Figure 26 – uploaded by Suzanne Wertheim

See full PDF downloadDownload figure

Table 2: Impact of various axiom sources on span-selected Metaphor Mapping performance. Learned and Hybrid ap- proaches generally outperform the Manual approach to axiom collection. — Table 2 Impact of various axiom sources on span-selected Metaphor Mapping performance. Learned and Hybrid ap- proaches generally outperform the Manual approach to axiom collection.

Related Figures (67)

Figure 2: Distribution of warrior metaphor centered around diagnosis event (x-axis: months from event, y- axis: average frequency of metaphor usage)

il Certain metaphor/event pairs show a peak around he event, or at 1 year after the event, for example on he anniversary of diagnosis, which is a significant event in the life of a cancer patient. However, the pattern does not hold across all such pairs, making it difficult to generalize. For example, in Figure 1, we see a peak of metaphor frequency occurring at he time of the event, but in Figure 2, we do not see such a peak at the time of the event, but see other peaks both before and after the event date. Another complicating factor is that different users experience different cancer treatment timelines. For instance, one user might experience these events over a long period of time, whereas another user may encounter these events in quick succession (Wen and Rosé, 2012). These factors motivated us to consider other methods, including hierarchical mixed models, for more in-depth analysis. Figure 1: Distribution of journey metaphor centered around diagnosis event (x-axis: months from event, y- axis: average frequency of metaphor usage)

Table 2: Metaphor candidates and their associated events

Table 1: Corpus-wide unambiguous popular metaphor use statistics (among posts where the user used the metaphor at least once) (M: posts that contain each metaphor, L: posts that do not contain each metaphor).

Table 4: Metaphor candidates and their associated events

Table 3: Metaphor use statistics of data used for MTurk (* indicates metaphor candidates for which the literal usage is more common than the non-literal one, N: nonliteral use L: literal use).

Table 5: Performance on metaphor disambiguation evaluation. (6) is significantly better than (5) [p=0.013] (fs.: used feature selection)

Table 2: Performance of a model with AutoWeighted training examples in comparison to the unweighted base- line, in terms of Precision (P), Recall (R), and F-1 score (F) for the positive (’metaphor’”) class. A-B and B-A cor- respond to training-testing scenarios where the system is trained on Set A and tested on Set B and vice versa, re- spectively. All other figures report average performance across the cross-validation folds.

Figure 1: Weights assigned to the different concreteness bins by the logistic regression classifier with the baseline feature set in an unweighted training regime. The bins span the 1-5 range with 0.25 increments; words falling in bin | are the most abstract, while words falling in bin 17 are the most concrete. Table 3: Performance of a model with optimally weighted training examples in comparison to the auto-weighted scheme, in terms of Precision (P), Recall (R), F-1 score (F), and the difference between Recall and Precision (D). A-B and B-A correspond to training-testing scenarios where the system is trained on Set A and tested on Set B and vice versa, respectively. All other figures report average performance across the cross-validation folds.

Table 4: Performance of a model trained with unweighted examples with and without DC (difference in concrete- ness) features.

Table 5: Performance of a model trained with auto- weighted examples with and without DC (difference in concreteness) features.

Table 7: Performance the baseline model UPT+CBins in the baseline configuration (unweighted) the UPT+CUpDown+DCUpDown model in opti-weighted configuration.

Table 8: Instances of the adjective full in Set B that are predicted to be non-metaphors by the UPT model trained on Set A in the unweighted regime, while the UPT+CUpDown+DCUpDown model classifies these as metaphors. The noun that is recognized as being in the amod relation with full is shown in square brackets. FULL (small caps) indicates an instance that is annotated as a metaphor; lowercase version corresponds to a non- metaphor annotation.

Table 1: Percentages of POS and metaphors per register, with standardised residuals in brackets (n=47934

Concreteness for literal (L) and nonliteral (NL) for heads vs. dependents, by P Figure |: Plot of concreteness scores for literal vs. nonliteral/metaphorical heads vs. their dependents, in the VUAMC grouped by parts of speech

Figure 2: Plot of imageability scores for literal vs. nonliteral/metaphorical heads vs. their dependents, in the VUAMC. grouped by parts of speech

Figure 3: Plot of anew scores for literal vs. nonliteral/metaphorical verbs in the VUAMC

Table 4: Rules extracted from random forest models for Nouns (len=length of condition, freq=proportion of data instances meeting condition, err=proportion of incorrectly classified instances as over instances meeting condition, SD=standard deviation) Table 3: Results (Accuracy, F1) for random forest models for different POS, for predicting non/literal items, for models with and without the “weight” term w

Table 2: Results (F1) of evaluating models with different combinations of features, for predicting non/literal items (n=3574)

Table 5: Rules extracted from random forest models for Verbs (len=length of condition, freq=proportion of data instances meeting condition, err=proportion of incorrectly classified instances for instances meeting condition, SD=standard deviation, depslist=list of dependents)

Table 6: Rules extracted from random forest models for Prepositions (len=length of condition, freq=proportion of data instances meeting condition, err=proportion of incorrectly classified instances for instances meeting condition) concreteness, imageability, w, and a small but inter- esting subset of the ANEW categories, across POS. For example, while nouns make good use of con- creteness, im ageability and w, combinations of im- ageability and w are prevalent for verbs and prepo- sitions. Fur words, seem tures, but pre her, nouns and verbs, being content to make good use of the ANEW fea- positions make no use of such features, perhaps due to their status as function words. More careful study of a wider range of rules, as well as possible conditioning environments is required, and such suggestions remain tentative. Note also that one complication in all of this is that there are exten- sive errors for most of the extracted rules, and close study of possible sources of such errors is planned

Figure 1: Directionality of sensory modalities as pro- posed by Williams(1976).

Table 1: Feature sets used in the experiments.

Table 2: The cross validation and test accuracies of the baseline with and without vector space features.

Table 3: Performance of the B’ baseline in combination with the different sets of sensorial features. Statistical significance: ***, p < .001; **, p < .01; *, p < .05.

Figure 2. Non-metaphoric frame network pattern The MetaNet project represents an effort to formal- ly represent and categorize metaphors and frames that comprise their source and target domains, and relations between them. Frames are coherent, con- ceptual gestalts organized in a hierarchical struc- ture. These range from experiential universal structures such as Motion Along a Path and Verti- cality, to more specific situations such as Physical Restraints and Disease; they also include less phys- ically concrete culturally-based frames like Pov- erty and Corruption. More-specific frames incorporate the semantics and internal structure of the more-general frames they inherit from, forming a complex network of related concepts. Relations between frames define how elements of a parent frame are incorporated by the child frame. For in- stance, the ‘subcase of’ relation indicates that the child fully inherits and elaborates the structure of the parent frame. In addition to traditional ontolog- ical relations, we also include relations specific to frame semantics and metaphor theory. A fragment of this network is illustrated in Figure 2.

Metaphors are essentially representations of mappings between frames. The structure of the source domain frame maps onto the structure of the target domain frame (Figure 3); hence, in POVERTY IS A DISEASE, the impoverished people of the Pov- erty frame are understood as the patient experienc- ing the disease in the Disease frame. Specifically, the roles of the Disease frame map onto their coun- terparts of the Poverty frame.

expression bear trap. Hence, disambiguating be- tween these two types of expressions requires a second step of metaphoricity evaluation.

specify document structural constraints, including grammatical constraints. To search texts for con- structional matches, we construct Resource De- scription Framework (RDF) models of each sentence in terms of an ontology defined in the Web Ontology Language (OWL). The ontology defines the classes Document, Sentence, and Word, and properties, some of which are shown in Table 2 and with their domain and range.

Figure 4. Structures accessed by phrase cure poverty.

Table 3. Performance over gold standard data To evaluate the accuracy of the metaphor extractor, linguists collected attested sentences and annotated metaphoric expressions for the target domains Government, Bureaucracy, Democracy, Poverty, Taxation, and Wealth; they annotated all in- domain metaphoric expressions in the sentences where both the target and source were explicitly realized. Sentences were manually annotated for source and target word forms, source and target frames, and the construc tional pattern used to ex- press the metaphor. The metaphor extractor was run on these col ected go d standard sentences, and the output compared to the annotations entered by the linguists. Table 3 sho tions in the gold standard ws the number of annota- , the recall (percentage of gold standard examples that were identified), and the precision ( percentag e of extracted examples that were correct) of the system for three lan- guages.

Figure 6. Number of extracted metaphor expressions that evoke various translational motion frames

The richness of the extracted data and the struc- tural relations identified via our repository enable us to analyze data at varying levels of specificity. For instance, a search of the English Gigaword corpus (Graff & Cieri 2003) for metaphor expres- sions with the target domain ‘poverty’ revealed several interesting patterns. Firstly, at the very general level of frame families, we observe that the most frequently occurring source terms were either location/motion related (e.g. being at a location, translational motion, motion impediments) or in- volved physical harm (e.g. disease, physical com- bat, or other harmful encounters). Figure 5 shows the relative frequency of these frame families.

Figure 7. Number of different lexical units in extracted data that evoke the Downward Motion frame. Looking at a yet more specific level, we can ex- amine which lexical items are used to evoke a giv- en frame. Figure 7, below, shows this data for the Downward Motion frame. Looking at a yet more specific level, we can ex- amine which lexical items are used to evoke a giv- en frame. Figure 7, below, shows this data for the Downward Motion frame.

Figure 1: Abduction-based metaphor processing pipeline.

Table 1: The number of metaphoric sentences in the devel- opment and test sets for each language. However, for a goal of high-precision mapping, sophisticated learning methods are not necessary. In- stead, we require that a chosen percent of the in- stances of a logical form fragment correspond to a single source or target concept, in which case we out- put a lexical axiom mapping the LF to the concept. We found it helpful to enforce the mutual exclusivity of the text fragments that map to source concepts and hose that map to target concepts. When a text frag- ment is ambiguous between a source mapping and a arget mapping, we produce an axiom for whichever correspondence was more frequent. This reduces the ikelihood of the axioms leading the system to iden- ify, e.g., two source concepts in a sentence but no arget concept. The results are axioms like

Table 2: Statistics for the corpus of annotations being released. Each category consists of one or more scenarios. For a variety of sentences, the corpus gives instantiations (slot—value pairs) of the slots in the schema for each scenario. Each linguistic metaphor is also tagged as being about one or more source and target concepts. All counts are for unique entries, except for slot—value pairs, which includes duplicates when they occur in the data. Some sentences contain more than one metaphor, so the number of unique sentences is less than the sum of unique sentences for each schema.

Table 1. Distribution of collected data across targets in Gun Case debate Step four is data collection from the sites closer to the middle of the opinion spectrum identified in the second step. When this data is processed for metaphors, they are labeled accordingly as “mod- erate”. We note that “moderate” positions in mul- ti-side conflicts may have different interpretations than in a largely binary conflict of Gun Case. In Table 1, the column Total Passages represents the sum total of passages processed from the extreme and moderate websites.

To determine predominant Conceptual Meta- phors for each protagonist (21% and 20% referred above), we rank order the Source Domains (SDs) for each side by number of LMs that use each SD. In Table 4, we show the predominant conceptual metaphors used for key targets by each protagonist. Table 4. The most representative CMs on both sides of the Gun Debate, by key Targets. Font size indicates relative frequency for top CMs for each target.

Table 3. Conceptual Metaphors used by protagonists on the GOVTO side Table 2. Conceptual Metaphors used by protagonists on the INDO side

Figure 6. The INDO side dominates debate with use of more action-oriented language. Size of pie chart repre- sents the proportion of metaphors in the source role cat- those source domains that are used at least 2% of the time overall by both sides and the count of distinct rela- tions in the LMs by each protagonist. Y-axis represents count of metaphors.

Figure 4. The INDO side (red bars) dominates debate with use of more LMs overall. Here we show those source domains that are used at least 2% of the time overall by both sides and the count of LMs for those Source Domains. Y-axis represents count of metaphors. used in LMs that are of the “taking action” type (primarily verbs describing events, such as “at- tack’) vs. the “passively observe” (primarily nouns and adjectives, such as “devastating’’).

Figure 5. The INDO side dominates debate with richer vocabulary suggesting greater influence. Here we show a oe ah a La a. ae Ss. + ae oe,

Using this method, we find that the extreme proponents of the INDO and GOVTO sides are far apart, approximately 0.55 of the maximum theoret- ical distance of 1.0. Using the same measures, the distance between the INDO moderate position and both INDO and GOVTO extremes is approximate- y half of the above, or 0.27. This places the INDO moderate position in the center of the spectrum of positions between the two extremes. On the other hand, the language used by the GOVTO moderate position places them closer to the GOVTO ex- reme. This finding is illustrated in Figure 8.

To establish topical positioning, we first identify meso-topics that are present in a debate, as dis- cussed in Section 4.1. We then distinguish multiple forms in which polarization or valuation is applied to meso-topics in protagonists’ utterances such as through express advocacy or disadvocacy or via supporting or dissenting information, and express agreement or disagreement with a polarized state- ment made in a statement by the same or another protagonist. We create Topical Positioning Vectors representing each protagonist. Table 5 shows a fragment of positional vectors for extreme GOVTO and INDO positions for five meso-topics. In these vectors, value in each cell represents a prevailing combined polarity and intensity towards a meso-topic. We note that meso-topics form a su- perset of metaphorical targets as explained earlier. Table 5. Topical Positioning vectors for extreme GOVTO and INDO positions in the gun debate Topical Positioning vectors can now be used to calculate distance between protagonists, using standard cosine measure. We used this method to compute 4-ways distances in the Gun Case: be- tween the extreme positions on each side; between the moderate and extreme positions within each side; as well as between moderates and extremes across the sides and compared the distances so ob- tained to those obtained from metaphorical matri- ces (Section 5.3). We note that both methods

results. With the frequency information, we can see the top 10 most frequent pairs that fit the intui- tion of Chinese native speakers (see Table 1). It might be surprising to see that “> it] fashionable” ranks top of all pairs. However, it makes sense because Apple (the brand) products are popular in China. “4K#§ spirit” often refers to a young female demon/spirit who seduces people in Chinese fairy tales. The remaining 8 words repre- sent ordinary things people experience in everyday life.

Table 2. Four Kinds of Word-Property Pairs But we find some important phenomena in the re- sults that will be introduced in section 4.2. 4 Statistics

Next we explore what the most frequent words and properties are in Chinese CogBank. This is important as we could learn what the most com- mon entities are that people tend to use as vehicles in similes, and the most common properties people prefer to express in everyday life. As shown in Ta- ble 4, nouns like flower, man, water, child, human, cat, angel, wolf and sunshine rank the highest in the database. These words are quite common in everyday life and they have hundreds of properties. But the top 10 properties of each word dominate more than half the occurrences of these words when employed in a simile. This indicates that people always rely more heavily on a word’s sali- ent properties to form a simile expression.

Table 5 shows the most frequent properties in CogBank: beautiful, bloom, fight, fly, convenient, warm, and painful. Each property is associated with hundreds of words. But the frequency of the top 10 concept words occupies more than half the occurrences. This indicates that people is tend to use the same kinds of vehicles to form a simile expression.

Table 5. Top 10 Most Frequent Properties in CogBank 4.2 Valuable Information from Uninforma- tive Data

The manual data check drops many uninformative data which on the surface seem to possess no value, for example, “as stupid as yow’’, “as cheap as be- fore’. The pronouns and time expressions have to be removed from CogBank. But through observing all the pronouns and time expressions through manual data check, we find something useful in Chinese sentences “X {& (as) Y —#¥ (same) A” (X is as A as Y) where Y is the reference object. As Indicated in Table 6, people prefer to use #¢(J) as the reference object rather than other pronouns.

People also prefer to reference recurring and concur- rent time frames over past or future ones. As shown in Table 7, usual (7£ 7%, #7) occurs more than past and before, while future(#X) occurs with even lower fre- quencies. q t S t J The usage patterns showing much higher fre- uencies for the pronoun #(/) and time expression tE i, Ff (usual) suggest that people prefer to use heir experienced everyday life knowledge to make imile or contrast sentences. This finding supports he Embodied Cognition Philosophy (Lakoff & ohnson 1980; Lakoff 2008), which hypothesizes t hat much of our conceptual structure is based on knowledge formed through physical and emotional experience.

The word 44(pig) is in the center surrounded by its properties. [he width of the line indicates the frenqucies of the “word-property” pair:

The 4#f(pig) has many properties(yellow nodes), and these yellow nodes have their words(orange nodes). So the extendted orange nodes share more properites with 44(pig) appear closer to the center.

The word 4f(pig) is in the center surrounded by its properties. Each property is linked to a bilingual sememe in HowNet (blue nodes).

E. The comparison graph of “sheep” with translation “==” extended by HowNet’s sememes