240 likes | 278 Vues
Explore how to manage structural ambiguities in natural language generation by selecting the optimal referring expressions. Utilizing corpus-based heuristics and empirical evidence, the study presents a method to estimate the likelihood of different readings of a phrase, enhancing the generator's output clarity. By considering psycholinguistic evidence and using word sketches to identify distractor interpretations, this research aims to determine the most effective approach in choosing between ambiguous expressions. Experiments demonstrate preferences for clear, concise referring expressions in different linguistic contexts. Discover the intricate balance between clarity and brevity in creating effective referring expressions.
E N D
Generation of Referring Expressions: Managing Structural Ambiguities I.H. Khan G. Ritchie K. van Deemter University of Aberdeen, UK
A natural language generator should avoid generating those phrases, which are too ambiguous to understand. But, how the generator can know whether a phrase is too ambiguous or not? We use corpus-based heuristics, backed by empirical evidence, that estimate the likelihood of different readings of a phrase, and guide the generator to choose an optimal phrase from the available alternatives.
Natural Language Generation (NLG) • Process of generating text in natural language (e.g., English) from some non-linguistic data (Reiter & Dale, 2000) • Example NLG system • Pollen Forecast: generates reports from pollen forecast data Grass pollen levels for Tuesday have decreased from the high levels of yesterday with values of around 4 to 5 across most parts of the country. However, in South Eastern areas, pollen levels will be high with values of 6. [courtesy E. Reiter]
Generation of Referring Expressions (GRE) • Referring Expression = Noun Phrase • e.g., the black cat; the black cats and dogs (etc.) • A key component in most NLG systems • Task of GRE: • Given a set of intended referents, compute the properties of these referents that distinguish them from distractors in a KB
GRE: An Example • Input:KB, Intended Referents R • Task: find properties that distinguish R from distractors KB • Output: Distinguishing Description (DD) • (Black Sheep) (Black Goat)
NP1: The black sheep and the black goats = {Object1,Object3,Object4,Object6} (Black Sheep) (Black Goat) NP2: The black sheep and goats (Black Sheep) Goat = {Object1,Object3,Object4,Object5,Object6,Object7} The Problem • Linguistic ambiguities can arise when DDs are realised • NP1 unambiguous and long; NP2 ambiguous and brief • Question: How the generator might chose between NP1 and NP2?
Our Approach • Psycholinguistic evidence • Avoidance of all ambiguity is not feasible (Abney, 1996) • Avoid only distractor interpretations • An interpretation is distractor if it is more likely or almost as likely as the intended one. • Question • How to make distractor interpretation precise? • Our solution • Getting likelihood using word sketches (cf. Chantree et el., 2004) • Word sketches provide detailed information about word relationships, based on corpus frequencies • Relationships are grammatical
Pattern: the Adj N1 and N2 • Hypothesis 1 • If Adj modifies N1 more often than N2, then a narrow-scope reading is likely (no matter how frequently N1 and N2 co-occur). bearded men and women handsome men and women • Hypothesis 2 • If Adj does not modify N1 more often than N2, then a wide-scope reading is likely (no matter how frequently N1 and N2 co-occur).. old men and women tall men and trees
Experiment 1 Please, remove the roaring lions and horses.
Experiment 1: Results • Hypothesis 2(i.e., predictions for WS reading) is confirmed • Hypothesis 1(i.e., predictions for NS reading) is not confirmed • Tendency for WS (even though results are not stat. sig.) • Tentative conclusion • An intrinsic bias in favour of WS reading • BUT: The use of *unusual* features may have made people’s judgements unreliable
Experiment 2 Please, remove the figure containing the young lions and horses.
Experiment 2 (cont.) • Results: Both hypotheses are confirmed Please, remove the figure containing the barking dogs and cats.
The black sheep and the black goats (Black Sheep) (Black Goat) The black sheep and goats • Word Sketches can make reasonable predictions about how an NP would be understood. • But we need more to know from generation point of view: which of the following two NPs is best? (Black Sheep) Goat • We seek the answer in next experiment
Clarity-brevity trade-off • Recall the pattern: the Adj Noun1 and Noun2 • Brief descriptions (+b) take the form • the Adj Noun1 and Noun2 • Non-brief descriptions (-b) take the form • the Adj Noun1 and the Adj Noun2 (IR = WS) • the Adj Noun1 and the Noun2 (IR = NS) • Clear descriptions (+c) • Which have no distractor interpretations • Non-clear descriptions (-c) • Which have some distractor interpretations
The Hypotheses (Readers’ Preferences) • Hypothesis 1 • (+c, +b) descriptions are preferred over (+c, -b) • Hypothesis 2 • (+c, -b) descriptions are preferred over (-c, +b) • Each hypothesis is tested under two conditions • C1:intended reading is WS • C2: intended reading is NS
Experiment 3: NS Case • Which phrase works best to identify the filled area? • The barking dogs and cats • The barking dogs and the cats
Experiment 3: WS Case • Which phrase works best to identify the filled area? • The young lions and the young horses • The young lions and horses
Experiment 3: Results • Both hypotheses are confirmed: • (+c, +b) descriptions are preferred over (+c, -b) • (+c, -b) descriptions are preferred over (-c, +b) • Role of length: • In WS cases preferences are very strong • In NS cases preference is not as strong as in WS cases
Summary of Empirical Evidence • For the pattern the Adj Noun1 and Noun2 • Word Sketches can make reliable predictions • Keeping clarity the same, a brief NP is better than a longer one
Algorithm Development • Main knowledge sources • WordNet (for lexicalisation) • SketchEngine (for predicting the most likely reading) • Main steps • Choose words • Use these to construct description in DNF • Use transformations to generate alternative structures from DNF • Select optimal phrase
Transformation Rules • Input • Logical formula in DNF • Rule Base • (A B1) (A B2) A (B1 B2) • (X Y) (Y X) [A = Adj, B1=B2=Noun, X=Y=(Adj and/or Noun)] • Output • Set of logical formulae
Select optimal phrase • (black sheep) (black goats) DNF • (black goats) (black sheep) • black (goats sheep) • black (sheep goats) Optimal (4):Adj has high collocational frequency with N1 and N2, so the intended (wide-scope) reading is more likely. Therefore, (4) is selected.
Conclusions • GRE should deal with surface ambiguities • Word sketches can make distractor interpretation precise • Keeping clarity the same, brief descriptions are preferred over longer ones • A GRE algorithm is sketched that balances clarity and brevity