Circumscribing is not excluding: A response to Manning Richard Crouch, Lauri Karttunen, Annie Zaenen Palo Alto Research Center Palo Alto, California 94305, USA Manning [1] presents an extended critique of a paper by Zaenen, Karttunen and Crouch [2] (henceforth ZKC), which was a commentary on the PASCAL Recognizing Textual Entailment (RTE) challenge [3]. The ZKC paper was enti- tled “Local textual inference: can it be defined or circumscribed?” and argued that the PASCAL data should have been better defined and circumscribed so as to distinguish between different forms of textual inference. Some misunder- standing has arisen, and Manning takes us to be in favor of excluding certain forms of textual inference rather than just distinguishing between them. In this note we clarify how the considerations put forward in ZKC serve not to narrow the textual inference task, as Manning claims, but to broaden it. We illustrate this by discussion of the AQUAINT Knowledge-based Evaluation (KBEval) data and annotation guidelines [4] 1 , which we played a large role in formulating. Manning’s principal complaint is that we seek to narrow the definition of textual inference “so as to exclude many of the inferences that humans make and many of the inferences that are needed for operational use of robust tex- tual inference” ([1] p. 2). In fact the opposite is true. In ZKC we argued only that the impact of world knowledge and plausible inference should have been circumscribed in the PASCAL RTE data, not that it should have been excluded. That is, the data annotation would have benefited from drawing lines around and distinguishing between inferences based solely on linguistic knowledge and those also based on world knowledge, and between strictly valid inferences and merely plausible inferences. It was not our intent to exclude all but one form of inference, and this should be evident from the AQUAINT KBEval annotation guidelines. A comparison of the PASCAL and AQUAINT annotations schemes moreover shows that the distinctions ZKC argued for supports a broad and open definition of textual inference. Recall that the PASCAL data provides Text/Hypothesis pairs, annotated with whether the hypothesis does, or does not, follow from the text. The data is also divided into seven classes (IR, CD, RC, QA, IE, MT, PP) reflecting the kind of application the data is relevant to. 1 ZKC only offers two paragraphs of recommendations on improving the PASCAL data annotation. However, the annotation guidelines for the AQUAINT KBEval are spelled out in detail in [4].