Generation of Referring Expressions: Modeling Partner Effects

Generation of Referring Expressions: Modeling Partner Effects Surabhi Gupta Advisor: Amanda Stent Department of Computer Science

Outline • Introduction • Data • Previous work • Modeling partner effects • Generating NP postmodifiers • A little statistical experiment • Discussion and Future Work

Referring Expressions • A referring expression denotes (or points to) an object in the world of a discourse. • Examples of referring expressions include the red chair, the 400 dollar red chair and 5 red chairs. • Referring expressions are usually noun phrases • Improper construction of a referring expression can result in • referring expressions that are ambiguous (e.g. the book when there are two books). • referring expressions that are too descriptive and lead to false implicatures (e.g. the 400 dollar chair when there is only one chair)

Structure of a Noun Phrase • A definite/indefinite noun phrase is constructed of: • An (optional) determiner or quantifier e.g. a, three • A number of premodifiers (adjectives, adverbs, noun modifiers) e.g. red • A number of postmodifiers (prepositional phrases, relative clauses) e.g. worth 400 dollars, that is red • Other noun phrases include pronouns, proper nouns, deictics green

Adaptation in Conversation • When people talk with each other, they adapt to the other’s choice of referring expression (Clark 1996, Levinson 1983, Brennan 1987). • Example: • (A): Lets buy the 400 dollar red chair • (B): That’s a good idea. The chair matches with the red table. • (A): The chair it is then.

Generation of Referring Expressions in Dialog • When a computer constructs human language, it is called generation • NewsBlaster summaries, or Google translation • Generation for dialog must involve consideration of the dialog partner (the human)

Good Generation of Referring Expressions • The algorithm should generate a referring expression for which the human reader can identify the referent. • The algorithm should generate referring expressions that do not lead the human reader to make false implicatures (Grice 1968). • The algorithm should model how conversational partners adapt to each other. • The algorithm should be able to generate the whole range of referring expressions observed in discourse. • The algorithm should be computationally feasible.

Our Objective • We are building a model of referring expression generation that captures adaptation to partners in conversation. • Related work in this field does not include partner adaptation for dialog (Dale and Reiter 1995, Siddharthan and Copestake 2004).

Data • Two corpora of spoken dialog rich in noun phrases: • Maptask - Speaker A giving instructions to Speaker B about following directions in a map • Coconut - Two participants trying to buy furniture by using both of their inventories and money. • For each corpus, we: • Automatically extracted the noun phrases • Annotated the noun phrases by hand for referent (in a knowledge representation we built), type (noun phrase or pronoun), and to indicate whether the noun phrase was embedded in another noun phrase.

Algorithms Compared • Rule Based • Dale and Reiter 1995 • With partner effects (x 2) • With postmodifier ordering (x 4) • Siddharthan and Copestake 2004 • With partner effects (x 2) • With postmodifier ordering (x 4) • Statistical • Support Vector Machines

Rule-Based Algorithms • Terms used: • Contrast Set: contains information of all the objects in the world. • Preferred list of attributes: the attributes that are known for the objects. • For Coconut: type, quantity, cost, color, state • E.g. three green high tables worth $400 • Intended Referent: The object from the world, which we are trying to describe.

Dale and Reiter • Basic idea: • Specify the preference list by hand • Repeat until all members of the contrast set are gone: • Add the value for the next attribute from the preference list for the intended referent to the noun phrase to be generated

200 dollar green couch 300 dollar red couch 250 dollar brown table • Example: • Preference list: {Type, Color, Cost, Quantity, State} • Contrast set: {300 dollar red couch, 200 dollar green couch, 250 dollar brown table } • Intended referent: 200 dollar green couch • Generated NP: green couch

Siddharthan and Copestake • Basic idea: See Dale and Reiter • Preference list is reordered by using synonyms and antonyms of words in each attribute

Benefits to Rule Based Algorithms • They consider the way humans actually converse ie. humans use unnecessary attributes, they also begin mentioning a referring expression without scanning the entire list of distractors. • They do not attempt to look for the optimal number of attributes. They just go through the list of preferred attributes and iteratively includes those attributes that rule out at least one distractor from the contrast set. • There is no backtracking and the head noun is always included.

Disadvantages to Rule Based Algorithms • They don’t generate the whole range of referring expressions • Ones with postmodifiers • Pronouns • Deictics • They don’t model adaptation to partners.

Adding Partner Effects • A rule based algorithm • Basic idea: See Dale and Reiter, Siddharthan and Copestake • Preference list is reordered to match selection of attributes in previous mentions of the intended referent. • Variant to this where those attributes mentioned previously are definitely included even if all the competitors have been eliminated.

Evaluation • Metric: Correct / Correct + Inserted + Deleted + Moved • Example: • Human: the big fat green cat • Computer: the green happy cat • Correct: the, cat • Inserted: happy • Deleted: big, fat • Moved: green • Score: 2 / 6

Results • The variant to our partner effects algorithms performs significantly better that our Baseline, Dale and Reiter and Siddharthan and Copestake for both the cropora used.

Discussion and Conclusions • The corpus you choose makes a difference • Maptask: Few distractors, no significant different between Baseline, Dale and Reiter and Siddharthan and Copestake • Do partner effects make a difference?

References • Advaith Siddharthan and Ann Copestake. 2004. Generating Referring Expressions in Open Domains. In Proceedings of the 42th Meeting of the Association for Computational Linguistics Annual Conference (ACL 2004), Barcelona, Spain. • Grice, H P (1975). Logic and conversation. In P. Cole and J. Morgan, editors, Syntax and Semantics: Vol 3, Speech Acts, pages 43-58. New York: Academic Press. • Grosz, B and Sidner, C (1986). Attention, intention, and the structure of discourse. Computational Linguistics, 12: 175-206. • Robert Dale and Ehud Reiter. 1995. Computational interpretations of the Gricean maxims in the generation of referring expressions. Cognitive Science, 19:233–263.

Acknowledgements • Dr. Amanda Stent, for all her time and efforts during the last three years. • The Natural Language Processing Lab in Computer Science. • The Honors College for giving me the chance of working on this year long project. • NSF

Questions?

Generating with Postmodifiers • Why? -- because previous algorithms don’t but it’s a big part of the corpus we have used. • Random - randomly decide whether the attribute selected should be a post modifier or premodifier • Unigrams - see where the attribute is in relation to the ‘type’. • Bigrams - statistics of pairs of attributes. E.g probability of finding an attribute given another.

Results

Support Vector Machines • SVMs are a set of machine learning algorithms for binary classification that have been applied to NLP. • We used a set of SVMs, one per attribute, that voted “yes” or “no” to use this attribute at this point in the noun phrase. • Maptask: 6 attributes, Coconut: 5 attributes • We evaluated using: • 10-fold cross-validation for Maptask. • 4-fold cross-validation for Coconut.

Evaluation

Results

Generation of Referring Expressions: Modeling Partner Effects