We propose a Hawkish-Dovish (HD) indicator that measures the degree of 'hawkishness' or 'dovishness' of the media's perception of the ECB's tone at each press conference. We compare two methods to calculate the indicator: semantic orientation and Support Vector Machines text classification. We show that the latter method tends to provide more stable and accurate measurements of perception on a labelled test set. Furthermore, we demonstrate the potential use of this indicator with several applications: we perform a correlation analysis with a set of interest rates, use Latent Dirichlet Allocation to detect the dominant topics in the news articles, and estimate a set of Taylor rules. The findings provide decisive evidence in favour of using an advanced text mining classification model to measure the medias perception and the Taylor rule application confirms that communication plays a significant role in enhancing the accuracy when trying to estimate the bank's reaction function. JEL codes: C02, C63, E52, E58 In recent years communication became increasingly important in central banks. In particular, after the financial crisis, communication has increasingly qualified as a genuine policy tool able to steer interest rates in financial markets and drive expectations on the course of monetary policy. With official interest rates approaching zero and thereby reducing their effectiveness, various forms of forward guidance (i.e. a verbal commitment on the future course of monetary policy unconditional or conditional to some economic event) were added to the standard monetary policy toolkit existing out of interest rates and refinancing instruments for the banking sector. The growing relevance of communication in the conduct of monetary policy was mirrored by a rising interest of academicians and practitioners. In particular, a branch of economic research increasingly focused on the role of communication in adding valuable information besides what is already contained in macroeconomic variables, and in revealing policy makers preferences on the course of monetary policy with a view to enhance predictability. If the traditional approach consists in analysing how information events impact financial market developments and expectations of future policy moves, more recently analysis shifted towards analysing the language used by the central bank in its statements and how such message is perceived by its stakeholders. This paper contributes to the latter with a numerical indicator, called HD index (after the initials of hawkish and dovish), derived from media reports on the ECB press conference. Combining concepts and techniques developed in the context of computational linguistics and data mining, the indicator extracts relevant information on ECB monetary policy as reported by external observers and may therefore be interpreted as how media perceive the central banks monetary policy messages. For practical reasons, the perception is expressed on a numerical interval between-1 (most dovish) and +1 (most hawkish). In other words, the indicator indicates whether the perceived tone on monetary policy communication is predominantly on the tightening side (hawk-ish perception) or rather on the loosening side (dovish perception). Although the approach to quantify communication is not new, the approach proposed in this paper is original in various dimensions. First, the indicator does not measure directly official central bank communication but how such communication is received and interpreted with the crucial support of data mining techniques. Second, two different techniques are employed to compute the HD index: one based on the semantic orientation (SO) and the second on text classification using Support Vector Machines (SVM). The former, the most commonly used by researchers, measures how often the ECB is mentioned in a news article together with a number of given hawkish and dovish words or expressions, while the latter, computationally more complex, uses a classification model to predict the tone of an article. Both methods are applied to a data set of around 9,000 articles published between January 1999 and March 2016 in order to assess which methodology produces better results. Based on various criteria (event ECB Working Paper 2085, July 2017 2 analysis, correlations with actual interest rates and classification method) the SVM methodology tends to produce better and more reliable results. Third, in addition to its superiority on the SO, the SVM classification model can be used to analyse the terms most frequently employed by media in relation to a likely future course of monetary policy. Finally, an expanded Taylor rule framework including the HD index alongside traditional variables measuring inflation expectations and economic slack, is presented. Results suggest that the significance of the HD index as well as a relatively better fit confirm the positive role of ECB communication in enhancing the accuracy when trying to estimate the bank's reaction function, and thus that, on average, the ECB messages are correctly understood by its media watchers.