(***) Build on unification based Kwiatkowski 2010. Key observation is groups of words show same syntactic/semantic tag variation. So learn the variation for the whole group, more robust to data sparsity. i.e. They have discovered that word classes exist :) This helps generalize the language-independent unification approach to unedited sentences like in atis. Mentions Clarke10, Liang11, Goldwasser11 as going from sentences to answers without LF. Mentions Branavan10, Vogel10, Liang09, Poon09, 10 as learning from interactions. Results: (ubl: Kwiatkowsky10, fubl: Kwiatkowski11) atis-exact-f1: zc07:.852 ubl:.717 fubl:.828 geo880-f1: zc05:.870 zc07:.888 ubl:.882 fubl:.886 geo250-en: wasp:.829 ubl:.826 fubl:.837 geo250-sp: wasp:.858 ubl:.824 fubl:.857 geo250-jp: wasp:.858 ubl:.831 fubl:.835 geo250-tr: wasp:.781 ubl:.746 fubl:.731

Conversations provide rich opportunities for interactive, continuous learning. When some- thing goes wrong, a system can ask for clari- fication, rewording, or otherwise redirect the interaction to achieve its goals. In this pa- per, we present an approach for using con- versational interactions of this type to induce semantic parsers. We demonstrate learning without any explicit annotation of the mean- ings of user utterances. Instead, we model meaning with latent variables, and introduce a loss function to measure how well potential meanings match the conversation. This loss drives the overall learning approach, which in- duces a weighted CCG grammar that could be used to automatically bootstrap the semantic analysis component in a complete dialog sys- tem. Experiments on DARPA Communica- tor conversational logs demonstrate effective learning, despite requiring no explicit mean- ing annotations.

(*) Learning with logical forms provided for only part of the data (clarification questions in dialogues). Loss-sensitive perceptron (Singh-Miller and Collins 2007) DARPA communicator corpus. (Walker 2002) Access to logs of conversations where system utterances annotated, user utterances not. Annotations include speech acts (Walker and Passonneau (2001)), these are not predicted for unannotated sentences. Seems like the annotated part of the dialogue (system utterances) can be seen as training set, the rest (user utterances) as test set, is there anything new here? Uses ZC05, ZC07 (template based, not unification). Further reading: Clarke et al. (2010) and Liang et al. (2011) describe approaches for learning semantic parsers from questions paired with database answers, while Goldwasser et al. (2011) presents work on unsupervised learning. Semantic analysis tasks from context-dependent database queries (Miller et al., 1996; Zettlemoyer and Collins, 2009), grounded event streams (Chen et al., 2010; Liang et al., 2009), environment interactions (Branavan et al., 2009; 2010; Vogel and Jurafsky, 2010), and even unannotated text (Poon and Domingos, 2009; 2010). Uses BIU Number Normalizer http://www.cs.biu.ac.il/˜nlp/downloads/

(****) Sentence to logical form mapping using CCG and unification (UBL) instead of GenLex. Geo dataset, four languages, two meaning representations (funql, lambda). Start with single lex item for each sentence mapping it to LF. Introduce vertical bar | to ccg which can match / or \. (ZC07 similar?) Understand the SGD gradient possibly reading CC07. Starting with single lex item and trying splits look much less ad-hoc than ZC05,07 with Genlex and initial lexicon. Only the proper noun NPs (e.g. Texas) are in the initial lexicon. 4.1 splitting constraints interesting, can learn them from data? The split-merge process seem a bit ad-hoc, a more principled Bayesian approach may be possible. Good related work discussion in Sec 6. UBL geo880: p=.941 r=.850 f=.893 UBL-s (2pass): p=.885 r=.879 f=.882

We consider the problem of learning to parse sentences to lambda-calculus representations of their underlying semantics and present an algorithm that learns a weighted combinatory categorial grammar (CCG). A key idea is to introduce non-standard CCG combinators that relax certain parts of the grammar—for example allowing flexible word order, or insertion of lexical items— with learned costs. We also present a new, online algorithm for inducing a weighted CCG. Results for the approach on ATIS data show 86 % F-measure in recovering fully correct semantic analyses and 95.9% F-measure by a partial-match criterion, a more than 5 % improvement over the 90.3% partial-match figure reported by He and Young (2006).

(*) Solving the same problem as ZC05 paper on atis and geo. Geo, jobs, restaurant are artificially generated, atis is natural! New CCG combinators and new online algorithm more flexible with realistic language. atis exact 1-pass p=.9061 r=.8192 f=.8605 atis exact 2-pass p=.8575 r=.8460 f=.8516 atis partial 1-pass p=.9676 r=.8689 f=.9156 atis partial 2-pass p=.9511 r=.9671 f=.9590 (He and Young 2006 atis partial f=90.3%) geo880 1-pass p=.9549 r=.8320 f=.8893 geo880 2-pass p=.9163 r=.8607 f=.8876 (ZC05 p=.9625 r=.7929 f=.8695) Still uses GENLEX with two additional rules. Still uses initial lexicon with nouns and wh-words! CCG additions include: 1. function application with reverse word order. 2. function composition with reverse word order. (do we even need the syntactic cats with slashes?) 3. additional type raising and crossed-comp rules that need more careful reading. Two important differences from learning algorithm of ZC05: 1. online updates instead of batch. 2. perceptron updates instead of SGD on NLL. Learning algorithm: 1. skip example if parsed correctly with current lex. 2. introduce all genlex and find maxscore parse with correct semantics. 3. add the new entries in maxparse to lex and try parsing again. 4. do a perceptron update.

(***) Zettlemoyer and Collins 2005: map sentences to typed lambda expressions: which rivers run through states that border the state with the capital austin (lambda $0:e (and:<t*,t> (river:<r,t> $0) (exists:<<e,t>,t> (lambda $1:e (and:<t*,t> (state:<s,t> $1) (next_to:<lo,<lo,t>> $1 (the:<<e,t>,e> (lambda $2:e (and:<t*,t> (state:<s,t> $2) (capital2:<s,<c,t>> $2 austin_tx:c))))) (loc:<lo,<lo,t>> $0 $1)))))) In addition to the training set of pairs like above they have two additional sources of information given by hand: a. GENLEX rules: to generate candidates for the lexicon, they produce a cross product of (all substrings of the sentence) x (all categories that can be produced from the lambda expression). The second part is done using 10 heuristic rules that convert parts of the lambda expression to possible categories e.g. N:(lambda (x) (major x)). b. Initial lexicon: this includes two types of manually added entries: domain specific, database derived entities such as (Utah => NP:utah) and domain independent entries such as “what” => (S/(S\NP))/N:(lambda f (lambda g (lamba x (and (f x) (g x)))). Both these look like cheating. How sensitive are the results to these manual inputs? What kinds of words are actually learnt at the end? Is it possible to learn entries like “what” at all? Why do we need syntactic categories, can’t we just have lambda expressions? Should look at how future work handled these. Geo880 p=.9625 r=.7929 f=.8695 (TM01 p=.8992 r=.7940) Jobs640 p=.9736 r=.7929 f=.8740 (TM01 p=.9325 r=.7984)