Learning Context-Dependent Mappings from Sentences to Logical Form

Proceedings of the 47th Annual Meeting of the ACL and the 4th IJCNLP of the AFNLP, pages 976–984, Suntec, Singapore, 2-7 August 2009. c⃝2009 ACL and AFNLP Learning Context-Dependent Mappings from Sentences to Logical Form Luke S. Zettlemoyer and Michael Collins MIT CSAIL Cambridge, MA 02139 {lsz,mcollins}@csail.mit.com Abstract We consider the problem of learning context-dependent mappings from sen- tences to logical form. The training ex- amples are sequences of sentences anno- tated with lambda-calculus meaning rep- resentations. We develop an algorithm that maintains explicit, lambda-calculus repre- sentations of salient discourse entities and uses a context-dependent analysis pipeline to recover logical forms. The method uses a hidden-variable variant of the percep- tion algorithm to learn a linear model used to select the best analysis. Experiments on context-dependent utterances from the ATIS corpus show that the method recov- ers fully correct logical forms with 83.7% accuracy. 1 Introduction Recently, researchers have developed algorithms that learn to map natural language sentences to representations of their underlying meaning (He and Young, 2006; Wong and Mooney, 2007; Zettlemoyer and Collins, 2005). For instance, a training example might be: Sent. 1: List flights to Boston on Friday night. LF 1: λx.flight(x) ∧ to(x, bos) ∧ day(x, fri) ∧ during(x, night) Here the logical form (LF) is a lambda-calculus expression defining a set of entities that are flights to Boston departing on Friday night. Most of this work has focused on analyzing sen- tences in isolation. In this paper, we consider the problem of learning to interpret sentences whose underlying meanings can depend on the context in which they appear. For example, consider an inter- action where Sent. 1 is followed by the sentence: Sent. 2: Show me the flights after 3pm. LF 2: λx.flight(x) ∧ to(x, bos) ∧day(x, fri) ∧ depart(x) > 1500 In this case, the fact that Sent. 2 describes flights to Boston on Friday must be determined based on the context established by the first sentence. We introduce a supervised, hidden-variable ap- proach for learning to interpret sentences in con- text. Each training example is a sequence of sen- tences annotated with logical forms. Figure 1 shows excerpts from three training examples in the ATIS corpus (Dahl et al., 1994). For context-dependent analysis, we develop an approach that maintains explicit, lambda-calculus representations of salient discourse entities and uses a two-stage pipeline to construct context- dependent logical forms. The first stage uses a probabilistic Combinatory Categorial Grammar (CCG) parsing algorithm to produce a context- independent, underspecified meaning representa- tion. The second stage resolves this underspecified meaning representation by making a sequence of modifications to it that depend on the context pro- vided by previous utterances. In general, there are a large number of possi- ble context-dependent analyses for each sentence. To select the best one, we present a weighted lin- ear model that is used to make a range of parsing and context-resolution decisions. Since the train- ing data contains only the final logical forms, we model these intermediate decisions as hidden vari- ables …

Content truncated to 500 words