Report on the JSPS Award for Eminent Scientists FY2013

Robert Kowalski, Professor Emeritus, Imperial College, London, UK

JURISIN October 27-28.

I attended and participated in the discussions of the workshop. Most of the talks were related to my research background and interests. In particular, I contributed to the discussion about the following talks:

A Fine-grained Approach to Statutory Interpretation in Legal Decision Aids by Padmaja Sasidharan and Dr Jeroen Keppens.
Legal statutes can be regarded as having a pyramid (or triangular) structure whose top represents a goal concept, and whose bottom is a collection of vague (or open-textured) concepts. This paper presented a novel analysis of different kinds of vague concepts. Deciding how to interpret these vague concepts can be done by using case-based reasoning or by reasoning about the values promoted by a decision in a particular case. The paper gave rise to a more general discussion concerning case-based reasoning and its relationship with value-based reasoning. We held many further discussions about this topic during Ms Sasidharan’s subsequent stay at NII.

Invited Talk, Legal Communication without Shared Knowledge or Framework: An Approach from Forensic Psychology by Kotaro Takagi.
Professor Takagi presented his work on the interaction between lay and professional judges in legal cases, identifying deficiencies in existing systems, as well as proposals for reducing such deficiencies. I observed that similar problems arise in other group decision-making contexts, such as those in which members of a committee judge research grant applications. In my experience, sitting on such committees, it is often the case that individual members of the committee will state opinions, often very strongly, without presenting any argument to support their opinion. Often a decision will be made without a sufficient summary of the arguments for the decision. I suggested that the proposals that Professor Takagi presented could also be useful in such other kinds of group decision-making contexts.

Invited Talk, Modal patterns in abstract argumentation by Davide Grossi.
Dr. Grossi showed that abstract argumentation graphs can be interpreted as graphical representations of possible world accessibility relationships in modal logic. This means that certain properties of arguments can then be expressed in modal logic terms. This is an interesting mathematical observation. However, it is not clear whether it has any practical benefits.

Answering Yes/No Questions in Legal Bar Exams by Mi-Young Kim, Ying Xu, Randy Goebel and Ken Satoh.
This paper showed how logical representations can be extracted from legal texts and used to answer yes-no questions in natural language. This work contributes to the more general problem of extracting knowledge in logical form from large bodies of text, which are now available on the web. I mentioned some of my own experiences in translating English texts into logical form. This topic became one of the main themes that we discussed further, especially during my visit to Kyoto with Yamamoto-sensei and Kurohashi-sensei.

Computing Preference-based Argumentation in Answer Set Programming by Toshiko Wakaki and Masahiro Tatsuzawa.
This talk showed how argumentation can be implemented in ASP. In later discussions with Satoh-sensei, he showed me how a similar approach can also be used to implement in ASP the abductive logic programming agent framework in my 2011 book. This opened up an avenue of research which we plan to continue to explore in the future.

Belief Re-revision in Chivalry Case by Pimolluck Jirakunkanok, Shinya Hirose, Katsuhiko Sano and Satoshi Tojo.
This paper showed how modal logic can be used to represent communications and belief change among agents. The approach uses possible world semantics to represent an agent’s state of belief. I suggested an alternative approach, using sets of sentences to represent an agent’s beliefs. During my later visit to Jaist, we discussed with Tojo-sensei the relationships between the different approaches.

Invited Talk, A Brief Overview of Formal Argumentation Theory by Martin Caminada.
This talk was very closely related to my own research in argumentation. It was interesting to see how Caminada’s approach compares with my work with Toni and Dung. I had many comments, but mostly about technical details.

Invited Talk (hosted by AAA2013): Argumentation in Assurance Cases - Themes and Current Directions.
Tim Kelly I was not aware of this area of research before the talk. It was interesting both to see how close it is to my own work using logic programming for goal-reduction, and to see how successful this work has been in its own field. I discussed the relationships between our work, and plan to look at these relationships more closely in the future.

Special Lecture, Optimal and Conservative Social Laws for Multi-agent Systems by Thomas Agnotes. This was another talk about applying modal logic to multi-agent systems. I argued that modal logic is too weak for such purposes.

ArgPROLEG: A Normative Framework for The JUF Theory by Zohreh Shams, Marina De Vos and Ken Satoh.
This paper concerned the relationship between norms, arguments, and rules and exceptions. Norms share many characteristics with personal goals, and it seems that similar criteria can be used to decide how to resolve conflicts between goals and norms. Argumentation is one way of making these decisions. I suggested that decision theory, with its treatment of utilities and probabilities, is an alternative approach, which may be more appropriate in many situations. I presented a number of examples to support my argument against argumentation, and we had many fruitful discussions about this topic during Ms Shams stay at NII.

Hokudai, 22-25 October.

I gave six one hour lectures and ran a four hour workshop on Logic and English.

Lecture 1: The clausal form of logic as the language of thought (called LOT).

The discussion focused on the relationship between natural languages, such as English and Japanese, and the language, if any, of human thought. The main controversy is whether the LOT is independent of the natural languages that we speak, or whether the languages that we speak influence the way we think. Because of the differences between the English and Japanese languages, this proved to be a popular and lively topic of discussion.

Lecture 2: Psychological studies of human reasoning and thinking.

A number of psychological studies of human thinking support the argument that human thinking does not have a logical character. I presented a number of psychological tests to the audience, and asked for their responses. Some of these responses were consistent with the results of psychological studies, whereas others were consistent with classical logic. This gave rise to a discussion about whether or not the psychological tests were correctly formulated, or whether in some cases they might have been poorly formulated.

Lecture 3: Logic and legal reasoning.

I presented a number of examples, illustrating the relationship between the syntax of computational logic and the structure of legislation written in English. I also argued that computational logic can help to improve the style of poorly written English. These examples led to an interesting discussion about the logical structure of Japanese law. Professor Ken Satoh, who has investigated the logic of Japanese law was in the audience and contributed to a lively discussion.

Lecture 4: The logic of production systems.

Production systems are widely regarded as the most convincing computational model of human thinking. However, they do not have a straight-forward logical semantics. I presented the production system model and proposed a logical formulation of production systems. The audience was not very familiar with production systems, so it may not have been very obvious that it is useful to give them a logical semantics and logic-based proof procedures.

Lecture 5 Abduction.

Abduction is the problem of generating and deciding between different candidate sets of hypotheses to explain observations. Abductive Logic Programming (ALP) generates these hypotheses in a top-down manner that guarantees that the hypotheses are relevant to the observations to be explained. I also discussed a number of other criteria that can help to decide between different candidate sets of hypotheses. This led to an interesting discussion about the relationship between top-down and bottom-up reasoning.

Lecture 6: Logic and Decision Theory.

ALP can also be used to generate and decide between different candidate sets of actions to achieve higher-level goals. The problem of deciding between different candidate sets of actions is the topic of classical decision theory. I argued that ALP unifies and generalises both abduction and decision theory, as well as other applications such as production systems, complex event recognition and complex process generation. Tanaka-sensei, who attended all the talks, contributed greatly to the discussion, and supported the idea of unifying so many, otherwise diverse areas.

Workshop on Logic and English.

I asked PhD students to prepare abstracts of their PhD theses, or of research papers. These were sent to me before my visit to Hokudai. I commented extensively on the abstracts and sent them back to the students, so they could have time to revise them before the workshop. During the workshop, we spent a little more than half an hour discussing each abstract. The students and other researchers contributed greatly to the discussions. In my opinion, the discussions confirmed that logic can significantly improve the clarity and coherence of written communications. The workshop also had the beneficial side effect that students were able to explain the topic of their research to other students. Hopefully, this may lead to greater interaction among the students themselves in the future.

University of Kyushu, 6-8 November.

I gave two lectures of one and a half hours each, and ran a three hour workshop on Logic and English.

Lecture 1: Computational Logic and the Language of Thought (LOT).

This lecture included topics from my first and fourth Hokudai lectures.

Lecture 2: Computational Logic and Decision Theory.

This lecture included topics from my fifth and sixth Hokudai lectures. I argued that abductive logic programming (ALP) unifies many otherwise diverse areas of computing, including decision theory and abductive explanation of observations. This claim goes against the mainstream belief that different application areas require different approaches. During the discussion, it was noted that the approach that I advocate seems to share a number of features with ASP (answer set programming). We discussed the relationships and the differences.

Workshop on Logic and English.

I asked PhD students to prepare abstracts of their PhD theses, or of research papers. These were sent to me before my visit to Kyushu. I commented on the abstracts and sent them back to the students before the workshop.

Differently from Hokudai, I introduced the workshop with a short talk, providing some general ideas about Logic and English for participants who did not attend my first lecture. Also, there were many participants in the workshop who did not submit abstracts. We spent between 20-30 minutes discussing each abstract. One of the general conclusions to emerge from the discussions is that it is hard to achieve a good balance between writing too much about a topic and writing too little. We used the analogy between the logical structure of a text and the shape of a pyramid or triangle. It is important to make sure that the top of the pyramid is presented at the beginning of the text. An abstract should then present some upper portion of the pyramid, without going all the way down to the bottom. This analogy seemed to be helpful in suggesting a better structure for many of the abstracts.

As in Hokudai, one of the beneficial side effects of the workshop was that the students were able to explain their research to other students, with whom they may have had no previous contact.

Presentation of research in Professor Yokoo’s laboratory.

In the morning of my lectures, Professor Yokoo and members of his lab presented an overview of their research in auction mechanism design and coalition formation. I was not familiar with either of these areas before the presentations. I found both topics very interesting. Moreover, I could see strong similarities between the work on coalition formation in Professor Yokoo’s group and my own work on abductive logic programming. We discussed these similarities, and it would be interesting to look at them more closely in the future.

Presentation of research in Professor Yamashita’s laboratory.

In the morning of the Logic and English workshop, Professor Yamashita and members of his lab presented an overview of their research in robot networks and probabilistic algorithms. As in the case of Professor Yokoo’s research. I was not familiar with either of these areas before the presentations. I was very interested in the robot network research, because it is related to some of my own work on reactive multi-agent systems. Because my work is related to a number of other areas, I was able to suggest that some of these areas, such as the area of coordination languages, and especially the language Gamma might also be relevant to the work on robot networks. Unfortunately, the other area of research presented at the meeting was too far removed from my research background for me to be able to provide any useful comments or suggestions.

Tokyo Institute of Technology, 13 November.

I gave a short talk on Logic and English followed by a two and a half hour workshop.

During the workshop, we discussed five abstracts.

They dealt with different topics, but illustrated similar issues. Without a doubt, the most common problem was that many of the abstracts did not explicitly state at the beginning of the abstract the problem that the paper was addressing. Sometimes, the problem was mentioned later in the abstract, either in one place or scattered about. In other cases, it was missing completely. In one case, it was not possible to identify the problem even after about 20 minutes of discussion.

In all the abstracts, we were able to observe a goal-subgoal relationship between different parts of the work reported in the paper and summarised in the abstract. Sometimes this relationship was presented in an orderly manner. Sometimes, it was presented in a more haphazard fashion, which made the abstract harder to understand. I argued that in general it is helpful to present the goal-subgoal relationship top-down, presenting the top-level goal before lower-level subgoals. This is analogous to drawing a triangle starting at the vertex, and working downwards.

We also observed that a good logical structure is to first present previous work, identify its shortcomings, and then state how the new work presented in the paper addresses these shortcomings. This is an extension of Joesph William’s guidelines for good writing style, which state that a well-constructed sentence will typically begin with familiar information and end with new information. Looked at from this point of view, we observed that one of the abstracts was written backwards, presenting new information and subgoals before old information and higher-level goals. Much to my amazement, one of the abstracts was written perfectly, with previously mentioned information referred to at the beginning of sentences, followed by new information, which was introduced at the end of the sentences.

I argued, and I think that many of the participants agreed, that by analysing the logic of the abstracts, and thinking about how to rewrite them to make them easier to understand, we also achieved the side effect of improving the logic of our understanding of the topics of the abstracts.

JAIST, 20-22 November.

I met Professor Tojo and President Professor Katayama in the morning on 20 November.

We discussed the organisation of JAIST, and work on Artificial Intelligence and Knowledge Science at JAIST. We had a very interesting discussion about how the mind represents thoughts in relation to how humans express their thoughts in natural language. One interesting issue that we discussed was the suggestion that the relationship between brain and mind in humans and perhaps other animals is analogous to the relationship between computer hardware and software. This analogy, in turn, suggests that Brain Science needs to be related to “Mind Science”. Another interesting issue is whether natural language processing requires literal interpretations of texts and utterances, or whether it involves the use of expectations, which may need over-ride literal interpretations.

I gave two lectures and organised a workshop on Logic and English.

The first lecture on 20 November about Computational Logic and Human Thinking was a revised version of talks given earlier at other places. The talk was well attended, and there was a good discussion at the end. Several questions were mainly concerned with clarification. One of the comments drew attention to the similarity between the pyramid or triangle relationship between goals and subgoals on the one hand, and some of the methods developed in Design Science. This drew attention to a richer structure of goals and subgoals than the one presented in my talk: namely to a structure in which different solutions can have different outcomes, including different impacts on goals that are not currently the focus of attention.

Another comment about my talk challenged my proposed relationship between private and public language, in effect using Wittgenstein’s argument that thoughts in a truly private language cannot be communicated from one person to another in any public language. This is a deep issue, which supports the thesis that our public language communications only approximate their intended meanings. Professor Kunifuji, who is the vice-president of JAIST, attended my talk. He works on Discovery Science, and we had a very useful discussion about the criteria that can be used to decide between different solutions of goals. I advocated the criteria of expected utility, as employed in Decision Theory.

Professor Kunifuji advocated a simpler criterion, namely the use of an evaluation function.

We agreed that the evaluation function could be used, as in a branch-and-bound strategy, to generate the best solution possible within the computational resources available.

The second lecture on 21 November was a special lecture proposing a Logic-based Framework for Computing.

I had not given this lecture elsewhere during my visit this year. There were a few questions after the talk, mainly to clarify technical points. In general, the participants seemed to accept that the proposed framework has potential to unify a broad range of otherwise competing computer paradigms.

I especially emphasized the relationship between my proposal and MetaTem, which is a modal temporal logic for reactive computing. The logical semantics of MetaTem is a possible world semantics, and the operational semantics attempts to generate a model in which the reactive rules are true. Modal logic is the focus of much of the work in Tojo-sensei’s laboratory, and this provided a basis for interesting and useful discussions about the similarities and differences between my proposal and the modal logic approach of MetaTem.

On 22 November I organised a workshop on Logic and English.

This workshop was attended by a large number of students, as well as by a number of faculty, including three English language teachers. The workshop mainly involved a public discussion with the writers of abstracts, asking them to explain more clearly what they wanted to explain in their abstracts. I also asked other participants in the workshop to confirm whether or not sentences in the abstracts were intelligible to them. In many cases, we discovered that the abstracts introduced ideas that were not obviously related to earlier sentences in the abstract. This forces serious readers to fill the gaps themselves, by generating the missing links without help from the writer. I was very pleased that the English teachers, not only agreed with my approach, but also recognised that its focus on logical relationships provides a previously neglected technique for improving the clarity and coherence of English writing. I hope to explore this topic further during the next year, before my next visit to NII in the autumn of 2014.

University of Kyoto, 27-29 November.

I gave a one and a half hour talk about Logic and English. The talk provided an introduction to the workshop, similar to the shorter talk I gave before the workshop at TiTech. I expanded the Titech talk significantly, by including a greater amount of technical detail, and a greater number of examples.

I argued that, among the many ways in which Logic can help to improve writing, it can help to structure ideas and their expression in a pyramid (or triangular) form, in which higher-level goals are near the top of the pyramid, and lower-level goals and evidence are at the bottom. Good writing typically starts at the top, and works its way down the pyramid. Yamamoto-sensei suggested, in the discussion, that this pyramid (or tree-like) structure seems to conflict with the more flexible connections that exist among human thoughts. I agreed, but suggested an analogy with graph theory, where any node in a graph can be picked up, and the rest of the graph will then automatically fall into place into a hierarchy of lower levels.

Satoh-sensei also suggested that the top-down presentation of information leaves open the important question about how to order the presentation of ideas and goals at the same level of the pyramid. This is an important point, which is worth exploring further.

I organised a three and a half hour workshop on Logic and English.

The workshop focussed on discussing ways of improving five abstracts written by Master’s and PhD students. Three of the abstracts were clear enough that we were able to have a good discussion about how to improve them further. Interestingly and somewhat surprisingly, it became evident that improving the English writing also uncovered a number of deficiencies in the Logic of the work as it was presented in the abstracts. This showed that improving the logical presentation of writing can also help to improve the Logic of the thoughts that the writing is intended to convey.

Three of the abstracts were almost impossible to understand without intensive interrogation of the students, asking them to explain what they wanted to communicate. The three students seemed to have little idea of what needs to be included in their abstracts for the abstracts to be understood by a reader who is not an expert in the field. After much discussion, we were able to extract the necessary information from the students. I hope that the lessons will have a lasting effect, and that the students’ future writing will be improved.

I had several meetings with Yamamoto-sensei and Satoh-sensei. We discussed the application of unsupervised, case-based, machine-learning to a variety of different fields, including legal reasoning and the problem of extracting knowledge in logical form from natural language texts. Kurohashi-sense joined us for one of our meetings, and we agreed that there is much potential for this work, which is closely related to the problem of extracting knowledge from “big data”.

Other meetings:

Satoh-sensei and I also had a meeting with Dr. Fujita and Mr. Goto, to discuss their work on teleo-reactive robotics. I pointed out that their actions are not durative in the sense of Nilsson’s teleo-reactive programs. I also argued that they need to pay more attention to the relationship between their approach and the current state of the art.

NII Lectures: Computational Logic and Human Thinking

NII, I gave four one and a half hour lectures.

Lecture 1,  11 November.  

The clausal form of logic as the language of thought (LOT). This lecture was a revised version of my first Kyushu lecture. I gave greater emphasis in this talk to the production system model of human thinking, partly because I knew that there would be researchers interested in production systems who would be attending this talk.

There was a lively discussion about the relationship between logic and production systems, and about the difference and relationship between beliefs and goals. In some cases, goals and beliefs can have the same syntax, and it can be hard to tell them apart. We also discussed the role of decision theory and uncertainty. Satoh-sensei, in particular, asked whether decision making can be represented in the same language as the language of goals and beliefs, or whether the two systems, thinking and deciding need to be separated. An interesting suggestion to emerge from this discussion is that perhaps decision making can be compiled into lower-level heuristic rules in a dual process model that also employs a separate decision making module at a higher-level.

Satoh-sensei also raised the interesting question about whether a set of actions, needed to make an agent’s goals true, need to be a minimal set. This is an important issue, which we discussed at length, and about which there is more work to be done, to characterise more appropriate restrictions than simple minimality.

Lecture 2, 18 November.

I attempted to replicate a number of psychological experiments, designed to test whether people interpret English sentences in seemingly logical form according to the rules of classical logic. The participants in the lecture readily cooperated with the experiments. The results were mixed, because they only partly conformed with the results published in the psychological literature. The results of the experiments carried out in the lecture mostly confirmed the hypothesis that people reason differently with beliefs than they reason with goals. However, they showed greater conformity with the rules of classical logic than the results of most psychological studies, perhaps because many of the participants have had previous training in classical logic.

Perhaps more interestingly, the experiments carried out in the lecture also supported the hypothesis that some of the English sentences presented in the experiments did not correctly express their intended meaning. In many cases, the participants interpreted the sentences as expressing their intentions rather than as expressing their literal meaning. This result supported the thesis of the first lecture that understanding the logic of the language of thought can help people to communicate more clearly.

In the second half of the lecture, I returned to the topic of production systems, which are the most influential computational model of human thinking. I proposed a logical semantics for production systems. In the discussion, it was noticed that I had over-simplified the semantics, for the sake of making it easier to understand. A more precise version was presented in the fourth lecture.

Lecture 3, 25 November.

This talk investigated three applications of logic programming to the representation of legal information. The first application was the representation of the British Nationality Act (BNA), whose English sentences are written in a style that is close to logic programming form. The BNA also illustrates the importance of rules and exceptions, whose representation is the main focus of Satoh-sensei’s Proleg language.

The second application was the University of Michigan lease termination clause. This is a poorly written text, whose meaning is made clear by rewriting it into logic programming form. However, the resulting form does not seem very sensible, and it motivated an interesting discussion about its likely intended meaning.

The third application was a collection of rules that I developed with staff in the World Health Organisation in Geneva. The rules are used to estimate yearly infant immunisation coverage in all the countries belonging to the United Nations. They improve upon the previous informal rules, and make the decision of official estimates more consistent and more transparent.

Because Satoh-sensei is an expert in legal reasoning, the talk provoked a very interesting discussion, in which he played an important part.

Lecture 4, 2 December.

Computational Logic can be used to formalise the process of reducing goals to subgoals, including alternative courses of actions. Deciding between different courses of action, in turn, is the topic of classical Decision Theory, which evaluates the expected utility of the consequences of actions and recommends the choice of a course of actions that maximises expected utility.

Computational Logic and classical Decision Theory are, therefore, complementary, and their different strengths can be combined. Moreover, their combination results in a more powerful and more comprehensive framework of human and machine intelligence than either system on its own.

The discussion raised a number of issues, including the fact that the system I presented in the talk does not explicitly incorporate learning. I mentioned that the system can be extended, but is currently restricted for the sake for simplicity. Inoue-sensei asked whether the system can be implemented by means of neural networks, and I replied that conclusions and conditions can be connected by links, which can also be weighted by how often they have contributed to useful outcomes in the past. The resulting network can be interpreted in connectionist terms.