We present a new problem for the society of AI, that is to collect evidence useful for explaining the meaning of an observed event. Evidence here is a set of pieces of useful and new information, which may come from the open real world, for understanding the target observation. The information may not be necessarily accepted widely enough to be learned by machine learning but is novel knowledge or claim within personal or local messages, which supports a query positively or negatively. We address this presentation to the proposal of Evidence-based Semantics, which means to obtain a novel explanation of the meaning of an event, situation, action, utterance, message, etc., that is critically required in various real-world application domains.
Learning reusable patterns, extracting useful parts of data, and searching for data hitting user’s interest are already daily tasks for a machine. By combination, it is easy to collect datasets and use them for knowledge discoveries if the user can express his/her own interest. The recent Chat GPT, where the user enters the query in natural language, returns an answer which may appear to come from the machine’s understanding of the query and relevant information in collected various datasets.
However, it is still an open problem to respond
to quite a simple and widely perceived requirement,
that is to collect useful information from the open
space of information for explaining the meaning of
a query i.e., an event observed in the real life. The
required information may not be collected from the
open data or public data market processed by
machine learning, but may be novel knowledge or
claims within personal or local messages. We call
this kind of information evidence and address this
short paper to the proposal of Evidence-based
Semantics, which means to explain the meaning of a
target observation (an event, situation, action,
message, etc) by collecting evidence. In contrast to
the literature [
EBS refers to the problem to obtain E and h’, given and h as in Clauses (1) through (3), where G and respectively refer to the target observation and prepared knowledge. can be probabilistic, but it is consistent with new knowledge or observation. ⋃ h
⊢ G ⋃ h’ ⊢ G ⋃ E ℎ ≠ ℎ′ ⋃ h’ ⊬ □ (1) (2) (4) h, and h’, are respectively hypotheses to entail G, and G together with new observation E. E is the evidence. a collection of additional observations (ei’s in Fig.1) supporting h’. G can be entailed by the previous hypothesis h, but the new hypothesis h’ is preferred due to its ability to entail both G and E.
Two approaches to EBS are shown in Figure 1, where (a) represents the entailment structure above, where e and h are observed events other than G and hypotheses respectively. The arrows show parts used for explaining (entailing) G and e. By use of data visualization in (b), h’s are invisible (not observed) but a hypothesis of higher confidence is reflected here as one at a shorter distance from observations derived from it. Here we regard these two figures as the bases of approaches toward EBS.
Using the multi-layer KeyGraph [
EBS can be regarded as a restoration of the
situational semantics [
This study is supported by JST COI-NEXT JPMJPF2013, MEXT Q-Leap JPMXS0118067246, JSPS Kakenhi 20K20482. We thank Lawyer R. Uchino for advices about evidence collection in disputes and statements in this paper.