Articles Comments

The WebGenre Blog: The power of genre applied to digital information. By Marina Santini » computational models, signed posts » Mining genres with lexical affect sensing?

Mining genres with lexical affect sensing?

Post signed by: Alexander Osherenko, University of Augsburg

I gained a comprehensive knowledge in emotion recognition in texts in my PhD thesis “Opinion mining and lexical affect sensing” (http://www.informatik.uni-augsburg.de/~osherenk/promotionsvortrag_english.pdf). In my opinion, this knowledge can be utilized for identifying genres of texts — I don’t think identifying emotions differs much from identifying genres.

There are two basic categories of approach to recognize emotions in texts: a semantic approach and a statistical approach. There are also another categories of approach, for example, information fusion. However, I discuss only the first two for simplicity!

In the semantic approach specific patterns of text parts are identified and used as cues of opinions. Jan Wiebe has done much work on identifying affective patterns: she uses emotional corpora to compose a list of emotional patterns. In the genre-related case, you would use a genre-related corpora to construct this list and identify genre-specific patterns. For example, the pattern “<verb> roses” can be identified as the cue “throw roses” of the genre “love story” (excuse me for this example).

Another case is the statistical approach. Here, you compose datasets with features and classify the datasets using mathematical algorithms, for example, NaiveBayes or SVM. The only difference in contrast to emotion recognition: you have to consider genres as results of recognition and look what happens. I would extract the same features as in the emotion recognition for identifying genres: lexical (Bag-Of-Words), grammatical, stylometric, deictic. If you ask why I suggest to use, for example, stylometric features to identify genres I remind you on the work of psychologist Pennebaker who argues that also function words can be used for expressing meaning.

What is better for analysis of genre? In the context of opinion mining, the semantic approach was more beneficial to identification of emotions in short texts (a sentence). The statistical approach was beneficial for analysis of long texts (more than 200 words). I assume that the same applies to genre identification: genre of short texts should be analyzed by semantic approach; statistical approach can classify genres of long texts.

Written by

Computational Linguist, PhD

Filed under: computational models, signed posts · Tags: , , ,

11 Responses to "Mining genres with lexical affect sensing?"

  1. Leslie Barrett says:

    You might find a paper I wrote some years ago of interest. My colleagues and were able to find that (in our corpus of documents) that genre was better identified by syntactic than by lexical features:

    A Syntactic Feature Counting Method for Selecting Machine Translation Training Corpora
    Authors: Barrett, Leslie; Greenberg, David F.; Schwartz, Marc
    Source: Language and Computers, Corpus Linguistics Beyond the Word: Corpus Research from Phrase to Discourse. Edited by Eileen Fitzpatrick , pp. 1-19(19)
    Publisher: Rodopi

    1. Alexander Osherenko says:

      Leslie,

      I read your reference — thank you it was very interesting. Here are my questions.

      As far as I understood you want to analyze genres using PoS-tags. You consider, for instance, army etc. Do you think there are enough PoS-tags to classify every domain?

      Are results using lexical features really not comparable? PoS-Tagger classify tags also not perfectly.

      I mentioned different feature groups, for example, stylometric with features for authorship attribution or deictic with stopwords. Do you think these features can be extracted for genre identification?

      1. Alexander,
        You raise a number of interesting questions. I would say that there are most likely limits to the genre that can be identified using only syntactic features such as POS-densities. For example, differences in spoken corpora I imagine might fall into that category and in such cases I have seen interesting work on dialogue features that might serve this purpose. I have not seen any research specifically in genre detection in speech corpora however – an interesting thing to consider I suppose. I would very much like to repeat this study possibly expanding the features set and using different genre.

        1. Alexander Osherenko says:

          Leslie,

          now I understand better what you mean: you extract PoS-features because you analyze SPEECH corpora.

          The approaches that analyze speech corpora I read about for my PhD thesis also extract PoS-features as an ultimate mean of analysis. I also tested them for my opinion mining. But be warned! Evidently, “speech-people” like to talk not only about extracting pitch, energy, MFCCs but also about PoS features.

          Alexander

  2. marinasantini.ms@gmail.com says:

    Hi Leslie, thanks very much for the reference. I am very interested in the behaviour of syntactic features for genre classification and I will have a look at your chapter. If you have any other suggestion about genre, pls post a comment or send me an email any time.

  3. Ismaïl El Maarouf says:

    Hi,
    Exploring features for genre classification clearly is an interesting research avenue. I am myself interested in how semantic categories and relations can contribute to such a task.
    Perhaps the following paper could be of interest :
    El Maarouf I., Villaneau J., Saïd F. and Duhaut D., 2009, “Comparing child and adult language: exploring semantic constraints” in Wocci’09 ICMI-MLMI, Boston (Ma).

    We used syntactic verb-noun couples combined with an ontology of semantic types to discriminate two genres (fairy tales/press) presented here as adult/child differences for the purposes of the workshop. Though this is a preliminary work, I believe it opens interesting possibilities and I’d be glad for any criticism or similar work.

    Regards

  4. marinasantini.ms@gmail.com says:

    Thanks, Ismail!

  5. marinasantini.ms@gmail.com says:

    LinkedIn Group: Natural Language Processing

    Andrés Hohendahl

    • I was aware of this problems in HCI, thus I’ve modeled an emotional response simulator, according to Plutchick’s emotional model theories, they are blended into my NLP dialog platform by using fuzzy logic to allow some math on words and the results are amazing!

    Alexander Osherenko

    • Andrés, could you be more specific? Do you consider statistical or specific approach? What do you understand under “some math”?

    Andrés Hohendahl

    • @Alexander
    I used a specific algorithmic approach (multi-vectorial) with 8 independent vectors, coupled in antagonistic pairs, from which some ‘contradictions’ arise and hence there is also a number of general numerical (real & signed) measures for grading some common feelings, which happens to be some strange combinations of the basic ones.
    By “math on words” I will mean that: “very awful” is greater than “awful” (in some way) and to get the “numeric” representation into the model we did a sort of math-conversion on every modifier, in this way anytime a person uses Pre or Post modifiers or enhancers, even twice or more times, there is a math conversion running inside the model, in order to get the feeling in the right way. Also if you want to get the ‘status” of the model, the Fuzzy Logic conversion uses the same system to obtain “very happy” or even more for “extremely happy”, depending on the internal model state.
    Hope this may clear up your question!
    best regards!

    Alexander Osherenko

    • @Andrés
    Thank you for your answer.

    I suppose, you are analysing short texts using Fuzzy logic. How do you get weights of modifiers to transform assessment in emotion vectors? What is with emotion words, intensifiers, negations, super-/subordinate clauses in your approach? Did you tested your approach on some corpus?

    And most important: how would you modify your approach to classify genres?

    LinkedIn Group: Corpora

    Pedro Marcal

    • @Alexander,Marina, Two methods are proposed. The statistical one for short text and the semantic one for long text. Is this because of the difference in computational effort?

    Marina Santini

    • Alexander, Petro’s question is interesting… what is the motivation for two different models, one for short texts and another for longer texts? I think that longer texts might create much more noise…

    Alexander Osherenko

    @Pedro. Vice versa. Semantic methods analyze short texts; statistical methods analyze long texts.

    The motivation is — statistical methods calculate much better results on long texts than semantic (60% vs. 30% on a corpus with 5 classes). In my opinion, the reason — statistical methods dont care much, eg about particular words and compensate with the high number of features, eg Bag-of-Words. And vice versa, semantic approaches are more successful in analyzing short texts because they parse the text and consider its grammatical elements: grammatical words, intensifiers, emotion words, negations, super-/subordinate clauses. If a statistical approach analyzes long texts it CAN learn something). In contrast, if it analyzes short texts it DOESNT learn — you can see it in the confusion matrix.

    If you wonder, if it makes sense to combine these two methods (statistical/semantic) — I mixed them in my thesis and made a “hybrid” method out of them as a combination. Although I was not very successful, I think it is worth experimenting with.

    I strongly recommend you to look over the slides of my thesis (http://www.informatik.uni-augsburg.de/~osherenk/promotionsvortrag_english.pdf). I am pretty sure you will have questions. For example, I found out that classifiers and also the evaluation method in the statistical approach (NaiveBayes, SVM) dont significantly influence the classification results probably due to the overall high number of features. You will also find different hybrid methods. I also experimented with fusion — I didnt mention it in my post to simplify the discussion.

    @Marina I am not sure what you mean with “noise”. There is no noise in the analysis of short texts by a statistical approach because there is no data to to learn.

    LinkedIn Group: Text Analytics

    Ed Kool

    • Lingvistica has produced and can producre grammatical and semantical wordlists in all major languages. We are interested to cooperate for all EU languages.

    http://www.lingvistica.nl

    West European languages: English, Afrikaans, Danish, Dutch (including Flemish), French, German, Italian, Norwegian, Portuguese (including Brazilian Portuguese), Spanish (including Latin American Spanish), Swedish.

    East European languages, including Slavic and Baltic: Bulgarian, Croatian, Polish, Romanian, Russian, Rusyn, Serbian, Ukrainian; Latvian, Lithuanian,Hungarian, Czech.

    Ex-USSR languages: Armenian, Azerbaijani, Georgian, Moldavian.

    Turkish, Arabic, Korean, Chinese (Simple and Traditional), Japanese

    Alexander Osherenko

    • Ed, could you explain me what do you understand under grammatical/semantic wordlists?

  6. @Ismail

    it seems that Ismail would also use PoS-tags or their combinations for discriminating between genres similarly as Leslie proposed. Maybe PoS-approach is correct. However, I assume that automatic (statistical) approaches would reach the limits of acceptable success if you have many genres.

    I gained comparable experience with opinion mining where there are also approaches that extract PoS-features to differentiate between emotion classes. However, the achieved results do not meet expectations since they do not significantly improve classification.

    Regards
    Alexander

  7. marinasantini.ms@gmail.com says:

    LinkedIn Group: Semantic Web

    Glenn Mungra • Thanks for the great update.
    Impressive attempt to fuze statistical and semantic approach to find the emotional meaning of text/speech and to identify personality (model) traits, using the strong points of both approaches.
    The statistical approach may be more effective to determine hidden traits and conclude about function or taboo words in a larger context, but is possibly biased or outcomes may vary depending on input variables. The semantical approach may be more effective to reason about a fragment in it’s situational and conceptual context but may only hold for a specific case or domain and may lead to unexperienced conclusions.

    Alexander Osherenko
    • Thanks, Glenn! If you are talking about personality and similar things, I can point you to a very short paper of mine that studies personality traits in the context of emotion recognition. I assume it is very important — Deducing a Believable Model for Affective Behavior from Perceived Emotional Data (http://www.informatik.uni-augsburg.de/lehrstuehle/hcm/publications/2008-CAFFEi/).

  8. marinasantini.ms@gmail.com says:

    LinkedIn Group: Semantic Web

    Glenn Mungra • @Alexander: Thanks for the link. I was interested because I was following a fascinating discussion about “development of a digital being” (http://www.linkedin.com/groupItem?view=&gid=49970&type=member&item=79040200&commentID=59366968&report.success=8ULbKyXO6NDvmoK7o030UNOYGZKrvdhBhypZ_w8EpQrrQI-BBjkmxwkEOwBjLE28YyDIxcyEO7_TA_giuRN#commentID_59366968).
    I hope you don’t mind that I referred to your interesting link (about: Deducing a Believable Model for Affective Behavior from Perceived Emotional Data) in this discussion.

Leave a Reply

*


*

You may use these HTML tags and attributes: <a href="" title=""> <abbr title=""> <acronym title=""> <b> <blockquote cite=""> <cite> <code> <del datetime=""> <em> <i> <q cite=""> <strike> <strong>