 require deep level of the natural language processing concepts.

Low level expression is converted to higher level expression when an agent reports information to a user or a user wants to know the contents of communication between agents. In this case, ambiguity is low and process of expression is fast.

In the opposite case, a user communicates with an agent, higher level expression is converted to lower level expression (for offering information to the agent). There may be some problems caused by ambiguities related to the user or user's sentences.

It's a trade-off problem between knowledge and expressive power.

It should be left as system implementation problem whether selection of level is automatic or manual.

Agent can not afford the all functions for level conversion, therefore there should exist an agent or message broker [3] which intercepts , manages for other agents.

1. Machine Translation

In multilingual environment, necessity of Machine Translation occasionally arises. Especially using agents in other countries or getting along with foreign languages will require Machine Translation Technologies. We will think the notion of MT as one general case of exceptional cases.

For example, a certain user - even if he/she is familiar with English - wants to order some goods. The user has few idea of it the first time. It is necessary to get the help of agents. The agent called will do good jobs and negotiate with other agents. But the information in the real world is very changeable and large, so agents cannot keep up with them. Even a human cannot do completely. In this case agent will give the message such as "I cannot handle it, I give up..." to its user and will be terminated.

In this situation user can negotiate with other agents or person directly. Very critical problems should not be up to the agents. Users will be exposed to the strangeness of foreign language and custom.

To solve the problem, it is necessary to use Machine Translation. We will explain the concept of MT, and then FMT which is appropriate to the Interface Level.

Machine translation systems produce translated output without any human intervention. They need access to a range of programs in order to analyze and interpret the input, look-up terms in a dictionary or terminology database and generate the target language output.

The linguistic modules in machine translation systems are responsible for the analysis of the source text, transfer between two languages and the generation or synthesis of the target text. The analysis produces a complete parsing of a source-language sentence whereby all the words and lexical items in a sentence are reduced to their basic grammatical components. The output of the analysis stage is used to create the translation in the target language.

In the analysis and generation stages, most systems have clearly separated components for dealing with different levels of linguistic description; morphology, syntax and semantics. Analysis is therefore divided into morphologic analysis (identification of word endings), syntactic analysis (identification of sentence structure and parsing) and semantic analysis (resolution of lexical and structural ambiguities). These phases may also be applied to the generation stage.

There are 3 approaches to machine translation; direct translation, interlingua, transfer. Direct translation systems were 'word-for-word' translation systems, perhaps with some local word-order adjustment. Direct systems were limited to the minimum work necessary to do the translation for a single language pair. Direct systems are linguistically very weak, offering only word-for-word translation resulting in poor quality output.

In interlingual systems, source-text analysis and target-text generation are kept separate; conversion from one language to another is achieved via abstract 'interlingua', representations of meaning that are common to several languages. Translation therefore has two stages; analyzing the source language to an interlingual representation and then generating the target language. Interlingual systems are considered theoretically superior to other approaches, since they require fewer language-dependent modules. However, the reduction in modules is generally negated by the effort required and difficulty involved in defining an interlingual representation for all languages.


Figure 2 Transfer Machine Translation

The third approach to machine translation is 'transfer'. There are three stages involving underlying syntactic representations for both the source language and target language texts. The first stage converts the source language into 'deep' intermediate representations, in which ambiguities have been resolved without reference to the target language. In the second stage, these are converted into equivalent 'deep' representations in the target language. The final target language text is then generated. In transfer systems, analysis and generation programs are independent and are specific for particular languages. There may also be separate components to handle lexical transfer (selection of vocabulary equivalents) and structural transfer (transformation of source-text structures to equivalent structures in the target text). The transfer approach is the one most commonly used in current commercial systems, as it can be easier to implement than interlingual system.

1. Flexible Machine Translation

Now we propose Flexible Machine Translation(FMT) architecture for the machine translation systems used in agent technology.

In fact, the process of translation does not assume the understanding. In some occasion, a set of translation templates is enough to translate. For example, articles in stock news use only the special usage and special domain expression. That does not invoke the process to recover the complete document. If such process fails, then the next process in the deeper level of document architecture starts to analyze and translate. That process fills up the next level of document architecture, their result is transferred to the reader's language side by using the appropriate transfer knowledge. The sequence of processes to recover the complete document is awaken incrementally in a demand-based way. That is flexible in a sense that the process evocation is flexible. We call this "Flexible Machine Translation"(FMT).

Figure 3 Flexible Machine Translation : Configuration

As shown in the above figure, after failure from the first process "morphology analyzer", that process either suggests the alternative solution or passes to the next process "syntactic analyzer". The result of syntactic analyzer is transferred by based on syntactic pattern transfer knowledge. The evaluator of the corresponding node of syntactic generator in the reader's side measures whether the output of syntactic transfer is possible to be generated upto the final surface form. These processes continue whenever the reader's side sends the rejection signal. The feedback information is also logged. The process will progress until the interlingua meets. However, if the reader's side accepts the transferred document, the translation ends without going to the direction toward the point of interlingua.

Flexible machine translation is embodied in a "distributed" way. Every module can be a node in network. They are fault-tolerant because each level is competing nodes of the so-claimed same function. Every node is also competing and compensating processes or entities.

1. Use of FMT to Interface Level

As a solution of the previous problem, after adapting interface level, translation can be efficient by using the concept of flexible MT. In the case of level 1 and 2, it is possible to solve by using word-to-word translation (morphological analysis), and in level 3 syntactic processing will be needed. In level 4, it is a problem of general Machine Translation.
Figure 4 Adaptation of FMT to Interface level

In the higher level, expressive power increases but the problem of processing time and correctness will also arise. But these have related to MT problem itself which will be improved.

For language support the ideal form is the full natural language support.

1. Conformance to the FIPA CFP

Above technologies can describe the following criteria

• Messages can be understandable
• extensive set of message types
• formal, computational framework for representing protocols
• allow for general failure to meet expectations :

In addition, we propose

• the necessity of message level,
• and the importance of Natural Language Processing

1. Application Model

Consider the following Example

Figure 5 Example

1. message level
   • User "a" in the country "K" wants to get the information of country "I" and reserves if necessary. "a" meets the agent in the "K". This agent meets other agents in the "I" and do the necessary jobs. In this case, message level between user and agent is level 4, and the agent knows the intention of "a". It communicate other agents in "I" at the message level 2. The result at the level 2 will be converted to the report describe at level 3. The report will be shown to the "a".

1. FMT example
   • In the previous example, if a part of report is described in other than native language, it may be possible to translate it.
   • When user "c" from "K" who travels in the country "I" asks agents in "I", and adapting message level and the translation of the message may help "c".

1. Conclusion and Future Extension

So far we have said that the necessity of not only simple format protocol, but also flexible type of message according to the environment, and that translation technology must aid the user if languages differ. Discriminating the type of message is necessary to improve the efficiency of information processing between agent and person. For the human interface in the agent technology, natural language must be the main goal and natural language processing is very important. In other words, Natural Language Processing Unit should be treated as important issues. To translate in the different language environment means that the Agent System is able to help the user in the exceptional case. So, aids by agent-management system in the exceptional case should be exist. In the view of the total system, user should be aided by agents even if most of agents around cannot help him/her, and system is responsible for it.

1. Reference

[1] Association for Foundation for Intelligent Physical Agents, FIPA's First Call for Proposals, Tokyo, Japan, October 1996

[2] Key-Sun CHOI, Agents in information Architecture for Intelligent Distributed Multilingual Document Retrieval Service, Fipa 3^rd Meeting, Tokyo, Japan, October 1996

[3] Key-Sun CHOI, S.Y KIM, W.J LEE, J.S PARK, and N.I KIM, Structured Knowledge Description Language to Represent Objects and Knowledge of an Agent, Fipa 4^th Meeting, Turin, Italy, January 1997

[4] Jane Mason and Adriane Rinsche, Translation Technology Products, Ovm Ltd, 1995