Computing Point-of-View: Modeling and Simulating Judgments of Taste
Download 438.05 Kb.
|
- Navigate this page:
- Syntactic frame extraction.
- Step 2—topic extraction.
- Step 3—textual affect sensing.
- Step 4—statistical estimation.
- 3.5 Technology: commonsense reasoning
- Structure and mechanism.
- 3.5 Technology: textual affect sensing
| Part-of-speech tagging. Tokens are assigned part-of-speech tags such as VB (verb, root form), NN (common noun singular), and JJR (adjective, comparative form) using the Penn Treebank tagset [], based on Eric Brill’s transformation-based learning tagger for English []. Lemmatization. The lemma, or normal form, for nouns and verbs are generated. Lemmas are important supplemental information added as annotations to tokens. Lexical features such as number are stripped from nouns (e.g. “robots” ==> “robot”), and tense is stripped from verbs (e.g. “went” ==> “go”). Anaphora resolution. An anaphor is a referring expression such as a pronoun (e.g. “he”, “they”) whose referent usually lies in the immediately prior sentences. As the reader scans the textual tokens sequentially, a deixis stack of possible referents such as noun phrases, are maintained. When an anaphor is encountered, it is resolved with the aid of the deixis stack, according to the knowledge-poor resolution strategy outlined in [mitkov]. Phrase chunking. From a flat sequence of tagged tokens, phrases will emerge as the boundaries of phrases are identified, e.g.:
“John/NNP likes/VBZ to/TO play/VB board/NN games/NNS” ==> (NX John NX) (VX likes to play VX) (NX board games NX) Here, NX denotes noun chunks, and VX denotes verb chunks. Moving from words to the level of chunks allows text to be regarded on the conceptual level. Phrase chunking is accomplished by a set of regular expression patterns operating over the stream of words and tags. Phrase linking. To inch toward a syntactic parse tree, verb chunks need to be linked to their noun phrase and prepositional phrase arguments. Accomplishing this requires some heuristics about verb-argument structure, as well as semantic selectional restrictions, gotten from common sense knowledge in ConceptNet. The following example illustrates the successful resolution in light of ambiguity: (NX John NX) (VX robbed VX) (NX the bank NX) with (NX the money) ==> (John (robbed
(the bank (with (the money)) Note that “the money” was linked to “the bank” and was not implicated as the second argument to the verb “robbed”. This mechanism makes the common sense assumption that “the money” was not the instrument used by John to perform the robbery, though such a scenario is certainly possible, though not pragmatic. More challenging cases of linking arises in the encounter of surface movement phenomena such as subject-verb inversion (e.g. “have you the money?”), topicalization (e.g. “to the bank I go”), and passive voice (e.g. the subject is nested in an agentive “by” phrase in the utterance “The bank was robbed by John”). Syntactic frame extraction. Finally, the event structure of each sentence can be captured concisely as a syntactic frame, e.g.: SENTENCE FRAME [verb] “robbed” [subject] “John” [direct object] “the bank” “with the money” A sentence frame may contain any number of direct and indirect objects. A sentence frame is constructed for each clause in the text. A dependent clause has a frame which is linked to the frame of the clause upon which it depends. Step #2—topic extraction. After open texts are normalized into phrases and events, one major task central to psychoanalytic reading is to identify significant topics which are talked in textual passages. Ascending from textual tokens into topics is moving from unstable lexemes up into the scale of more stable classemes. Topics constitute various classemes in the schemas of self-expressive texts, such as the ‘topic-classeme’ for attitude viewpoint, ‘subculture-classeme’ for cultural taste viewpoint, and ‘cuisine-classeme’ for gustation viewpoint. Note that ‘perception-classemes’ cannot be arrived at via topic extraction. The term ‘topic’ is a bit underspecified since each granularity of text has a different sort of topic. The topic of the utterance “George Washington was great” is “George Washington,” but if the utterance occurred in the context of giving examples of lots of presidents, then “president” is a more inclusive topic. If the discussion of presidents occurs in a history book, then the over-arching topic is “history.” Thus, every granularity of text can be assigned a set of relevant topic phrases, which may be explicit in the text, or absent from the text. Assessing what topics are central to a segment of text is akin to summarization. Topic extraction is performed using various folksonomies and the ConceptNet common sense reasoning toolkit. For example, AllMusicGuide (AMG) [] is a folksonomy about music—its artists, genres, albums, and moods. Suppose a textual passage talks about the ‘rhythm and blues’ musical genre, but does mention the genre by name, instead mentioning various artists and albums in the genre. During the semantic recognition process, the AMG folksonomy is used to identify artist names and album names in the text passage. Sentence-level topic extraction concludes that artists “Al Green,” “Peven Everett,” “Alicia Keys” and albums “My Brazil” and “Songs in A Minor” are topicalized in the sentences. Each recognized entity entails a set of weighted metadata annotations—albums entail their artists, and artists entail the genre. All recognized entities entail the ‘rhythm and blues’ genre, so a simple scoring approach allows ‘rhythm and blues’ to be assigned as a salient topic of the whole passage. Various folksonomies are layered atop ConceptNet, and the latter is capable of sensing generic topics, which do not belong to the specialized discourse of music, books, films, and such. The ConceptNet topic extraction mechanism is presented in Section 3.5. Step #3—textual affect sensing. The goal of textual affect sensing is to scan over text and annotate it—at the various scales of phrase, sentence, paragraph, and document—with its affect valence score, given in the Pleasure-Arousal-Dominance format of Mehrabian []. The semiotic action of overlaying a textual passage’s extracted topics with its sensed affect is to project semes and classemes into the emotive register. Harkening to Jakobson’s model of communication, seme and classeme are born of ‘message’ and ‘context’ respectively, while textual affect is the emotive trace of the message ‘sender’, which is impressed upon each message and context. The textual affect sensor built for psychoanalytic reading is constituted from three sorts of knowledge—1) a database of sentiment words (e.g. ‘cry’, ‘sad’) annotated with affective valence; 2) a database of non-sentiment words (e.g. ‘homework’, ‘recess’) annotated with their typical affective connotation as measured in psychological focus groups; and 3) a database of commonsense knowledge affording typical affects for everyday concepts and events (e.g. ‘be(person,fired)’). By combining lexical and eventual treatments, affect can be sensed as a more robust combination of both surface language and deep meaning. A common textual affect sensing serves all five viewpoint schemas and work over all sorts of self-expressive texts. However, each schema appropriates the sensed affect differently, because each requires its own genre of ‘emotive register’. Attitudes schema maps the text’s extracted topics into the full PAD space. Humor schema combines Displeasure, high Arousal, and Dominance into a psychic tension. Cultural taste schema uses textual affect to infer which lexemes in the text are probably the writer’s interests, chiefly by measuring the Arousal context of each entity. Finally, the perception schema utilizes the sensed affect which cathects around the writer, ‘alters’, and incoming and outgoing affective transactions between writer and alters. Section 3.6 presents details for the textual affect sensing technology. Step #4—statistical estimation. Having transformed a corpus of raw self-expressive texts into a bag of recognized lexemes and classemes, each contextualized by their emotive valence, statistical methods are applied to infer a stable isotopy from these data points. Recall Greimas’ principle of monosemization—when persons read a narrative, interpretation converges upon an overarching coherent meaning, called an isotopy. In self-expressive texts, this isotopy is an image of the writer’s viewpoint. But how would a machine reader enact monosemization? Under an information theoretic framework, if the text’s parsed lexemes and classemes constitutes a pointillistic impression of a writer’s viewpoint, then perhaps the first-order moment of this cloud of points can be regarded as the isotopy. An attitude isotopy is the set of all stable attitude-classemes—each attitude-classeme is the first-order moment of all the emotive valences associated with its topic-classemes. After taking the first-order moment, self-contradictory topics (e.g. topic X talked about negatively and positively in the text) will fail to exceed a minimum threshold of emotiveness; thus, thresholding, is used to eliminate unstable attitude-classemes. Likewise, for humor isotopy. A cultural taste isotopy is the set of all stable interest-semes and subculture-classemes—each seme and classeme is the first-order moment of the emotive valences of their corresponding lexemes. Likewise, for gustation isotopy. Perception isotopy is the set of all stable perception-classemes (i.e. intuiting, sensing, thinking, feeling) are not first-order moments because the inference is nontrivial; rather, a machine learning algorithm, Boostexter [], learns a classificatory mapping from perception-lexemes into perception-classemes. The application of statistical estimation to assign stable affective characteristics to textual entities builds upon the technique known as semantic orientation [] in the computational linguistics literature. [grefenstette] exhibits numerous applications of semantic orientation to the estimation of stable emotive valences for words from a corpus. The application of semantic orientation for the estimation of sophisticated models of viewpoint in this thesis is believed to be unique in the literature. That statistics can converge upon isotopy assumes that the self-expressive texts being considered 1) can be unified under a coherent meaning that is the writer’s viewpoint (i.e. monosemization) and 2) that the vast majority of textual fragments emanates the overarching spirit of the writer (i.e. aesthetic self-consistency). These assumptions may be especially fair to self-expressive texts such as weblog diaries, social network profiles, and homepages because these are candid, stream-of-consciousness, and non-literary. Greimas, in [], talked about certain texts, such as jokes and sarcastic texts, which cause ruptures of isotopy. For example, a joke narrative leads a reader to converge upon a false isotopy, only to have that isotopy overturned by a punch-line, which forces systematic re-semization of the text, resulting in the actual isotopy. Statistical convergence certainly would draw wrong conclusions from such texts, as a statistical reader is unable to recognize the rupture, and unable to selectively overturn things that it has already read. So, it is assumed that sarcasm and jokes within the chosen self-expressive texts are limited to short episodes within the textual corpus. The statistical estimation mechanism can tolerate a certain amount of these episodes as ‘noise’ in the text. 3.4 Culture mining What is culture? In semiotics, Roland Barthes (1964) proposed culture to be the set of symbols salient to the unconscious of a population. He said also that these symbols are organized into semantic systems and have valence, or degrees of privilege. Similarly, Clifford Geertz (1973) remarked that cultures were ‘webs of significance’ which implicated people into them. What sort of textual corpus affords the capture of culture? A cultural corpus can be one of two things – 1) a large-scale aggregate of self-expressive texts for persons in the cultural population; or 2) a single archetypal self-expressive text of indistinct authorship, which could be considered to epitomize the attitudes, tastes, and dispositions of the culture. The former kind of corpus is phenomenal, while the latter is essential. Thus, the computational technique of culture mining is corollary to psychoanalytic reading, as described above. A most straightforward case of culture mining is exhibited for attitude viewpoint. Just as an individual’s viewpoint is represented as a ‘sheet’ of emotive valences overlaying the sheet of possible topics, the attitudes of a whole culture can likewise be represented as a sheet—at each of its points, the sheet shows the average phenomenal or essential affect cathected by the culture unto a topic—depending upon the genre of the cultural corpus. For example, the What Would They Think? (WWTT) implemented acquisition for attitude viewpoint considered the members of a weblog community called Xanga. Attitude sheets were created for each individual, and an attitude sheet was created for the whole ‘culture’, as the embodiment of the average phenomenal attitude in the culture. By examining the alignments and differences between an individual’s and the culture’s sheet, a sense of the individual’s ‘normalcy’ can be had. Deviation from norms can itself lead to a more potent characterization of a person’s viewpoint than attitudes taken as scalars. Another implemented example is political attitude. Two essential corpora—one of politically liberal texts, and one of politically conservative texts—were assembled from known ideological publications, and WWTT produced sheets for the ‘liberal culture’ and ‘conservative culture’. By examining the degree of alignment of an individual’s attitude viewpoint to either liberal or conservative sheets, it’s possible to characterize that individual’s ideological leaning. Or alternatively, the pair of liberal and conservative cultures can be applied to interpret arbitrary fodder texts—acting as ‘virtual pundits’, so to speak. A more non-trivial case of culture mining is exhibited for cultural taste viewpoint. The set of social network profiles of all available individuals constitutes a phenomenal corpus representation of the whole ‘culture’. The first three steps of psychoanalytic reading, including natural language normalization, topic extraction, and textual affect sensing are performed as usual. However, in the statistical estimation step, rather than learning just the stable seme and classeme profile for an individual, the taste fabric which underlies the whole culture must be learned. The weaving of the taste fabric is a machine learning feat. By regarding the semes and classemes possessed by each individual as containing mutual information, a machine learning algorithm finds the pointwise mutual information (PMI) affinity score which connects every possible pair of interest semes and subcultural classemes. After atrophying away the weak connections, what is left is a semantic fabric of stable interconnections between semes and subcultures. An individual’s profile of semes and classemes can then be considered in relation to the cultural taste fabric, as occupying a region, or location, atop the fabric. 3.5 Technology: commonsense reasoning Commonsense reasoning via ConceptNet [] is a key technological subsystem invoked to support both viewpoint acquisition systems in psychoanalytic reading and the building and simulation of the judgmental apparatus; it also provides one of several technical foundations for another key technology: textual affect sensing. This section 1) briefly reviews ConceptNet’s origin with respect to the computational genre of common sense reasoning systems, 2) the structure and mechanism of commonsense reasoning in ConceptNet, and 3) the specific affordances of ConceptNet for textual topic extraction, and conceptual analogy. § Background. ConceptNet is a freely available commonsense knowledge base and natural-language-processing toolkit which supports many practical textual-reasoning tasks over real-world documents including topic-gisting, analogy-making, and other context oriented inferences. The knowledge base is a semantic network presently consisting of over 1.6 million assertions of commonsense knowledge encompassing the spatial, physical, social, temporal, and psychological aspects of everyday life. ConceptNet is generated automatically from the 700,000 sentences4 of the Open Mind Common Sense Project [] (as of — a World Wide Web based collaboration with over 14,000 authors. What is commonsense knowledge? Of the different sorts of semantic knowledge that are researched, arguably the most general and widely applicable kind is knowledge about the everyday world that is possessed by all people — what is widely called ‘commonsense knowledge’. While to the average person the term ‘commonsense’ is regarded as synonymous with ‘good judgment’, to the AI community it is used in a technical sense to refer to the millions of basic facts and understandings possessed by most people. A lemon is sour. To open a door, you must usually first turn the doorknob. If you forget someone’s birthday, they may be unhappy with you. Commonsense knowledge, thus defined, spans a huge portion of human experience, encompassing knowledge about the spatial, physical, social, temporal, and psychological aspects of typical everyday life. Because it is assumed that every person possesses commonsense, such knowledge is typically omitted from social communications, such as text. A full understanding of any text then, requires a surprising amount of commonsense, which currently only people possess. The ConceptNet common sense reasoning toolkit was created as a way to afford machine readers some basic contextual reasoning faculties needed to achieve a humanesque understanding of text and narrative. The size and scope of ConceptNet make it comparable to, what are in our opinion, the two other most notable large-scale semantic knowledge bases in the literature: Cyc [] and WordNet []. However, there are key differences. While WordNet is optimized for lexical categorization and word-similarity determination, and Cyc is optimized for formalized logical reasoning, ConceptNet is optimized for making practical context-based inferences over real-world texts. That it reasons simply and gracefully over text is perhaps owed to the fact that its knowledge representation is itself semi-structured English (a further discussion of reasoning in natural language can be found in []). ConceptNet is also unique from Cyc and WordNet for its dedication to contextual reasoning. Of the 1.6 million assertions in its knowledge base, approximately 1.25 million are dedicated to different sorts of generic conceptual connections called k-lines (a term introduced by Minsky []). Contextual commonsense reasoning is highly applicable to textual information management because it allows a computer to broadly characterize texts along interesting dimensions such as topic and affect; it also allows a computer to understand novel or unknown concepts by employing structural analogies to situate them within what is already known. The former affordance is invoked to enable psychoanalytic readings of text. The latter affordance is invoked in viewpoint simulation as a way to anticipate taste judgment responses to previously unseen fodder. §
The ConceptNet knowledge base is formed by the linking together of 1.6 million assertions (1.25 million of which are k-lines) into a semantic network of over 300,000 nodes. The present relational ontology consists of twenty relation-types, distributable into eight genres—k-lines (1.25 million assertions); things (52,000 assertions) e.g. is-a, property-of, part-of, made-of, defined-as; agents (104 assertions) e.g. capable-of; events (38,000 assertions) e.g. prerequisite-event-of, first-subevent-of, subevent-of, last-subevent-of; spatial (36,000 assertions) e.g. location-of; causal (17,000 assertions) e.g. effect-of, desirous-effect-of; functional (115,000 assertions) e.g. used-for, capable-of-receiving-action; and affective (34,000 assertions) e.g. motivation-of, desire-of.
(ConceptuallyRelatedTo ‘bad breath’ ‘mint’ ‘f=4;i=0;’) (ThematicKLine ‘wedding dress’ ‘veil’ ‘f=9;i=0;’) (IsA ‘horse’ ‘mammal’ ‘f=17;i=3;’) (PropertyOf ‘fire’ ‘dangerous’ ‘f=17;i=1;’) (PartOf ‘butterfly’ ‘wing’ ‘f=5;i=1;’) (MadeOf ‘bacon’ ‘pig’ ‘f=3;i=0;’) (CapableOf ‘dentist’ ‘pull tooth’ ‘f=4;i=0;’) (FirstSubeventOf ‘start fire’ ‘light match’ ‘f=2;i=3;’) (SubeventOf ‘play sport’ ‘score goal’ ‘f=2;i=0;’) (LocationOf ‘army’ ‘in war’ ‘f=3;i=0;’) (EffectOf ‘view video’ ‘entertainment’ ‘f=2;i=0;’) (DesirousEffectOf ‘sweat’ ‘take shower’ ‘f=3;i=1;’) (UsedFor ‘fireplace’ ‘burn wood’ ‘f=1;i =2;’) (CapableOfReceivingAction ‘drink’ ‘serve’ ‘f =0;i =14;’) (MotivationOf ‘play game’ ‘compete’ ‘f =3;i=0;’) (DesireOf ‘person’ ‘not be depressed’ ‘f=2;i=0;’) Through graph-based inference techniques such as structure-mapping and spreading activation, ConceptNet can perform these semantic tasks—finding contextual neighborhoods, making structural analogies, extracting topics from text, and estimating the emotive valence of a text. § Two specific mechanisms of ConceptNet in support of psychoanalytic reading and building the judgmental apparatus—topic extraction and structural analogy, respectively—are now reviewed. Topic extraction. ConceptNet’s topic extraction mechanism is built upon a more basic premise. With all of the complexities associated with the term ‘context’, we can begin at one very simple notion. Given a concept and no other biases, what other concepts are most relevant? The ConceptNet API provides a basic function for making this computation, called get_context(). For example when fed the input ‘go to bed’, get_context() returns the following contextual neighborhood of nodes, with relevance scores omitted: sleep
take off clothes close eye dream go to sleep brush tooth have nightmare be tire have sex take nap snore relax insomnia Technically speaking, the contextual neighborhood around a node is found by performing spreading activation radiating outward from that source node. The relatedness of any particular node is not simply a function of its link distance from the source, but also considers the number and strengths of all paths that connect the two nodes. Topic extraction via guess_topic() is a straightforward extension of the get_context() feature to accept the input of real-world documents. Its value to information retrieval and data mining is immediately evident. Using MontyLingua, a document is gisted into a sequence of verb-subject-object-object (VSOO) frames. Minor transformations are applied to each VSOO frame to massage concepts into a ConceptNet-compatible format. These concepts are heuristically assigned saliency weights based on lightweight syntactic cues, and their weighted contextual intersection is computed by get_context(). The get_context() function used in this way serves as a naïve topic spotter. To improve performance it may be desirable to designate a subset of nodes to be more suitable as topics than others. For example, we might designate ‘wedding’ as a better topic than ‘buy food’ since ConceptNet has more knowledge about its subevents (e.g. ‘walk down aisle’, ‘kiss bride’), and its parts (e.g. ‘bride’, ‘cake’, ‘reception’). Previous to the addition of this feature to ConceptNet, Eagle et al [12] used get_context() in a similar fashion to extract topics from overheard conversations. Researchers in text summarization such as Hovy and Lin have also recognized the need for symbolic general world knowledge in topic detection, which is a key component of summarization. In SUMMARIST [], Hovy and Lin give the example that the presence of the words ‘gun’, ‘mask’, ‘money’, ‘caught’, and ‘stole’ together would indicate the topic of ‘robbery’. However, they reported that WordNet and dictionary resources were relationally too sparse for robust topic detection. ConceptNet excels at this type of natural language contextual task because it is relationally richer and contains practical rather than dictionary-like knowledge. In psychoanalytic reading, ConceptNet’s guess_topic() function is leveraged to extract the chief topics from textual passages. It is a necessary way to generalize meaning in text, from a large bag of lexemes, into a smaller number of topic-classemes.
Gentner’s structure-mapping theory of analogy emphasizes formal, shared syntactic relations between concepts. In contrast, Hofstadter and Mitchell’s ‘slipnets’ [] project emphasizes semantic similarities and employs connectionist notions of conceptual distance and activation to make analogy more dynamic and cognitively plausible. Analogy in ConceptNet can be coaxed to resemble either structure-mapping or slipnets depending on whether weakly semantic relations (e.g. ‘LocationOf’, ‘IsA’) or strongly semantic relations (e.g. ‘PropertyOf’, ‘MotivationOf’) are emphasized in the analogy. Analogy in ConceptNet also has a slipnet-like connectionist property in that connections between nodes are heuristically weighted by the strength or certainty of a particular assertion. Stated concisely, two ConceptNet nodes are analogous if their sets of back-edges (incoming edges) overlap. For example, since ‘apple’ and ‘cherry’ share the back-edges, [(PropertyOf x ‘red’); (PropertyOf x ‘sweet’); (IsA x ‘fruit’)], they are in a sense, analogous concepts. Of course, it may not be aesthetically satisfying to consider such closely related things analogous (perhaps their shared membership in the set, fruit, disqualifies them aesthetically), our simple discussion will not indulge such considerations here. As with the get_context() feature, it may also be useful to apply realm-filtering to dimensionally bias the get_analogous_concepts() feature. For example, preferring to variously emphasize functional similarity versus affective similarity versus attribute similarity, certain relation-types can be weighed numerically more heavily than others. As discussed in Section 2.3, in the building of the judgmental apparatus, an individual’s location model needs to be semantically relaxed to include taste interpretations of symbols which are not in the location model—for example, an individual’s attitude about ‘macrome’ are known, but her attitude about ‘crafts’ are not known. Structural analogy allows the precedented attitude about ‘macrome’ to be relaxed unto the unprecedented attitude about ‘crafts’. Or in other words, the meanings of novel symbols are resolved within the known semantic space via structural analogy. The effect of structural analogy is to expand the semantic reach of an individual’s location model. As discussed earlier, one pitfall to this approach is that judgments expanded by structural analogy are sometimes invalid when they take place in semantic areas which lack aesthetic consistency—for example, although ‘dogs’ and ‘cats’ are structural analogs along many taxonomic dimensions, such as both being ‘pets’ and ‘animals’, they are aesthetically quite inconsistent. According to empirical data in WWTT, ‘dogs’ and ‘cats’ tended to form an aesthetic opposition—dog lovers tend toward a distaste for cats, and vice versa. Pet preference seemed particularly to be a heated and ideological space. One heuristic solution that is used when invoking get_analogous_concepts(), then, is to weight the aesthetic and affective dimensions of commonsense knowledge (e.g. motivation-of, desire-of, property-of relations) more heavily than the formal, taxonomic dimensions (e.g. defined-as, is-a, part-of relations) of knowledge. 3.5 Technology: textual affect sensing Textual affect sensing is a critical aspect of psychoanalytic reading because it illuminates the emotive context underlying each speech-act. The textual affect sensor devised to accomplish textual affect sensing in this thesis’ implemented systems attempts to characterize text from a combined analysis of its surface sentiment, the affective lexical affinities of non-sentiment words, and the common sense affective implications of everyday concepts and events. Sensed affect is measured within the Pleasure-Arousal-Dominance affect format of Mehrabian []. Each of the three textual affect sources and mechanisms—surface sentiment, lexical sentiment, and deep sentiment are treated below. But first, the selection of Mehrabian’s representation for affect is supported. Directory: ~hugo -> publications -> drafts drafts -> Rendering aesthetic impressions of text in color space publications -> English: The Lightest Weight Programming Language of them all ~hugo -> Ph. D. General Examinations General Area Exam (Prof. Pattie Maes, Examiner) Xinyu Hugo Liu Download 438.05 Kb. Share with your friends:
|