NLog-like Inference and Commonsense Reasoning

NLog-like Inference and Commonsense Reasoning Lenhart Schubert1 Recent implementations of Natural Logic (NLog) have shown that NLog provides a quite direct means of going from sentences in ordi- nary language to many of the obvious entailments of those sentences. We show here that Episodic Logic (EL) and its Epilog implementation are well-adapted to capturing NLog-like inferences, but beyond that, also support inferences that require a combination of lexical knowl- edge and world knowledge. However, broad language understanding and commonsense reasoning are still thwarted by the “knowledge ac- quisition bottleneck”, and we summarize some of our ongoing and contemplated attacks on that persistent difficulty. 1 Introduction: Natural Logic, Episodic Logic, and Natural Logic (NLog, as we will abbreviate it) provides an attrac- tive framework for implementing many “obvious” entailment inferences, starting with little more than a syntactically analyzed sentence (e.g., MacCartney and Manning 2008). But in itself, NLog is certainly in- sufficient for language understanding or commonsense inference. We advocate here a formalized (not necessarily fully disambiguated) rep- resentation that is very close to language, along with an inference methodology that easily accommodates NLog inferences, but beyond that enables more general reasoning based on both lexical knowledge and world knowledge. This approach to representation and inference has been pursued by our group at the University of Rochester (and 1University of Rochester 45 LiLT Volume 9 Perspectives on Semantic Representations for Textual Inference. Copyright c ⃝ 2014, CSLI Publications. 46 / Lenhart Schubert prior to that, at the University of Alberta) for over two decades, and the outcome so far is the Episodic Logic (EL) representation and its implementation in Epilog systems 1 and 2. We have also built up very substantial amounts of general knowledge,2 much of it rough, but a good deal of it precise enough for Epilog-based inference. In the next section we review the EL representation and the Epilog inference architecture. In section 3 we illustrate NLog-like inferences in Epilog, and touch on some evaluation results for such inferences. Section 4 provides examples of commonsense inferences performed by Epilog that lie beyond the capabilities of NLog, in part because they involve radical di↵erences in phrase structure between the premises and the conclusion and in part because they draw on world knowledge as well as lexical semantic knowledge. However, enhanced inference ca- pabilities alone cannot take us very far towards achieving broad un- derstanding and commonsense reasoning: We still face the knowledge acquisition bottleneck, and in section 5 we discuss multiple facets of our e↵orts to provide broad knowledge to Epilog. In section 6 we as- sess the prospects for general knowledge acquisition through genuine understanding of WordNet glosses, Simple Wikipedia entries, or the Open Mind collection of general statements. We sum up the relation- ship between NLog and Epilog in section 7, also briefly reiterating the status of our knowledge acquisition e↵orts and the prospects for moving beyond them through deeper understanding of general statements. 2 EL and Figure 1 provides a glimpse of the EL representation and …

Content truncated to 500 words