Automating the Encoding for the LISA Model of Analogy from Raw Text

Abstract

Analogy is an integral part of human cognition. Consequently researchers would like to produce computational models of analogy to test implementations of theoretical claims. Such computational models have relied almost exclusively upon hand-coded representations, making the resulting models too dependent upon the modeler's choices, and thus di cult to interpret in isolation of the modeler. In an e ort to combat this dependency, we present here a means of automatic encoding for the LISA model of analogy and inference.

1 INTRODUCTION

Analogy, and more generally, relational reasoning are core cognitive capacities [ 6, 8, 10 ] that, along with language, seem to separate human cognition from the cognitive abilities of all the other great apes [ 22 ]. Many of the computational models of analogy arising from machine learning revolve around solving semantic, or 4-way, analogies of the form “king:queen::brother:?” where the model lls in “sister” [ 19 ]. ese models fail to account for the more substantial form of analogy – narrative analogy – wherein agents and actions must all be analogically mapped to one another. Numerous models have been developed to account for our ability to make such analogies. Among the most in uential have been Falkenhainer et al.’s [ 3 ] Structure Mapping Engine (SME), Holyoak and agard’s [ 9 ] Analogical Constraint Mapping Engine (ACME) and Hummel and Holyoak’s [ 11, 12 ] Learning and Inference with Schemas and Analogies (LISA) (see Gentner & Forbus, 2011[ 7 ], for a review).

Although these and other models account for many aspects of human relational thought, they share a common limitation: All of them rely on hand-coded knowledge representations, placing a great deal of control in the modeler’s hands. As such, it is o en unclear how many of the successes of these models re ect the models’ fundamental tenets and how many re ect arbitrary representational and modeling choices made by the modeler at the time of creating the simulations (see, e.g., Lu et al. 2008 [ 17 ]). Since such models are only as good as their modeler, the support or challenge to any theoretical claims they make must be interpreted carefully.

Ideally, these models would produce their representational inputs directly from natural language, removing their dependence upon hand-coded inputs. ere are some models that approach this goal for speci c narrow domains, such as intuitive physics (Friedman SE & Forbus KD, 2009 [ 5 ]) and moral decision making (Dehghani M et al., 2008 [ 2 ]). However, the task of generalizing open domain natural language for analogy remains elusive. To address this problem, we describe recent e orts applying several Natural Language Processing (NLP) techniques c 2016 ACM. /2016/1-ART1 $ DOI: 1:2 • to the automatic generation of knowledge representations for LISA. When combined, these techniques allow us to transform English sentences into semantically-rich propositional representations usable to the LISA model of analogy (i.e., representations in “LISAese”). 1.1

Representation

LISAese is a collection of semantic encodings and structural representations which, together, allow LISA to produce semantically informed, structurally consistent analogies. LISA employs semantic representations for both objects and predicates and encodes structure in a predicated logical form. Semantics for an object in LISAese consist of a distributed list of abstract “semantic units” and corresponding weights. For instance, if we were to hand-code the representation for a taxi driver, “Bob” and a bank teller “Jill”, we may end up with the following semantics and corresponding weights:

Bob: human=1.0, adult=0.7, male=1.0, employed=0.8, taxi driver=1.0 Jill: human=1.0, adult=0.7, female=1.0, employed=0.8, bank teller=1.0

Bob’s semantics could then overlap with the semantics for our bank teller, “Jill”, who is also human, adult, and employed, but would di er on male and taxi driver. LISA treats these “semantic units” as universal across all representations, that is, if two “semantic units” have the same name, they are the same unit. Aside from this universality, the names are arbitrary and the choice to give them relevant names with regard to the objects they are a ached to only serves to facilitate human readability.

Similarly, predicate semantics consist of “semantic units” and weights; however, predicate semantics are de ned at the level of their relational roles rather than the predicate itself. For example, the two-place predicate love is de ned by two lists of semantics. e rst list represents the agent role of love (the lover); the second represents the patient role of love (the beloved). is could be represented as follows:

love: [ lover=1.0, emotion=1.0, strong emotion=0.8, passionate=0.9 ] [ beloved=1.0, receives emotion=1.0, receives strong emotion=0.8, receives passion=0.9 ] With de ned semantics on objects and predicates, LISA is able to ground its structural representation, given in a predicated logical format, to represent “Bob loves Jill”: love(Bob, Jill). is logical representation is robust to recursive references such as “Jill knows that Bob loves her.” which yields: know-(Jill, love(Bob, Jill)). Indeed, this is su cient to represent the necessary semantic and structural information LISA needs to make analogies and inferences. 1.2

Task

Our task in this paper is to outline a technique to produce semantic and structural representations automatically from text. Subsequently, we will show that LISA can use the resulting LISAese to successfully nd analogies and inferences. More speci cally, we will present a method which achieves the following: (1) Generates Semantics • Object semantics that account for modi cation by adjectives, noun-phrases, and copular verb a achments • Predicate semantics with role-by-role representation (2) Generate Structure

Automating the Encoding for the LISA Model of Analogy from Raw Text • 1:3

• Produces logic form • Handles recursive references • Associates multiple references to a single entity By using NLP techniques to automatically pre-process and encode text into LISAese, the LISA model will no longer need to rely on hand-coded knowledge representations. Freeing it of this requirement will clarify how LISA’s successes as a model re ect its fundamental tenets rather than choices made by the modeler. 2

GENERATE SEMANTICS

Ultimately, we need to generate semantic representations for both objects and predicates, but the requirements for automated generation of the two di er slightly. is is because, as previously noted, in LISAese semantics are assigned not to the predicates themselves, but rather to each predicate role independently. 2.1

Object Semantics

Objects may be represented in a fairly straightforward manner using word embedding. Word embedding is a means of mapping words into a semantic space (a vector space) such that words which are ‘similar’ are closer in the resulting space than words which are ‘dissimilar’. ere are many means of producing word embeddings. One of the most commonly used semantic embedding tools is Google’s Word2Vec [ 19 ] which produces vectors based o carefully cra ed word co-occurrence statistics. Using these vectors, we can produce embeddings for individual words such as “cat”, “January”, or “happy”. However, this means of embedding can break down upon encountering proper nouns such as “Microso ” or “Bob”. In order to properly encode proper nouns, we employ an NLP technique called Named Entity Recognition (NER) which identi es what kind of entity the proper nouns are: “Microso ”!ORGANIZATION and “Bob”!PERSON. Using these NER tags, we can encode the proper nouns by the semantic embedding of their associated tag (for added speci city, where known, we also incorporate the gender of each proper noun).

In order to accommodate various modi ers, we additively combine the semantic vectors corresponding to each object. ese vectors consist of the embeddings of the words contained within or associated with an entity, for example, the words in a noun-phrase, an NER label, adjective modi ers, or copular modi ers. Copular verbs are stative verbs such as is/are/was/were, seem, or appear. For example, “smelly Sam” and “Sam is smelly” both result in the incorporation of the semantics for “smelly” in the representation of Sam.

Using the resulting vectors, we can produce LISAese by creating an arbitrary semantic label, call it “SemUnit i”, to every ith index of the vector where the weight on the generated semantic unit is the value at the ith index of the vector. For example, if our word embedding was 4 dimensions (instead of 300), we could have the following embedding for “happy”: happy! [0.14, 0.0024, 0.81, -0.37]

Which would then yield a LISAese representation of “happy”:

happy! SemUnit 0=0.14, SemUnit 1=0.0024, SemUnit 2=0.81, SemUnit 3=-0.37 us allowing us to produce object semantics in LISAese from the Word2Vec word-embeddings. 1:4 • Predicate semantics in their most general form are more di cult to automatically encode than objects. Direct role-by-role representation via word-embedding is certainly non-obvious, although we will explore methods to do so in the future work section. Here, we will simplify our semantic representation to the easiest usable form wherein we assign semantics to each role corresponding to word embeddings for the predicate. For example, in our 4-dimensional embedding from before, we can imagine the following vector for “love”: ”love” ! [0.23, 0.01, 0.76, -0.45]

Copying the embedding onto the agent and patient roles of love gives us:

love: [SemUnit 0agent=0.23, SemUnit 1agent=0.01, SemUnit 2agent=0.76, SemUnit 3agent=-0.45] [SemUnit 0patient=0.23, SemUnit 1patient=0.01, SemUnit 2patient=0.76, SemUnit 3patient=-0.45] While this representation meets all of our requirements, it does have a glaring weakness. ere will never be any semantic overlap between two di erent predicate roles (e.g. agent and patient). e resulting LISAese will make LISA preferentially map agents and patients onto other agents and patients respectively. Again, we will explore alternatives to this encoding which do not su er from the same weakness in the future works section below.

Using the above semantic representations, we meet all the requirements for LISAese’s semantic encoding and are ready to explore how to structurally represent text in LISAese. 3

GENERATE STRUCTURE

In order to encode language in the kind of predicated logic LISAese requires, we need to identify predicate roles and the objects which ll them. In the simplest case, we may be presented with text such as “Bob loves Jill”. In which there is one predicate with two roles, agent and patient, lled by “Bob” and “Jill” respectively. As mentioned earlier, this structure becomes more di cult when predicates ll the roles of other predicates as in “Jill knows Bob loves her”. In this later case, the structure we hope to capture is that “Jill” is the agent and love is the patient of know. To discover these a achments, we may employ an NLP technique called Semantic Role Labeling (SRL). SRL identi es predicates and their roles as well as portions of the text which ll those roles. Using SRL, we can label “Jill” as the agent and “Bob loves her” as the patient of know. Similarly, “Bob” is identi ed as the agent, and “her” the patient of love. For our work, we use the SRL engine laid out in Punyakanok, Roth, and Yih [ 24 ] which uses joint inference over several SRL systems and Integer Linear Programming for global constraint satisfaction to optimize its labeling. A er using our SRL engine, we have the following structural encoding: know(Jill, love(Bob, her)). is is nearly enough to properly structure the input for LISA since we are able to encode the predicates in a LISA’s logical form and are able to handle recursive references to predicates as arguments to other predicates. e last remaining piece is to properly associate “Jill” with “her”. In order to achieve this, we use another NLP technique called ‘co-reference resolution’ [ 18 ]. Co-reference resolution associates multiple references to the same entity throughout a text. Using this technique, we are able to associate “her” with “Jill”, yielding our nal representation: know(Jill, love(Bob, Jill)).

e automated structural encoding, along with the semantic encoding, is su cient to produce functional LISAese without the involvement of any hand-coding and li le modeler input. However, this automatic encoding is useless if LISA is not able to function on the resulting LISAese. In order to test this, we explore three benchmark simulations LISA has successfully completed in the past: Latin to Logic, Love Triangle, and Driving a Vehicle.

Automating the Encoding for the LISA Model of Analogy from Raw Text • 1:5

4

SIMULATIONS

To test our automated representations, we will compare LISA’s discovered mappings and inferences on them against the hand-coded LISAese. Since our automatic encoding produces several hundred semantics for each object and role, we have omi ed them from this paper. However, since there were some di erences in some of the structural representations between the hand-coded and automatic representations, we have provided both for each benchmark below.

For “Latin to Logic”[14], the benchmark representation is shown below:

Analog 1: is incomplete(Second Order Logic)

Analog 2: is self referential(Natural Language) has unprovable statement(Natural Language) is incomplete(Natural Language)

LISA can use this produced representation to nd the analogy between Second Order Logic and Natural Language and infer that Second Order Logic has unprovable statements and that it is incomplete.

In order to test our automatic representation, we need to identify English sentences which convey a similar meaning. Running LISA over the representations generated by these sentences, we can check if it produces the correct mappings and inferences. For that purpose, we chose the following stories: “Second Order Logic is incomplete.” and “Latin is able to self reference. Latin proves and disproves a statement. Latin is incomplete.” Using the automatic representational system described here, we produced the following representation:

Analog 1: be(Second Order Logic, incomplete)

Analog 2:

be(Latin, able to self reference) prove(Latin, a statement) disprove(Latin, a statement) be(Latin, incomplete)

Running LISA on this representation does in fact produce a mapping between Second Order Logic and Latin. LISA makes both of the desired inferences, namely that Second Order Logic is self referential and that it can prove and disprove a statement. Worth noting is that the structure generated by our automated system is quite di erent from that of the original representation. However, the correct mapping and inference was still produced.

For “Love Triangle” [ 12 ] the hand-coded benchmark structure and the automatic structural encoding were identical, and it is shown below:

Analog 1:

love(Sally, Jack) 1:6 •

“Latin to Logic” is the easiest of our three benchmarks since the structure is simple and semantic similarity alone is enough inform mapping. In “Love Triangle”, while the structure remains simplistic, the object semantics are inversely correlated with the ideal mapping. at ideal mapping is from “Sally”!“John”, “Jack”!“Mary”, and “Megan”!“Bob”, which is counter-indicated by semantic similarity on the subset of semantics corresponding to gender. In the hand-coded version, LISA nds this ideal mapping and produces the correct inference, namely hates(John, Bob) in Analog 2. A er running LISA on the automatically produced LISAese, the same mapping and inference was found.

Lastly, we will look at the most complicated benchmark, “Driving a Vehicle” [ 12 ]. Its hand coded structure is shown below:

Analog 1: has(Joan, car) want(Joan, goto(Joan, airport))

Analog 2:

has(Sam, car) want(Sam, goto(Sam, beach)) drive(Sam, car, beach) e automatic encoding produced a very similar structure: Analog 1: have(Joan, a car) want(Joan, go(Joan, the airport))

Analog 2: have(Sam, a jeep) want(Sam, go(Sam, to the beach)) drive(Sam, a jeep, to the beach)

To produce this, the NLP pre-processor was given “Joan has a car. She wants to go to the airport.” for Analog 1 and “Sam has a jeep. He wants to go the beach. He drives the jeep to the beach.” for Analog 2.

On both the hand-coded and auto-coded LISAese, LISA found the correct mapping that “Joan” is like “Sam” and the “jeep” is like the “car”. LISA also managed to produce the correct inference that Joan drives her car to the airport.

Automating the Encoding for the LISA Model of Analogy from Raw Text • 1:7

5

CONCLUSION

Using NLP techniques to pre-process and encode text into LISAese has proven to be a fruitful approach to incorporating the encoding process into LISA as a model. With the incorporation of this linguistic front-end, LISA is now the rst computational model of analogy capable of encoding simple open-domain text and performing analogical mapping and inference over it. More importantly, the LISA model no longer needs to rely on hand-coded knowledge representations. With this, it is now clearer how LISA’s successes as a model re ect its fundamental tenets rather than the representational and modeling choices made by the modeler. 6

FUTURE WORK

We hope to improve both the semantic and structural encoding in the future with be er semantic libraries and relational capture. For object semantics, we can use sparse over-complete vectors [ 4 ] for the word embeddings, since it has previously been shown that LISA performs be er over sparse encodings. For predicate semantics, we can explore role speci c semantics so that predicate role semantics can overlap between labeled roles (some agents overlap with some patients). Towards that end, we intend to explore several di erent tracks for embedding. We will look at PropBank [ 15 ] labels for verb speci c roles and use the embeddings of the words in their labels (agent of love is “lover”, patient of love is “loved”). We will also look at embedding the roles directly, either by forming role co-occurence chains and using Word2Vec on those chains directly, or by replacing the spans of text by their SRL label and running Word2Vec over the replaced text. However, regardless of what embedding techniques we use, the representation technique as we’ve described it falls pray to a common word embedding weakness – inability to di erentiate phrasal cconstructs. at is, “the cat on the blanket” vs “the blanket on the cat” have the same semantic representation. is pitfall can be avoided by capturing only the head of noun phrase (with adjectives) and instead predicating any prepositional relationships contained within the phrase. We will also explore additional ways to extract structure from language in the form of explicitly stated intra-sentential cause and e ect: “Bob loves Susan because she is kind.” We will store the extracted information in “Cause-E ect” groups [ 13 ].

In addition to improvements of the encoding, we hope to expand our use of LISA. With the availability of computational resources as well as abundant text, auto-encoding LISAese allows for an explosion of analogical modeling. As research progresses, we hope to use auto-coded LISA’s mappings and inferences as learning features on a range of traditional NLP tasks. Success in these domains would demonstrate the importance of analogy as a core cognitive capacity and its salience to a host of higher level computational tasks, for example schema generation [ 23 ][ 1 ], the Story Cloze Task [ 20, 21 ], story classi cation or even provide some insight into the Winograd Schema challenge [ 16 ].

ACKNOWLEDGMENTS

is material is based on research sponsored by the Air Force Research Laboratory, under agreement number FA9550-12-1-0003. e U.S. Government is authorized to reproduce and distribute reprints for Governmental purposes notwithstanding any copyright notation thereon.

e authors would also like to thank Hani Awni and Megan Emigh for their input throughout the process. 1:8 •

Wilner & Hummel

References

[1] Niranjan Balasubramanian , Stephen Soderland, Oren Etzioni Mausam, and Oren Etzioni . 2013 . Generating Coherent Event Schemas at Scale. . In EMNLP . 1721 - 1731 .

[2] Tomai E. Forbus K. Klenk M. Dehghani , M. 2008 . An Integrated Reasoning Approach to Moral Decision-Making. AAAI Conference on Arti cial Intelligence (AAAI) Twenty- ird ( 2008 ).

[3] Brian Falkenhainer , Kenneth D Forbus , and Dedre Gentner . 1989 . e structure-mapping engine: Algorithm and examples . Arti cial intelligence 41 , 1 ( 1989 ), 1 - 63 .

[4] Manaal Faruqui , Yulia Tsvetkov, Dani Yogatama, Chris Dyer, and Noah A. Smith . 2015 . Sparse Overcomplete Word Vector Representations . CoRR abs/1506 . 02004 ( 2015 ). h p://arxiv.org/abs/1506.02004

[5] Sco E Friedman , Kenneth D Forbus , and Jason Taylor . Learning and Reasoning with alitative Models of Physical Behavior. ( ).

[6] Dedre Gentner . 1983 . Structure-mapping: A theoretical framework for analogy . Cognitive science 7 , 2 ( 1983 ), 155 - 170 .

[7] Dedre Gentner and Kenneth D Forbus . 2011 . Computational models of analogy . Wiley interdisciplinary reviews: cognitive science 2 , 3 ( 2011 ), 266 - 276 .

[8] Mary L Gick and Keith J Holyoak . 1983 . Schema induction and analogical transfer . Cognitive psychology 15 , 1 ( 1983 ), 1 - 38 .

[9] Keith J Holyoak and Paul agard. 1989 . Analogical mapping by constraint satisfaction . Cognitive science 13 , 3 ( 1989 ), 295 - 355 .

[10] Keith James Holyoak and Paul agard. 1995 . Mental leaps .

[11] John E Hummel and Keith J Holyoak . 1997 . Distributed representations of structure: a theory of analogical access and mapping . Psychological Review 104 , 3 ( 1997 ), 427 .

[12] John E Hummel and Keith J Holyoak . 2003 . A symbolic-connectionist theory of relational inference and generalization . Psychological review 110 , 2 ( 2003 ), 220 .

[13] John E Hummel and David H Landy . 2009 . From analogy to explanation: relaxing the 1: 1 mapping constraint very carefully . In New Frontiers in Analogy Research: Proceedings of the Second International Conference on Analogy. New Bulgarian University, 211 - 221 .

[14] John E Hummel , John Licato, and Selmer Bringsjord . 2014 . Analogy, explanation, and proof . Frontiers in human neuroscience 8 ( 2014 ).

[15] Paul Kingsbury and Martha Palmer . 2002 . From TreeBank to PropBank .. In LREC. Citeseer.

[16] Hector J Levesque , Ernest Davis, and Leora Morgenstern . 2011 . e Winograd schema challenge .. In AAAI Spring Symposium: Logical Formalizations of Commonsense Reasoning , Vol. 46 . 47.

[17] Hongjing Lu , Alan L Yuille, Mimi Liljeholm, Patricia W Cheng, and Keith J Holyoak. 2008 . Bayesian generic priors for causal learning . Psychological review 115 , 4 ( 2008 ), 955 .

[18] Christopher D. Manning , Mihai Surdeanu, John Bauer, Jenny Finkel, Steven J. Bethard , and David McClosky . 2014 . e Stanford CoreNLP Natural Language Processing Toolkit . In Proceedings of 52nd Annual Meeting of the Association for Computational Linguistics: System Demonstrations . 55 - 60 . h p://www.aclweb.org/anthology/P/P14/P14-5010

[19] Tomas Mikolov , Ilya Sutskever, Kai Chen, Greg S Corrado, and Je Dean . 2013 . Distributed representations of words and phrases and their compositionality . In Advances in neural information processing systems . 3111 - 3119 .

[20] Nasrin Mostafazadeh , Nathanael Chambers, Xiaodong He, Devi Parikh, Dhruv Batra, Lucy Vanderwende, Pushmeet Kohli, and James Allen . 2016 . A corpus and cloze evaluation for deeper understanding of commonsense stories . Proceedings of NAACL HLT , San Diego, California, June. Association for Computational Linguistics ( 2016 ).

[21] Nasrin Mostafazadeh , Lucy Vanderwende, Wen-tau Yih , Pushmeet Kohli, and James Allen . 2016 . Story Cloze Evaluator: Vector Space Representation Evaluation by Predicting What Happens Next . ACL 2016 ( 2016 ), 24 .

[22] Derek C Penn , Keith J Holyoak , and Daniel J Povinelli. 2008 . Darwin's mistake: Explaining the discontinuity between human and nonhuman minds . Behavioral and Brain Sciences 31 , 02 ( 2008 ), 109 - 130 .

[23] Karl Picho a and Raymond J Mooney . 2016 . Statistical Script Learning with Recurrent Neural Networks . EMNLP 2016 ( 2016 ), 11 .

[24] V. Punyakanok , D. Roth , and W. Yih . 2008 . e Importance of Syntactic Parsing and Inference in Semantic Role Labeling . Computational Linguistics 34 , 2 ( 2008 ). h p://cogcomp.cs.illinois.edu/papers/PunyakanokRoYi07.pdf